Generative AI

Generative AI is a class of artificial intelligence systems that produces new data instances, such as text, software code, images, audio, video, molecular structures, or other representations, by learning statistical structure from examples and then sampling or decoding from the learned model. The technical core is a generative model: a model of a data distribution, often conditioned on instructions or other inputs. This differs from a purely discriminative model, which is trained primarily to predict labels, scores, or boundaries, although one deployed system may perform both generative and discriminative tasks. The United States National Institute of Standards and Technology (NIST) defines generative AI broadly enough to include systems that emulate the structure and characteristics of input data to create derived synthetic content.^[1]

Generative AI overlaps with, but is not identical to, foundation models, language models, or machine learning as a whole. Some foundation models are not used to generate expressive content, and many generative models are small or specialized rather than broadly adaptable. A model can also generate without understanding its output in a human sense. Generated material is not guaranteed to be factual, novel, safe, or free of protected material: research has demonstrated both plausible falsehoods and reproduction of training examples.^[1][19]

The field is old as a branch of statistics and young as a consumer technology. Mixture models, hidden Markov models, and n-gram language models were generative in the formal sense decades before deep learning, and the mathematical questions they raised (how to estimate a density, how to sample from it, how to evaluate the result) are the same questions the current systems face. What changed after roughly 2018 is scale, conditioning, and the fact that the outputs became good enough to sell. By late 2025 the Federal Reserve Board, drawing on the Real-Time Population Survey, put work-related generative AI use among United States adults at about 41 percent, a diffusion rate comparable to that of the personal computer measured from its own mass-market launch.^[87][88]

Scope and terminology

Generative versus discriminative modelling

In statistical learning, a generative model represents a joint distribution, a marginal distribution, or a conditional distribution from which new samples can be drawn. A discriminative model estimates the conditional distribution of a label given an input and needs no account of how inputs arise. Andrew Ng and Michael Jordan formalized the practical consequence in 2002 by comparing naive Bayes against logistic regression on the same data: the generative classifier converged to its asymptotic error with far fewer examples, while the discriminative classifier reached a lower asymptotic error once data was plentiful.^[29] That tradeoff, more assumptions for better sample efficiency versus fewer assumptions for a better ceiling, still describes the choice.

The generative label says what a model represents, not how well it works and not what it is used for. A model that estimates a density can be used to classify, detect anomalies, impute missing values, compress, or plan, and modern deployed systems routinely do several of these at once.

What "the model" actually models

Two systems can both be generative and still be doing different things. An autoregressive large language model factorizes the joint probability of a token sequence into a product of next-token conditionals and optimizes exact log-likelihood under that factorization. A variational autoencoder posits an unobserved latent variable, defines a decoder that maps latents to data, and optimizes a lower bound on log-likelihood because the exact marginal likelihood requires an intractable integral over latents.^[4] The language model has no latent variable and an exact objective; the VAE has a latent variable, an approximate objective, and a learned representation that the language model does not produce. Calling both "generative AI" is correct and uninformative.

A useful axis is what the model gives you access to. Some families provide an explicit, exactly computable density (autoregressive models, normalizing flows). Some provide an explicit but only bounded density (VAEs, diffusion models under their variational interpretation). Some provide no density at all, only a sampler (generative adversarial networks). Some provide an unnormalized density that can be compared between points but not evaluated absolutely (energy-based models). These distinctions determine which questions the model can answer, which evaluation metrics are even defined for it, and how it can be combined with other components.^[2][30]

Why the popular term drifted

Before 2020, "generative model" was a term of art among statisticians and machine learning researchers. The consumer meaning of "generative AI" arrived with a specific product pattern: a very large model pretrained on broad data with a self-supervised objective, adapted afterwards to follow instructions, and exposed through a text or image prompt. That pattern is what the Stanford Center for Research on Foundation Models named a foundation model in 2021, deliberately choosing a term that described general-purpose pretraining and adaptation rather than any one architecture.^[3]

The drift has two consequences worth stating plainly. First, most of the generative models in production are not the ones people mean by the phrase: recommender systems, speech recognizers, spam filters, and compression codecs contain generative components that nobody markets as generative AI. Second, the commercial phrase usually refers to an application or system, not only its underlying probability model. Such a system can include data pipelines, one or more trained models, retrieval tools, safety filters, user interfaces, and human review. NIST therefore treats generative AI risk as a property of a sociotechnical system and its context of use, not as a property that can be determined from model architecture alone.^[1]

A large language model is generative when it predicts and samples sequences, but language modelling is only one part of the field. Conversely, a foundation model is defined by broad pretraining and adaptation to downstream tasks, rather than by whether every use is generative.^[3]

History

Generative modelling predates deep neural networks. Claude Shannon's 1948 paper on communication included explicit n-gram models of English and demonstrated generation by sampling from letter and word statistics, which is the same procedure a modern language model follows with a much better conditional distribution.^[31] Speech recognition and machine translation ran on hidden Markov models, Gaussian mixtures, and smoothed n-grams for three decades. Deep learning changed the scale and fidelity of generation by providing trainable nonlinear representations and large-capacity conditional models, not by inventing the idea.

Scroll sideways for more →

Two threads run through this table. One is a steady move away from architectures that constrain the model class (invertibility, tractable normalization) toward objectives that constrain only the training signal (denoising, next-token prediction, velocity regression). The other is the arrival of conditioning as the main product surface: what made these systems commercial was not unconditional sampling but the ability to steer a sample with text.

Modelling families

Each family is defined by its training objective. Architecture is largely interchangeable; the objective is what determines the model's failure modes, what it can be asked, and how expensive sampling is.

Autoregressive sequence models

An autoregressive model factorizes a joint distribution over an ordered sequence into a product of conditionals and trains by maximizing exact log-likelihood, usually as a cross-entropy loss on the next element.^[2] The objective is stable, the gradient signal is dense (every position supplies a target), and the resulting likelihood is directly comparable across models on the same tokenization. This is why the family scales so reliably and why perplexity is a meaningful, if limited, number for it.

The costs follow from the same factorization. Generation is sequential, so latency grows with output length and cannot be parallelized across positions without approximation. Errors made early are conditioned on rather than corrected, a property sometimes called exposure bias. The chosen ordering is a modelling assumption: left-to-right text is natural, raster-order pixels are not, and the autoregressive model has no notion of infilling unless it is trained for one. PixelRNN and PixelCNN demonstrated the approach on images in 2016 and made the cost visible, since generating one image required as many forward passes as there were pixels.^[32] WaveNet did the same for raw audio waveforms.^[6]

Discretization matters more here than elsewhere. Continuous data has to be turned into tokens before an autoregressive model can be applied to it, which is what VQ-VAE introduced in 2017 (a vector-quantized latent code trained with a reconstruction objective) and what every image or audio token model since has depended on.^[33] The tokenizer sets a hard ceiling on reconstruction quality that no amount of sequence modelling recovers.

Latent-variable models and VAEs

A VAE assumes data is generated by first drawing a latent variable from a prior and then decoding it. Because the marginal likelihood integrates over all latents, training maximizes the evidence lower bound instead: a reconstruction term plus a divergence term pulling the approximate posterior toward the prior. Diederik Kingma and Max Welling made this practical in 2013 by amortizing inference with an encoder network and reparameterizing the sampling step so gradients flow through it.^[4]

What a VAE is good at is representation. It gives a low-dimensional, roughly continuous latent space that supports interpolation, attribute manipulation, and downstream use, and sampling costs a single decoder pass. What it is bad at is sharpness. The bound is loose, the divergence term penalizes using the latent space aggressively, and a powerful decoder can learn to ignore the latent entirely (posterior collapse). VAE samples on natural images were visibly blurry for most of the family's history.

The family's largest practical role today is not standalone generation but compression. Nearly every production image, video, and audio generator runs its expensive generative process in a VAE latent space rather than in pixels or samples, which is what latent diffusion introduced for images^[10] and what spatiotemporal VAEs do for video. In that role the encoder-decoder pair is a fixed component and the generative modelling happens elsewhere.

Normalizing flows

A normalizing flow builds a bijection between a simple base distribution and the data distribution out of composable invertible layers. Because the transformation is invertible with a tractable Jacobian determinant, the change-of-variables formula gives exact log-likelihood, and sampling is a single forward pass.^[2][34] Danilo Rezende and Shakir Mohamed introduced flows for variational inference in 2015; Real NVP and Glow made them competitive image generators by designing coupling layers whose Jacobians are triangular and therefore cheap.^[35][36]

Exact likelihood is the selling point and the constraint is the price. Invertibility forces the latent space to have the same dimension as the data, which is wasteful for images, and it rules out most standard architectural components. Flows lost the image-quality race to GANs and then to diffusion, and the family is now most used where exact density genuinely matters: variational inference, physics and lattice field theory sampling, and density estimation on tabular or scientific data. Its ideas did not disappear. Continuous normalizing flows are the object that flow matching learned to train efficiently, which is how the family re-entered mainstream image generation through the back door.

Generative adversarial networks

A GAN trains a generator against a discriminator in a minimax game: the generator maps noise to samples, the discriminator tries to separate generated from real, and at the theoretical optimum the generator matches the data distribution.^[5] There is no likelihood anywhere in the procedure, which is exactly why GANs produced sharp images years before likelihood-based models did. Nothing in the objective rewards covering low-density regions.

That is also the family's central failure. Mode collapse occurs when the generator concentrates on a subset of the data distribution that reliably fools the current discriminator, producing high-quality samples with too little diversity. It is a consequence of the objective rather than a bug, and much of the GAN literature between 2016 and 2019 consisted of attempts to fix it: minibatch discrimination and feature matching^[39], a Wasserstein objective with weight clipping or a gradient penalty^[40], progressive growing and the style-based generator architecture.^[41] These helped without eliminating the problem, and adversarial training remained sensitive to hyperparameters and initialization in a way that diffusion training is not.

GANs still hold two niches. They are the fastest high-quality samplers available (one forward pass, no iteration), so they appear in real-time and on-device settings, and adversarial losses remain a standard ingredient in the decoders of the autoencoders that other families rely on.

Energy-based models

An energy-based model assigns a scalar energy to each configuration and defines the probability as the negative exponentiated energy divided by an intractable normalizing constant.^[37] Nothing constrains the architecture, since any function that outputs a scalar defines a valid unnormalized density, which makes the family the most flexible of the group and the hardest to train. Maximum likelihood requires samples from the model itself, which in practice means running Markov chain Monte Carlo inside the training loop; contrastive divergence and its variants approximate this by truncating the chains.^[37] Yilun Du and Igor Mordatch showed in 2019 that Langevin-based training could scale energy-based models to image generation, compositional generation, and out-of-distribution detection.^[38]

Energy-based models are rarely deployed as standalone generators. Their importance is conceptual and infrastructural: score matching was developed to avoid the normalizing constant in exactly this setting, and the score function that diffusion models estimate is the gradient of an unnormalized log-density. Diffusion is, in this sense, the energy-based programme made tractable by giving up on a single energy landscape in favour of a family of noise-smoothed ones.

Diffusion and score-based models

A diffusion model defines a fixed forward process that gradually corrupts data with noise and learns a network that reverses it. Training reduces to a weighted denoising regression: corrupt a sample to a randomly chosen noise level and predict the noise, the clean sample, or an algebraically equivalent target.^[9][42] The loss is a simple mean squared error, there is no adversarial game and no intractable partition function, and the same objective works across images, audio, video, molecules, and robot actions.

The reason it works is that the network is estimating the score, the gradient of the log-density of the noised data, at every noise level at once.^[43] Yang Song and colleagues showed in 2021 that the discrete-time and continuous-time formulations are the same object: a forward stochastic differential equation, a reverse-time SDE whose drift contains the score, and a probability-flow ordinary differential equation with the same marginals but deterministic trajectories.^[44] This is why samplers can be swapped freely, why deterministic sampling is possible from a stochastically trained model, and why sampling can be treated as a numerical integration problem with all the solver machinery that implies.

Diffusion's strengths are training stability, mode coverage (the objective touches the whole distribution, not just the modes the discriminator notices), and controllability through guidance. Its weakness is inference cost: sampling requires many sequential network evaluations, historically hundreds, now often tens or fewer with better solvers, distillation, or few-step training. Latent diffusion cut the constant factor by moving the process into a compressed space, at the cost of whatever the autoencoder cannot reconstruct.^[10] The denoising backbone moved from U-Nets to transformers, and William Peebles and Saining Xie showed that within their tested range, more forward-pass compute in a diffusion transformer correlated with better sample quality.^[49]

Flow matching and rectified flow

Flow matching trains a continuous normalizing flow by regressing a velocity field along a chosen probability path between noise and data, without simulating the ODE during training.^[46] Gaussian diffusion paths are a special case, but the framework also admits paths that diffusion does not, notably straight-line optimal-transport interpolants. Rectified flow makes the straight-path choice explicit and shows that straighter trajectories can be integrated accurately in fewer steps.^[47]

The practical result is that the distinction between diffusion and flow matching has largely dissolved in production image and video systems. Stability AI's Stable Diffusion 3 report describes a rectified flow transformer trained with a conditional flow matching objective and a noise-level sampling schedule chosen for that path^[48], and Black Forest Labs' FLUX family is built on the same construction. The vocabulary difference persists in papers; the code is close to identical. Flow matching is the more general statement, and it is inaccurate to call every flow-matching model a diffusion model.

Masked and discrete diffusion

Gaussian noise is meaningless for categorical data. Discrete diffusion instead defines corruption as a transition matrix over a finite state space, with uniform replacement, structured neighbourhoods, or an absorbing mask token as the standard choices.^[50] The masked-token variant has proved most useful. Aaron Lou and colleagues introduced score entropy to estimate probability ratios in discrete spaces^[51], and Subham Sahoo and colleagues derived a simplified masked-diffusion objective that is a mixture of masked-language-model losses and reported perplexities approaching autoregressive baselines on their benchmarks.^[52]

The appeal is parallel decoding: a discrete diffusion model can revise many positions per step rather than emitting one token at a time, and it can infill natively. The costs are that likelihood is a bound rather than an exact value, that quality per unit compute has generally lagged autoregressive models, and that key-value caching (the optimization that makes autoregressive serving cheap) does not transfer cleanly. As of mid-2026 the approach has shipped but has not displaced anything. Google demonstrated Gemini Diffusion as a research preview in May 2025 and released DiffusionGemma as open weights in June 2026, reporting generation speeds above 1,000 tokens per second on a single accelerator while scoring below the comparable autoregressive Gemma model on standard benchmarks, and recommending the autoregressive model where output quality matters more than speed.^[105]

Comparison

Scroll sideways for more →

These categories are not mutually exclusive at the system level. A product may use a transformer as the denoiser in a diffusion model, an autoregressive model to produce a latent representation, adversarial losses inside a VAE decoder, and several specialized models in one pipeline. Architecture labels alone therefore do not establish a system's capabilities or risks.

Modalities

Scroll sideways for more →

Text

Autoregressive transformers remain the dominant approach and there is no serious challenger in production.^[7][8] The visible changes since 2023 are structural rather than paradigmatic: sparse mixture-of-experts layers that raise parameter count without raising per-token compute proportionally, longer context windows, and a post-training stage that has grown from a light instruction-tuning pass into a substantial reinforcement learning phase.

The most consequential recent shift is training models to produce long intermediate reasoning before answering. DeepSeek-AI reported in Nature in September 2025 that reasoning behaviours including self-verification and backtracking emerged from reinforcement learning against automatically checkable rewards, without supervised reasoning traces, and released the resulting model.^[56] Whether this constitutes reasoning in any deeper sense is contested; what is established is that the training procedure improves measured accuracy on mathematics and coding benchmarks and increases the number of tokens generated per answer by a large factor.

Code

Code generation is autoregressive language modelling with a decisive evaluation advantage: correctness can be tested by execution. The Codex study introduced HumanEval to measure whether generated Python programs pass functional tests and found that repeated sampling substantially raised the chance of obtaining a correct program, while documenting risks from insecure or misaligned code.^[14] Execution feedback is now used in post-training as well as evaluation, which is what makes code one of the domains where reinforcement learning has clear traction.

The benchmarks have moved because the old ones stopped discriminating. HumanEval is saturated, SWE-bench Verified scores from frontier systems are reported in the high eighties and above, and evaluation attention has shifted to harder repository-level, multilingual, and terminal-based suites. Reported scores on agentic benchmarks depend heavily on the scaffold, the retry budget, and the evaluation harness, so a number is only meaningful with its full configuration attached, and vendor-reported figures should be read as vendor claims.

Images

Production image generation converged on a specific stack: a VAE that compresses pixels into a latent grid, a transformer denoiser operating on latent patches with text cross-attention or joint attention, and a rectified-flow training objective.^[10][48][49] Stability AI's Stable Diffusion 3, Black Forest Labs' FLUX, and their open-weight competitors all describe variants of this design. Classifier-free guidance remains the standard control knob and remains a distribution-distorting one: raising guidance improves prompt adherence and measured fidelity while reducing diversity.^[45]

The visible 2025-2026 change was not in this stack but around it. Image generation and editing moved inside general multimodal assistants, which handle instruction-following edits, text rendering, and multi-turn revision better than prompt-only pipelines did. Public technical detail on how those systems combine autoregressive multimodal modelling with image decoders is thin, and claims about their internals should be treated as unconfirmed unless a technical report supports them. Meanwhile ControlNet-style spatial conditioning, LoRA adapters, and DreamBooth-style subject personalization remain the standard mechanisms for control in open pipelines.

Video

Video generation is latent diffusion applied to spatiotemporal patches, with a video autoencoder compressing across time as well as space and a transformer denoiser.^[104] The compute and data requirements are much larger than for images, and temporal consistency (objects persisting, physics remaining plausible, faces staying the same) is the differentiating problem rather than per-frame quality. Sora, Veo, Kling, Runway's Gen series, and the open-weight Chinese models are all diffusion transformers by their published descriptions.

Two changes landed in 2025. Several systems began generating synchronized audio in the same pass rather than dubbing it afterwards, and interactive world models such as Genie 3 began producing navigable environments frame by frame in response to user actions rather than rendering a fixed clip. The second is closer to a world model than to a video codec, and it is the direction that connects generation to robotics and simulation.

Audio and music

Two approaches coexist rather than one dominating. Audio-token language models discretize audio with a neural codec and run an autoregressive model over the resulting tokens, a lineage running from WaveNet through AudioLM and MusicGen.^[6] Latent diffusion models generate a compressed audio representation directly and decode it, as in Stability AI's Stable Audio. Both produce usable music; the token approach tends to handle long-range structure through the language model, and the diffusion approach tends to handle fidelity and speed.

The commercial change in this modality was legal rather than technical. Universal Music Group settled with Udio in October 2025 and Warner Music Group settled with both Udio and Suno in November 2025, in each case combining a compensatory settlement with a licence for a new platform trained on authorized recordings.^[81] Sony Music litigation and Universal's case against Suno continued, and the American Federation of Musicians sued Universal and Warner over whether those licences triggered compensation obligations to performers under existing collective agreements. The result is a partial shift from unlicensed training to licensed catalogues in music specifically, well ahead of any comparable shift in text or images.

Speech

Text-to-speech has consolidated on hybrid designs: an autoregressive model over semantic or coarse acoustic tokens, followed by a flow-matching or diffusion decoder that produces the fine acoustic detail. Neural codec language models established zero-shot voice cloning from seconds of reference audio, and flow-matching decoders improved fidelity and inference speed over purely autoregressive codec stacks. Production systems from ElevenLabs and the major model providers, and open systems such as F5-TTS and CosyVoice, follow variants of this pattern.

The deployment change is duplex streaming speech inside general assistants, where the model consumes audio and emits audio with low enough latency for interruption to work. On the recognition side, Whisper and its successors made robust multilingual transcription a commodity. The same capability that enables accessibility and localization enables impersonation, and voice cloning from short reference samples is the reason consent and verification questions attach to this modality more directly than to the others.

3D and scene generation

3D generation had no equivalent of ImageNet, so the field solved the data problem by borrowing from 2D. The dominant pipelines either generate multiple consistent views with an image diffusion model and reconstruct geometry from them, or run diffusion and rectified flow directly over a structured 3D latent (voxel grids, triplanes, or sets of primitives) trained on curated 3D asset collections. Gaussian splatting has largely displaced NeRF as the runtime representation for captured and generated scenes because it renders fast and edits reasonably, and support for it has been moving into mainstream asset pipelines. Text-to-3D quality is adequate for game and visualization prototyping and is not yet at the level where generated assets pass without artist cleanup.

Molecules, proteins, and materials

Scientific generative models operate on structured representations where geometry, symmetry, and physical constraints are part of the modelling problem, not decorations on it. A protein backbone generator must respect rotational and translational equivariance; a crystal generator must respect periodicity and space-group symmetry. This is why the same diffusion mathematics appears with entirely different state spaces and corruption kernels.

RFdiffusion adapted a structure-prediction network for denoising and reported experimental validation of a subset of designed proteins.^[15] AlphaFold 3 replaced its predecessor's structure module with a diffusion module and extended prediction to complexes involving ligands, nucleic acids, and ions.^[98] In November 2025 a team led by David Baker reported fine-tuning RFdiffusion to design antibody variable domains against chosen epitopes, combining computational design with laboratory screening and biophysical validation of selected designs.^[100] Microsoft's MatterGen applied a symmetry-aware diffusion process to inorganic crystals with property conditioning, and reported synthesizing one generated material and measuring its property within 20 percent of the target.^[99]

These are evidence for specific pipelines, not proof that generated scientific candidates are valid without simulation, laboratory testing, or domain review. Success rates for de novo design remain low in absolute terms, screening remains the bottleneck, and a single synthesized compound is a proof of concept rather than a throughput claim.

Tabular and synthetic data

Tabular generation is the least glamorous and most operationally common case: mixed continuous and categorical columns, strong inter-column dependencies, small datasets, and privacy as the usual motivation. Diffusion models adapted to mixed types and transformer-based autoregressive models over cell sequences are the current research leaders, with GANs and copula methods still in production use.

The evaluation problem here is unusually well defined and unusually often botched. A synthetic table must be judged on three axes at once: statistical resemblance to the source, downstream utility when a model is trained on synthetic and tested on real, and privacy leakage measured by membership inference or attribute disclosure attacks. Methods that score well on resemblance frequently do so by memorizing rows, and the 2025 benchmarking literature converged on reporting all three together, with differentially private training as the standard mitigation and a measurable utility cost. Synthetic data generated this way is a risk-reduction tool, not a free replacement for real data.

Training

Pretraining data and provenance

Pretraining consumes web crawls, digitized books, code repositories, scientific text, images with alt text, and licensed collections. Common Crawl, The Pile, and LAION are the best-documented public examples, and the composition of frontier proprietary corpora is generally not disclosed. Quality filtering, deduplication, and mixture weighting have as much effect on final capability as architecture does, and are usually the least documented part of a technical report.

Provenance is the field's most concrete unsolved problem, and it has three separate layers. Legally, the right to copy works for training is contested and jurisdiction-dependent (see below). Technically, the web is not a stable resource: the Data Provenance Initiative's longitudinal audit of 14,000 domains found that between 2023 and 2024, restrictions in robots.txt and terms of service rendered roughly 5 percent of tokens in the C4 corpus and about 28 percent of its most actively maintained sources fully restricted, with terms-of-service restrictions covering a much larger share, and documented frequent inconsistency between what sites say in robots.txt and what they say in their terms.^[68] Practically, the same audit noted that these protocols were never designed for this purpose, so consent signals are ambiguous even when present.

Data documentation practice improved unevenly. Model cards and system cards are now standard for major releases, but disclosure of training data composition moved in the opposite direction as litigation increased.

Scaling behaviour

Jared Kaplan and colleagues reported in 2020 that language model loss follows smooth power laws in parameters, data, and compute over several orders of magnitude, with the exponent stable enough to extrapolate.^[53] Jordan Hoffmann and colleagues corrected the allocation in 2022: for a fixed compute budget, parameters and training tokens should scale at roughly equal rates, implying about 20 tokens per parameter, which meant that the large models of the time were substantially undertrained.^[54] The Chinchilla result changed the industry's training recipe immediately.

Two qualifications matter. Scaling laws describe pretraining loss, not downstream usefulness, and the mapping between them is empirical and task-dependent. And the Chinchilla optimum minimizes training compute, not total cost of ownership: if a model will serve many billions of tokens, it is rational to train a smaller model on far more data than the compute-optimal ratio, which is what most deployed models now do. Since 2024 a second scaling axis has become economically significant, namely compute spent on post-training and on inference, which is discussed below.

Post-training

Broad pretraining is followed by fine-tuning that determines almost everything a user perceives. In instruction tuning, demonstrations teach the desired input-output format and the model's default register.^[11] Reinforcement learning from human feedback uses human comparisons to train a reward model and then optimizes a policy against it, as in the InstructGPT study.^[11] Direct preference optimization reformulates the same preference data as a classification objective on the policy itself, removing the separate reward model and RL loop.^[12] Both depend entirely on the quality and representativeness of their preference data, and both optimize measured preference rather than correctness. A reward model trained on human ratings will reward whatever raters rewarded, including length, confidence, and formatting.

Reinforcement learning with verifiable rewards replaces the learned reward model with a programmatic checker: a mathematics answer matcher, a unit test, a compiler, or a formal proof checker.^[57] This removes reward hacking against a learned proxy and restricts the method to domains where correctness is machine-checkable, which is why its gains concentrate in mathematics, code, and structured tool use rather than in open-ended writing. DeepSeek-R1 demonstrated that the approach can induce long-form reasoning behaviour from a base model without supervised reasoning data.^[56]

None of this creates a guarantee of correctness, harmlessness, or alignment. Post-training changes how a model responds to people; it does not install knowledge the pretrained model lacks, and it can degrade capabilities it was not designed to preserve.

Distillation

Knowledge distillation trains a smaller student to match a larger teacher's output distribution rather than the original hard labels, on the argument that the teacher's full distribution carries more information per example.^[55] It is now the standard way to produce the cheap tier of a model family, and in generative modelling it has a second role: distilling a many-step diffusion or flow sampler into a few-step or one-step one, which is how real-time image generation became possible.

Distillation from another provider's model raises a separate set of contractual and evidentiary questions, and terms of service for major APIs generally prohibit using outputs to train competing models. Detecting it is difficult, and public accusations have generally rested on circumstantial evidence.

Inference

Sampling and decoding

For an autoregressive model, decoding turns a sequence of distributions into text, and the choice is consequential. Greedy decoding and beam search maximize likelihood and produce degenerate, repetitive text for open-ended generation, because the most probable continuation of a long passage is not what human text looks like. Ari Holtzman and colleagues documented this and introduced nucleus sampling, which truncates to the smallest set of tokens whose cumulative probability exceeds a threshold and samples within it.^[58] Temperature rescales the logits, and top-p and top-k truncate the tail. Structured tasks with a single correct answer favour low temperature; open-ended tasks favour higher.

For diffusion and flow models, the equivalent choice is the solver and the step count. Deterministic ODE samplers give reproducible outputs from a fixed seed; stochastic samplers trade reproducibility for diversity. Speculative decoding accelerates autoregressive inference by drafting several tokens with a small model and verifying them in one pass of the large one, producing identical output distributions when implemented correctly.^[59]

Guidance, conditioning, and control

Classifier-free guidance trains conditional and unconditional behaviour jointly by dropping the condition on some fraction of training examples, then extrapolates between the two predictions at sampling time.^[45] It is the reason text-to-image models follow prompts as closely as they do, and it is not free: the guided sampler no longer targets the model's learned distribution, and high guidance reduces diversity and can amplify artifacts.

Beyond text prompts, control enters through spatial conditioning (edge maps, depth, segmentation, pose), reference images for subject or style, masks for inpainting, and lightweight adapters fine-tuned on a handful of examples. For language models the equivalent surfaces are the system prompt, structured output constraints and grammars, tool definitions, and retrieval-augmented generation, which conditions the model on retrieved documents.^[13] Retrieval can supply current or attributable evidence, but irrelevant, incomplete, malicious, or misread sources still produce wrong answers, so a deployed system needs controls for document provenance, access permissions, citation fidelity, and conflicts among sources.

Conditional generation does not guarantee that a requested constraint will be satisfied. Applications requiring exact physical, chemical, legal, or safety constraints need task-specific validation outside the generative model.

Inference-time compute

The observation that spending more compute at inference improves accuracy predates the current systems: sampling several answers and taking the majority, or sampling many programs and keeping those that pass tests, both work.^[14][60] Charlie Snell and colleagues studied the tradeoff directly and reported that on some problem distributions, additional inference compute is more effective than an equivalent increase in model size, and that the optimal allocation depends on problem difficulty.^[61]

Reasoning models internalize this by generating long chains of intermediate tokens before answering, trained with verifiable rewards to make that computation useful rather than decorative.^[56] The economic consequence is that inference cost is now variable per request in a way it was not in 2022, and that latency and cost budgets have become product design parameters rather than infrastructure details. The technique also has diminishing and sometimes negative returns: longer chains raise the chance of compounding an early error, and on easy problems the extra tokens buy nothing.

Evaluation

Why generative evaluation is hard

There is no single score that establishes the quality of a generative system, and the difficulty is structural rather than a matter of building a better metric. Generation is one-to-many: for most prompts many outputs are equally good, so any metric that compares against a single reference is measuring the wrong thing. The quantity a probabilistic model optimizes is not the quantity a user cares about. And the properties that matter most in deployment (factual accuracy, absence of copied material, appropriateness for a given audience) are properties of individual outputs, whereas most metrics are properties of distributions.

Lucas Theis, Aaron van den Oord, and Matthias Bethge established the sharpest version of this point in 2016: average log-likelihood, Parzen window estimates, and visual fidelity of samples are largely independent of each other in high dimensions, so good performance on one implies nothing about the others.^[30] A model can achieve excellent likelihood and produce poor samples, or produce excellent samples with terrible likelihood. Evaluation must therefore be matched to modality, task, user population, deployment environment, and harm model.

Where automated metrics fail

Perplexity measures how well a model predicts held-out text under its own tokenization. It is a valid measure of distributional fit and a poor proxy for quality: it is not comparable across tokenizers or corpora, it rewards memorization of the evaluation domain, and it correlates weakly with human judgements of usefulness on open-ended tasks. It also cannot be computed at all for GANs and is only a bound for VAEs and diffusion models.

BLEU compares n-gram overlap against reference translations.^[62] Chris Callison-Burch, Miles Osborne, and Philipp Koehn showed in 2006 that BLEU improvements do not reliably indicate translation quality improvements, that systems of different types cannot be fairly compared by it, and that the metric can be gamed by matching reference n-gram statistics without producing better translations.^[63] The criticism was published two decades ago and the metric is still reported. ROUGE inherits the same problems for summarization.

Fréchet Inception Distance compares Gaussian fits to Inception features of generated and reference images.^[16] Sadeep Jayasumana and colleagues documented three specific failures at CVPR 2024: the Inception feature space poorly represents the content modern text-to-image models produce, the normality assumption is wrong, and the estimator is biased at practical sample sizes, so FID can contradict human raters and fail to reflect genuine iterative improvement.^[64] Improved precision and recall metrics at least separate sample quality from distribution coverage, which a single aggregate score cannot do.^[17] All of these depend on the feature extractor and the reference dataset, and none establishes that individual outputs are original, appropriate, or correctly labelled.

The pattern is consistent. Automated metrics are useful for tracking relative progress within a fixed setup and for catching regressions. They are not evidence of quality, they degrade as models improve past the range the metric was designed for, and optimizing against them directly produces models that score well and behave worse.

Human evaluation and arena ranking

Human evaluation is the fallback, and it has its own failure modes: rater expertise, task ambiguity, position and length bias, fatigue, and cost. Careful protocols with defined rubrics, multiple raters, and reported agreement statistics are expensive enough that they are usually reserved for final claims.

Arena-style ranking is the scaled version. Chatbot Arena collects pairwise preferences from anonymous users on their own prompts and fits a Bradley-Terry model to produce a leaderboard.^[65] It solves the prompt-distribution problem (real users bring real tasks) and the reference problem (no reference needed). Its known biases are substantial. Shivalika Singh and colleagues analysed roughly two million battles across 243 models over sixteen months and reported that some providers were permitted to test many private variants and publish only the best score, which violates the sampling assumptions of the rating model; that sampling rates and data access differed systematically between proprietary and open models; and that prompt duplication across months was non-trivial.^[66] The platform disputed parts of the analysis and subsequently changed some policies. Separately, arena preferences reward style, length, formatting, and agreeableness independently of correctness, which means an arena ranking measures what anonymous users prefer, not what is true or useful.

Using a strong model as an automatic judge scales human-style evaluation further and inherits the same biases plus self-preference: judges tend to score outputs from their own family higher.

Contamination

Benchmark contamination occurs when evaluation data, or close paraphrases of it, appears in pretraining data. It inflates scores in a way that is invisible from the outside and that grows worse as benchmarks age and get discussed online. Hugh Zhang and colleagues quantified it by building GSM1k, a set of 1,205 new grade-school mathematics problems matched to GSM8K in difficulty, and reported accuracy drops of up to about 13 percent for some model families while others showed no gap, which is the signature of memorization rather than capability.^[67]

Contamination is why benchmark numbers should be read with the model's training cutoff in mind, why held-out and freshly-generated evaluation sets matter, and why saturation on a public benchmark is ambiguous between "the problem is solved" and "the answers are in the training data."

Safety and sociotechnical evaluation

Capability evaluation does not settle risk. A framework proposed by Laura Weidinger and colleagues at Google DeepMind separates model capability, human interaction, and systemic impact, on the grounds that the same capability produces different effects under different access controls, user populations, and social conditions.^[18] NIST recommends measuring risks in context, documenting assumptions and limitations, using independent evaluators where appropriate, and monitoring behaviour after deployment.^[1] For code, tests can be executed; for scientific output, simulators and experiments apply; for language, factuality checks should distinguish whether a claim follows from supplied evidence, agrees with a trusted source, or reflects current real-world knowledge. Passing a test suite does not prove the absence of defects, and a promising computational score does not replace empirical validation.

Failure modes and limitations

Hallucination

Generative systems produce confident but unsupported statements, usually called hallucination or confabulation. NIST describes confabulation as a natural consequence of generating from learned statistical patterns and notes that it becomes especially consequential in health, law, and finance.^[1]

Adam Tauman Kalai and colleagues at OpenAI argued in 2025 that the phenomenon has two distinct sources. In pretraining, hallucination arises from the same statistical pressures that produce errors in binary classification: if valid and invalid statements are not separable from the training signal, some invalid ones will be produced, and facts appearing once in the corpus are especially prone to it. In post-training, hallucination persists because most benchmarks score with binary accuracy, which rewards guessing over abstention, so a model optimized to score well learns to answer rather than say it does not know. Their proposed remedy is to change how mainstream evaluations are graded rather than to add another hallucination benchmark.^[69] The argument is a company-authored theoretical account, not a measurement, and the practical claim that scoring changes reduce hallucination in deployment remains to be established at scale.

The documented downstream cost is easiest to see in law, where fabricated citations are detectable and sanctionable. Damien Charlotin's AI Hallucination Cases database, which records only decisions where a court found or clearly implied reliance on hallucinated material, listed on the order of 1,600 cases worldwide by mid-2026, up from roughly 200 a year earlier.^[70] Retrieval, prompting, scaling, and human feedback reduce errors in defined settings; none removes the need to verify consequential claims.

Mode collapse and diversity loss

Mode collapse is the GAN-specific failure in which the generator maps many inputs to a narrow set of outputs that reliably fool the discriminator, losing coverage of the data distribution while retaining per-sample quality.^[39] It is diagnosable only with coverage-sensitive metrics, which is one reason precision and recall were separated from FID.^[17]

The general phenomenon is broader than GANs. Preference-tuned language models exhibit a related narrowing: optimizing toward average rater preference reduces stylistic and substantive diversity across samples, and high classifier-free guidance does the same for image models.^[45] In each case the mechanism is the same, namely an objective that rewards the centre of the distribution and does not penalize abandoning the tails.

Memorization and regurgitation

Models memorize training data, and the amount is measurable. Nicholas Carlini and colleagues extracted more than a thousand training examples from the image diffusion models and settings they studied, including photographs and identifying information.^[19] A separate CVPR study developed retrieval methods for detecting replicated training content and found replication across several evaluated datasets and models.^[102] Milad Nasr and colleagues showed that a simple divergence attack on a production chat model caused it to emit memorized training data verbatim.^[74]

The capacity side was quantified in 2025. Jack Morris and colleagues separated unintended memorization (information about a specific dataset) from generalization (information about the data-generating process) and estimated GPT-style models at roughly 3.6 bits of memorization capacity per parameter, with unintended memorization rising until capacity fills and then falling as generalization takes over.^[73] This reframes the question usefully: memorization is not a defect to be eliminated but a finite budget that is spent on whatever appears often enough or is distinctive enough, which is why duplicated and outlier records are the ones that come back out.

These results disprove the categorical claim that generative models never reproduce training material. They do not imply that every output is copied, and they do not determine the copyright status of any particular model or output. Deduplication, differential privacy, output filtering, and memorization audits address different parts of the risk and none of them solves it.

Bias

Training data encodes social stereotypes, historical inequities, measurement error, and uneven representation, and outputs differ in quality and harmfulness across languages, regions, demographic groups, and topics. The measurement literature for image generation is the most concrete. Federico Bianchi and colleagues found at FAccT 2023 that ordinary prompts mentioning traits, occupations, or objects produced stereotyped outputs, that the effect persisted for prompts that did not mention identity at all, and that explicitly countering a stereotype in the prompt did not reliably remove it.^[75]

Averages hide these disparities, so evaluation needs relevant subgroups and conditions without treating demographic attributes as simple or universally defined. Mitigations applied at the prompt or output layer can produce their own failures, including historically inaccurate outputs and inconsistent treatment across groups, which is an argument for measuring the intervention rather than assuming it helps.

Model collapse

Recursive use of generated material changes future training distributions. Ilia Shumailov and colleagues reported in Nature in 2024 that training successive generations of models on data produced by their predecessors caused progressive loss of distributional information: first the tails, then convergence toward a low-variance distribution bearing little resemblance to the original.^[20] They demonstrated it in language models, VAEs, and Gaussian mixture models, and decomposed the cause into statistical approximation error from finite sampling, expressivity error from a limited model class, and approximation error from the learning procedure itself.

The scope condition matters and is routinely dropped in popular coverage. The headline experiments replaced the training data at each generation rather than accumulating it. Matthias Gerstgrasser and colleagues showed that this distinction is decisive: with replacement, test error grows without bound in the number of iterations, but when real and synthetic data accumulate, test error has a finite upper bound independent of the number of iterations, and they confirmed this across model sizes and architectures for language models, diffusion models on molecules, and VAEs on images.^[71] Elvis Dohmatob and colleagues gave a complementary theoretical account, showing that synthetic data can change the exponent of a scaling law, degrading how performance improves with scale rather than causing outright collapse.^[72]

The honest summary is that indiscriminate recursive training on model output degrades models under conditions that are well characterized in controlled settings, that accumulating real data alongside synthetic data prevents the unbounded version of the failure, and that carefully curated synthetic data mixed with verified real data is a standard and effective technique used in current production training. Whether the open web's growing share of machine-generated text has measurably degraded real training runs is not established; nobody has published a controlled demonstration of it at frontier scale, and the estimates circulating for what fraction of the web is AI-generated come from detector-based methods whose error rates are themselves unvalidated.

Society and law

Copyright litigation

Copyright questions differ for training inputs, model outputs, and human-authored contributions, and the answers vary by jurisdiction. In the United States, the Copyright Office concluded in January 2025 that purely AI-generated material is not protected by copyright, while human-authored expression, creative selection or arrangement, and creative modifications can be protected; it also concluded that with then-current generally available technology, prompts alone usually do not provide sufficient human control, and that authorship remains a case-by-case inquiry.^[22] The Office's May 2025 pre-publication report on generative AI training did not establish a blanket rule that training on copyrighted works is lawful or unlawful; it described fair use as fact-specific, examined licensing and market effects, and noted unresolved litigation and policy choices.^[23] As of August 2026 the final version of Part 3 has not been published, and the Office has said it does not expect substantive changes to its analysis.

The litigation position as of August 2026 is that trial courts have split the question into two parts, and that no United States appellate court has ruled on the merits.

Scroll sideways for more →

Two generalizations survive the detail. First, courts have so far treated the act of training more favourably than the act of acquiring the corpus, which shifts legal risk toward data sourcing rather than model building. Second, the fourth fair use factor, market effect, is where the outcomes are being decided, and the version of it that matters (whether generated output dilutes the market for the works trained on) has not yet been litigated on a developed evidentiary record. Outside the United States, the UK judgment turned largely on where training occurred rather than on whether it was lawful, and the EU relies on a text and data mining exception with an opt-out that rightsholders have increasingly exercised. Dataset licences, contracts, privacy law, rights of publicity, and sector-specific rules impose obligations separate from copyright.

Provenance and watermarking

Provenance systems record how content was created or edited. The C2PA specification defines signed, tamper-evident manifests that travel with media and describe its origin and modification history, and successive revisions have been published as the coalition's membership and production implementations in editing software, cameras, and model providers have grown.^[24] Content Credentials can show that a participating actor made a signed assertion and whether the bound asset was altered. They do not prove that content is true, they can be stripped, and a re-encode or screenshot removes them unless a durable signal is also present.

Watermarking embeds a detectable signal in the content itself. Sumanth Dathathri and colleagues at Google DeepMind published SynthID-Text in Nature in October 2024, describing a scheme that modifies only the sampling procedure, preserves text quality by their measurements, and supports detection without access to the model; they reported deployment in Gemini and released the library.^[76] Text watermarking is weaker than image watermarking in principle, since paraphrasing or translation removes it, and published attacks recover or forge watermark signals under some conditions.

NIST's synthetic-content report treats provenance, watermarking, detection, and labelling as complementary techniques with different failure modes rather than as infallible detectors, and this is the correct framing.^[25] Detection classifiers for AI-generated text in particular have documented false-positive problems, with disproportionate error rates on text by non-native English writers, and should not be used as sole evidence in consequential decisions.

Deepfakes and non-consensual imagery

Generated media supports impersonation, fraud, harassment, non-consensual intimate imagery, disinformation, and automated influence operations. Non-consensual intimate imagery is the harm with the clearest legal response. The United States TAKE IT DOWN Act was signed in May 2025, criminalizing knowing publication of non-consensual intimate visual depictions including digital forgeries, and requiring covered platforms to provide a notice-and-removal process that acts within 48 hours; the Federal Trade Commission began enforcing the platform obligations on 19 May 2026 and issued warning letters ahead of the deadline.^[82] In the EU, Article 50 of the AI Act requires disclosure that content is artificially generated, with the machine-readable marking obligation for systems already on the market deferred to 2 December 2026 by the Digital Omnibus amendments.^[27][97]

Fraud is the harm with the clearest measured cost, and the measurement is poor. The FBI's Internet Crime Complaint Center reported roughly $893 million in adjusted losses across about 22,000 complaints carrying an AI-related descriptor in 2025, its first year tracking the category separately, out of about $20.9 billion in total reported cybercrime losses; the FBI states the AI figure is an undercount because victims often do not know AI was involved.^[103] Most widely circulated deepfake fraud totals come from vendor reports with undisclosed methodology and should not be cited as measurements.

Defensive and beneficial uses exist in the same technologies: simulation, translation, accessibility tools, restoration, red-teaming, and synthetic training data for detection systems. Risk depends on the model, user, scale, safeguards, and deployment context, so AI safety and AI alignment measures must be assessed against specified threats rather than described as universal solutions.^[1]

Labour

The empirical picture is genuinely mixed, and the mix is informative rather than a failure of the literature.

Scroll sideways for more →

The consistent pattern across the task-level studies is that generative AI compresses the performance distribution: it helps novices more than experts, helps most on tasks squarely inside the model's competence, and can hurt on tasks that require deep context the model lacks, which is the most plausible reading of the METR result on experienced maintainers of code they wrote themselves. The gap between the task-level gains and the firm-level and economy-level nulls is the central open question, and it has an obvious candidate explanation in the time required for organizational and process change.

Energy and water

Environmental cost is the topic where public numbers are least reliable, because training and inference are routinely conflated, per-query figures are quoted without their measurement boundary, and water withdrawal is confused with water consumption.

At the sector level, the IEA reported that global data centre electricity consumption reached roughly 485 TWh in 2025, growing 17 percent year on year, with AI-focused data centre consumption growing about 50 percent, and projects data centre consumption roughly doubling by 2030 to about 950 TWh, near 3 percent of global electricity demand.^[96] Its 2025 Energy and AI report gave a similar 2030 figure of about 945 TWh and treats these as scenario projections with wide uncertainty, not forecasts.^[21] For the United States specifically, Lawrence Berkeley National Laboratory estimated data centres at about 4.4 percent of national electricity in 2023 (176 TWh, up from 58 TWh in 2014) and gave a 2028 range of 6.7 to 12 percent, or 325 to 580 TWh, a range wide enough to signal how uncertain the inputs are.^[93]

At the per-query level, the only detailed production measurement comes from Google, which reported in August 2025 that the median Gemini Apps text prompt consumed 0.24 Wh of energy, 0.26 mL of water, and 0.03 gCO2e, and that the same figures had fallen 33-fold and 44-fold respectively over the preceding twelve months.^[94] The measurement boundary includes accelerator power, host CPU and memory, idle capacity, and data centre overhead, with the accelerator accounting for only 58 percent of the total. It excludes training, excludes embodied emissions from hardware manufacture and data centre construction, reports a median rather than a mean (so it is not the right number to multiply by query volume), covers text prompts rather than image or video generation, and uses market-based rather than location-based emissions accounting, a choice that credits purchased clean energy rather than reflecting the local grid mix. Those are the documented limits of the figure, and citing 0.24 Wh as "the energy cost of an AI query" ignores every one of them.

For a full life cycle, Mistral AI published an analysis of Mistral Large 2 conducted under the AFNOR Frugal AI methodology with external audit, reporting 20.4 ktCO2e, 281,000 cubic metres of water consumption, and 660 kg antimony-equivalent resource depletion over training plus 18 months of service, with training and inference accounting for about 85.5 percent of emissions and 91 percent of water, and impacts scaling roughly with model size for a fixed number of generated tokens.^[95] Sam Altman's June 2025 figures for ChatGPT (0.34 Wh and 0.000085 gallons per query) were published in a personal blog post without methodology and should not be treated as a measurement.^[108]

NIST reports that there is no agreed method for estimating generative AI environmental impacts, and the IEA treats projections as uncertain and distinguishes AI workloads from broader data centre demand.^[1][21] A single energy or emissions figure should not be generalized across generative systems, and the direction of travel is genuinely two-sided: per-query efficiency is improving quickly while total consumption is rising quickly, and neither trend cancels the other.

Economics and deployment

Adoption

Adoption of the consumer products has been fast by historical standards. Alexander Bick, Adam Blandin, and David Deming found that 39.4 percent of the United States population aged 18 to 64 reported using generative AI in August 2024, with 28 percent of employed respondents using it at work, and that work adoption matched the personal computer's diffusion rate measured from each technology's first mass-market product.^[87] Their quarterly survey put work-related use at about 41 percent by November 2025, with wide sectoral variation from about 70 percent in information industries to about 21 percent in accommodation and food services.^[88]

Enterprise adoption is a different measurement and a weaker one, because most published figures are self-reported vendor or consultancy surveys. Anthropic's Economic Index, which classifies patterns of use in its own traffic, reported that the share of Claude.ai conversations classified as augmentation rather than automation moved between roughly 41 and 55 percent over 2025 and into 2026, with automation more common in programmatic API traffic than in the consumer product.^[106] That is a measurement of one provider's traffic, not of the economy.

Where returns have been weak

The most-cited claim about enterprise returns is the MIT NANDA report "The GenAI Divide: State of AI in Business 2025," widely summarized as finding that 95 percent of enterprise AI pilots fail. What the report actually did was interview 52 executives, survey 153 leaders, and review roughly 300 public deployments, and conclude that 95 percent of the pilots it examined showed no measurable profit-and-loss impact, attributing the gap to organizational approach rather than model quality.^[91] It is not peer reviewed, the sample is small and not randomly drawn, "no measurable P&L impact" is not the same as failure, and the headline number has been repeated far beyond what the method supports. It is best read as a hypothesis about integration difficulty, and the Danish administrative-data study reaching precise null effects on earnings and hours is much stronger evidence for the same general conclusion.^[89]

The pattern that does hold up across sources is that measurable returns concentrate where output is verifiable and workflows are narrow: customer support with resolution metrics, software engineering with tests, document processing with ground truth, translation and transcription, and code-adjacent internal tooling. Returns are weakest where the output is open-ended, the quality bar is high, and verification costs as much as production.

The cost curve

Inference prices have fallen faster than almost any comparable technology, and the rate depends entirely on which capability level is held fixed. Epoch AI's analysis across six benchmarks found price declines for a fixed performance level ranging from about 9-fold to about 900-fold per year, with a median of about 50-fold, and noted that the fastest trends begin after January 2024; holding GPT-4-level performance on PhD-level science questions fixed, the price fell about 40-fold per year.^[92] The caveats are the usual ones: benchmarks are imperfect proxies for usefulness, models may be overfit to them, and reasoning models generate far more tokens per answer, so a lower price per token does not translate directly into a lower price per task.

Three structural facts follow. Capability at a fixed price improves faster than capability at the frontier, so the economically interesting question is usually what last year's frontier now costs rather than what this year's frontier can do. The shift toward inference-time compute partially offsets the price decline for the hardest tasks. And capital expenditure on training and serving infrastructure has grown far faster than revenue from generative AI products, which is the basis for the ongoing argument about whether current investment levels are justified; that argument is about forecasts, and no citation settles it.

Governance and standards

Governance combines technical testing with policies for data rights, access, human oversight, disclosure, procurement, incident reporting, and accountability. NIST AI 600-1 is a voluntary companion to the AI Risk Management Framework: it organizes generative AI risks and proposes actions across governance, mapping, measurement, and management rather than certifying a model as safe.^[1] ISO/IEC 42001:2023 specifies requirements for an organization-wide AI management system; it applies to AI broadly and is not a generative-AI product certification.^[26] In the United States, the AI Safety Institute established at NIST in 2023 was reorganized in June 2025 as the Center for AI Standards and Innovation, with a stated focus on demonstrable risks in cybersecurity, biosecurity, and chemical and biological weapons.^[107]

The European Union's AI Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024, with prohibitions applying from February 2025 and general-purpose AI obligations from 2 August 2025.^[27] The timetable was then amended. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was voted by Parliament on 16 June 2026, approved by Council on 29 June, published in the Official Journal on 24 July, and entered into force on 27 July 2026; it deferred the high-risk obligations for stand-alone Annex III systems from 2 August 2026 to 2 December 2027, and for AI embedded in regulated products under Annex I to 2 August 2028.^[97] Article 50 transparency duties still applied from 2 August 2026, covering disclosure that a user is interacting with an AI system, marking of synthetic audio, image, video and text as artificially generated, and labelling of deepfakes and AI-generated news content, with the marking obligation for systems already on the market deferred to 2 December 2026. The Commission published a voluntary Code of Practice on Transparency of AI-generated Content on 10 June 2026 to support compliance with those duties.^[28]

Compliance with one framework does not establish compliance everywhere. A responsible deployment identifies the applicable law and use-specific standards, records the model and data lineage, evaluates foreseeable misuse and affected groups, defines human authority to intervene, and monitors incidents and material changes throughout the system lifecycle.

References

1. ^National Institute of Standards and Technology. "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile." NIST AI 600-1, July 2024. nvlpubs.nist.gov/...NIST.AI.600-1.pdf
2. ^Bond-Taylor, Sam, Adam Leach, Yang Long, and Chris G. Willcocks. "Deep Generative Modelling: A Comparative Review of VAEs, GANs, Normalizing Flows, Energy-Based and Autoregressive Models." IEEE Transactions on Pattern Analysis and Machine Intelligence 44, no. 11 (2022). arxiv.org/...2103.04922
3. ^Bommasani, Rishi, et al. "On the Opportunities and Risks of Foundation Models." Stanford Center for Research on Foundation Models, 2021. arxiv.org/...2108.07258
4. ^Kingma, Diederik P., and Max Welling. "Auto-Encoding Variational Bayes." 2013. arxiv.org/...1312.6114
5. ^Goodfellow, Ian J., et al. "Generative Adversarial Nets." Advances in Neural Information Processing Systems 27 (2014). papers.nips.cc/...eb0234979a61f95710dbe25-Abstract
6. ^van den Oord, Aaron, et al. "WaveNet: A Generative Model for Raw Audio." 2016. arxiv.org/...1609.03499
7. ^Vaswani, Ashish, et al. "Attention Is All You Need." Advances in Neural Information Processing Systems 30 (2017). papers.nips.cc/...47dee91fbd053c1c4a845aa-Abstract
8. ^Brown, Tom B., et al. "Language Models are Few-Shot Learners." Advances in Neural Information Processing Systems 33 (2020). arxiv.org/...2005.14165
9. ^Ho, Jonathan, Ajay Jain, and Pieter Abbeel. "Denoising Diffusion Probabilistic Models." Advances in Neural Information Processing Systems 33 (2020). proceedings.neurips.cc/...67f1ab10179ca4b-Abstract
10. ^Rombach, Robin, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. "High-Resolution Image Synthesis With Latent Diffusion Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. openaccess.thecvf.com/...on_Models_CVPR_2022_paper
11. ^Ouyang, Long, et al. "Training Language Models to Follow Instructions with Human Feedback." Advances in Neural Information Processing Systems 35 (2022). arxiv.org/...2203.02155
12. ^Rafailov, Rafael, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. "Direct Preference Optimization: Your Language Model is Secretly a Reward Model." Advances in Neural Information Processing Systems 36 (2023). proceedings.neurips.cc/...6ce7-Abstract-Conference
13. ^Lewis, Patrick, et al. "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." Advances in Neural Information Processing Systems 33 (2020). proceedings.neurips.cc/...bc26945df7481e5-Abstract
14. ^Chen, Mark, et al. "Evaluating Large Language Models Trained on Code." 2021. arxiv.org/...2107.03374
15. ^Watson, Joseph L., et al. "De Novo Design of Protein Structure and Function with RFdiffusion." Nature 620 (2023). nature.com/...s41586-023-06415-8
16. ^Heusel, Martin, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. "GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium." Advances in Neural Information Processing Systems 30 (2017). proceedings.neurips.cc/...65871369074926d-Abstract
17. ^Kynkaanniemi, Tuomas, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. "Improved Precision and Recall Metric for Assessing Generative Models." Advances in Neural Information Processing Systems 32 (2019). proceedings.neurips.cc/...8c70ff313743079-Abstract
18. ^Weidinger, Laura, et al. "Sociotechnical Safety Evaluation of Generative AI Systems." Google DeepMind, 2023. deepmind.google/...45425
19. ^Carlini, Nicholas, et al. "Extracting Training Data from Diffusion Models." 32nd USENIX Security Symposium, 2023. usenix.org/...carlini
20. ^Shumailov, Ilia, et al. "AI Models Collapse When Trained on Recursively Generated Data." Nature 631 (2024). nature.com/...s41586-024-07566-y
21. ^International Energy Agency. "Energy and AI." April 2025. iea.org/...energy-and-ai
22. ^United States Copyright Office. "Copyright and Artificial Intelligence, Part 2: Copyrightability." January 2025. copyright.gov/...art-2-Copyrightability-Report.pdf
23. ^United States Copyright Office. "Copyright and Artificial Intelligence, Part 3: Generative AI Training, Pre-Publication Version." May 2025. copyright.gov/...eport-Pre-Publication-Version.pdf
24. ^Coalition for Content Provenance and Authenticity. "C2PA Technical Specification, Version 2.2." May 2025. spec.c2pa.org/...C2PA_Specification
25. ^National Institute of Standards and Technology. "Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency." NIST AI 100-4, November 2024. nvlpubs.nist.gov/...NIST.AI.100-4.pdf
26. ^International Organization for Standardization. "ISO/IEC 42001:2023: Information Technology - Artificial Intelligence - Management System." December 2023. iso.org/...42001
27. ^European Parliament and Council of the European Union. "Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artificial Intelligence." Official Journal of the European Union, 2024. eur-lex.europa.eu/...eng
28. ^European Commission. "Code of Practice on Transparency of AI-generated Content." Shaping Europe's Digital Future, June 10, 2026. digital-strategy.ec.europa.eu/...generated-content
29. ^Ng, Andrew Y., and Michael I. Jordan. "On Discriminative vs. Generative Classifiers: A Comparison of Logistic Regression and Naive Bayes." Advances in Neural Information Processing Systems 14 (2002). ai.stanford.edu/...01-discriminativegenerative.pdf
30. ^Theis, Lucas, Aaron van den Oord, and Matthias Bethge. "A Note on the Evaluation of Generative Models." International Conference on Learning Representations, 2016. arxiv.org/...1511.01844
31. ^Shannon, Claude E. "A Mathematical Theory of Communication." Bell System Technical Journal 27 (1948). people.math.harvard.edu/...entropy.pdf
32. ^van den Oord, Aaron, Nal Kalchbrenner, and Koray Kavukcuoglu. "Pixel Recurrent Neural Networks." Proceedings of the 33rd International Conference on Machine Learning, 2016. proceedings.mlr.press/...oord16
33. ^van den Oord, Aaron, Oriol Vinyals, and Koray Kavukcuoglu. "Neural Discrete Representation Learning." Advances in Neural Information Processing Systems 30 (2017). arxiv.org/...1711.00937
34. ^Rezende, Danilo Jimenez, and Shakir Mohamed. "Variational Inference with Normalizing Flows." Proceedings of the 32nd International Conference on Machine Learning, 2015. proceedings.mlr.press/...rezende15
35. ^Dinh, Laurent, Jascha Sohl-Dickstein, and Samy Bengio. "Density Estimation Using Real NVP." International Conference on Learning Representations, 2017. arxiv.org/...1605.08803
36. ^Kingma, Diederik P., and Prafulla Dhariwal. "Glow: Generative Flow with Invertible 1x1 Convolutions." Advances in Neural Information Processing Systems 31 (2018). arxiv.org/...1807.03039
37. ^LeCun, Yann, Sumit Chopra, Raia Hadsell, Marc'Aurelio Ranzato, and Fu Jie Huang. "A Tutorial on Energy-Based Learning." In Predicting Structured Data, MIT Press, 2006. yann.lecun.com/...lecun-06.pdf
38. ^Du, Yilun, and Igor Mordatch. "Implicit Generation and Modeling with Energy Based Models." Advances in Neural Information Processing Systems 32 (2019). arxiv.org/...1903.08689
39. ^Salimans, Tim, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. "Improved Techniques for Training GANs." Advances in Neural Information Processing Systems 29 (2016). arxiv.org/...1606.03498
40. ^Arjovsky, Martin, Soumith Chintala, and Leon Bottou. "Wasserstein Generative Adversarial Networks." Proceedings of the 34th International Conference on Machine Learning, 2017. proceedings.mlr.press/...arjovsky17a
41. ^Karras, Tero, Samuli Laine, and Timo Aila. "A Style-Based Generator Architecture for Generative Adversarial Networks." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. openaccess.thecvf.com/..._Networks_CVPR_2019_paper
42. ^Sohl-Dickstein, Jascha, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. "Deep Unsupervised Learning Using Nonequilibrium Thermodynamics." Proceedings of the 32nd International Conference on Machine Learning, 2015. proceedings.mlr.press/...sohl-dickstein15
43. ^Song, Yang, and Stefano Ermon. "Generative Modeling by Estimating Gradients of the Data Distribution." Advances in Neural Information Processing Systems 32 (2019). proceedings.neurips.cc/...1a96dcd947c7d93-Abstract
44. ^Song, Yang, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. "Score-Based Generative Modeling through Stochastic Differential Equations." International Conference on Learning Representations, 2021. openreview.net/forum
45. ^Ho, Jonathan, and Tim Salimans. "Classifier-Free Diffusion Guidance." 2022. arxiv.org/...2207.12598
46. ^Lipman, Yaron, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. "Flow Matching for Generative Modeling." International Conference on Learning Representations, 2023. openreview.net/forum
47. ^Liu, Xingchao, Chengyue Gong, and Qiang Liu. "Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow." International Conference on Learning Representations, 2023. arxiv.org/...2209.03003
48. ^Esser, Patrick, et al. "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis." Proceedings of the 41st International Conference on Machine Learning, 2024. arxiv.org/...2403.03206
49. ^Peebles, William, and Saining Xie. "Scalable Diffusion Models with Transformers." Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. openaccess.thecvf.com/...nsformers_ICCV_2023_paper
50. ^Austin, Jacob, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. "Structured Denoising Diffusion Models in Discrete State-Spaces." Advances in Neural Information Processing Systems 34 (2021). proceedings.neurips.cc/...e97125b70e6973d-Abstract
51. ^Lou, Aaron, Chenlin Meng, and Stefano Ermon. "Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution." Proceedings of the 41st International Conference on Machine Learning, 2024. proceedings.mlr.press/...lou24a
52. ^Sahoo, Subham Sekhar, et al. "Simple and Effective Masked Diffusion Language Models." Advances in Neural Information Processing Systems 37 (2024). proceedings.neurips.cc/...e0ad-Abstract-Conference
53. ^Kaplan, Jared, et al. "Scaling Laws for Neural Language Models." 2020. arxiv.org/...2001.08361
54. ^Hoffmann, Jordan, et al. "Training Compute-Optimal Large Language Models." Advances in Neural Information Processing Systems 35 (2022). arxiv.org/...2203.15556
55. ^Hinton, Geoffrey, Oriol Vinyals, and Jeff Dean. "Distilling the Knowledge in a Neural Network." 2015. arxiv.org/...1503.02531
56. ^DeepSeek-AI (Guo, Daya, et al.). "DeepSeek-R1 Incentivizes Reasoning in LLMs through Reinforcement Learning." Nature 645, no. 8081 (September 2025): 633-638. nature.com/...s41586-025-09422-z
57. ^Lambert, Nathan, et al. "Tulu 3: Pushing Frontiers in Open Language Model Post-Training." Allen Institute for AI, 2024. arxiv.org/...2411.15124
58. ^Holtzman, Ari, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. "The Curious Case of Neural Text Degeneration." International Conference on Learning Representations, 2020. arxiv.org/...1904.09751
59. ^Leviathan, Yaniv, Matan Kalman, and Yossi Matias. "Fast Inference from Transformers via Speculative Decoding." Proceedings of the 40th International Conference on Machine Learning, 2023. proceedings.mlr.press/...leviathan23a
60. ^Wei, Jason, et al. "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models." Advances in Neural Information Processing Systems 35 (2022). arxiv.org/...2201.11903
61. ^Snell, Charlie, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. "Scaling LLM Test-Time Compute Optimally Can Be More Effective than Scaling Model Parameters." 2024. arxiv.org/...2408.03314
62. ^Papineni, Kishore, Salim Roukos, Todd Ward, and Wei-Jing Zhu. "BLEU: A Method for Automatic Evaluation of Machine Translation." Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, 2002. aclanthology.org/P02-1040
63. ^Callison-Burch, Chris, Miles Osborne, and Philipp Koehn. "Re-Evaluating the Role of BLEU in Machine Translation Research." Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, 2006. aclanthology.org/E06-1032
64. ^Jayasumana, Sadeep, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. "Rethinking FID: Towards a Better Evaluation Metric for Image Generation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. openaccess.thecvf.com/...eneration_CVPR_2024_paper
65. ^Chiang, Wei-Lin, et al. "Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference." Proceedings of the 41st International Conference on Machine Learning, 2024. arxiv.org/...2403.04132
66. ^Singh, Shivalika, et al. "The Leaderboard Illusion." Advances in Neural Information Processing Systems 38, Datasets and Benchmarks Track, 2025. arxiv.org/...2504.20879
67. ^Zhang, Hugh, et al. "A Careful Examination of Large Language Model Performance on Grade School Arithmetic." Advances in Neural Information Processing Systems 37 (2024). arxiv.org/...2405.00332
68. ^Longpre, Shayne, et al. "Consent in Crisis: The Rapid Decline of the AI Data Commons." Advances in Neural Information Processing Systems 37, Datasets and Benchmarks Track, 2024. arxiv.org/...2407.14933
69. ^Kalai, Adam Tauman, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang. "Why Language Models Hallucinate." OpenAI, September 2025. arxiv.org/...2509.04664
70. ^Charlotin, Damien. "AI Hallucination Cases." Database, HEC Paris, accessed 2026. damiencharlotin.com/hallucinations
71. ^Gerstgrasser, Matthias, et al. "Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data." 2024. arxiv.org/...2404.01413
72. ^Dohmatob, Elvis, Yunzhen Feng, Pu Yang, Francois Charton, and Julia Kempe. "A Tale of Tails: Model Collapse as a Change of Scaling Laws." Proceedings of the 41st International Conference on Machine Learning, 2024. arxiv.org/...2402.07043
73. ^Morris, John X., et al. "How Much Do Language Models Memorize?" 2025. arxiv.org/...2505.24832
74. ^Nasr, Milad, et al. "Scalable Extraction of Training Data from (Production) Language Models." 2023. arxiv.org/...2311.17035
75. ^Bianchi, Federico, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. "Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale." Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency. dl.acm.org/...3593013.3594095
76. ^Dathathri, Sumanth, et al. "Scalable Watermarking for Identifying Large Language Model Outputs." Nature 634 (October 2024). nature.com/...s41586-024-08025-4
77. ^Bartz et al. v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal.). Summary judgment order on fair use, June 23, 2025; class settlement of at least $1.5 billion announced September 5, 2025. Official settlement administration site: anthropiccopyrightsettlement.com
78. ^Kadrey et al. v. Meta Platforms, Inc., No. 3:23-cv-03417 (N.D. Cal.). Order on partial summary judgment, June 25, 2025. Analysis: Norton Rose Fulbright, "Two US Decisions Find That Reproducing Works to Train Large Language Models Is Fair Use, Part 2: Kadrey v Meta." nortonrosefulbright.com/...roducing-works-to-train
79. ^Getty Images (US) Inc. and others v. Stability AI Ltd, [2025] EWHC 2863 (Ch), High Court of Justice of England and Wales, November 4, 2025. bailii.org/...2863
80. ^In re: OpenAI, Inc. Copyright Infringement Litigation, MDL No. 3143 (S.D.N.Y.), consolidating The New York Times Company v. Microsoft Corporation and OpenAI, Inc., No. 1:23-cv-11195 (S.D.N.Y.). United States Judicial Panel on Multidistrict Litigation, pending MDL dockets. jpml.uscourts.gov/pending-mdls
81. ^Universal Music Group and Udio. "Universal Music Group and Udio Announce Udio's First Strategic Agreements for New Licensed AI Music Creation Platform." October 29, 2025. prnewswire.com/...usic-creation-platform-302599129
82. ^Federal Trade Commission. "Take It Down Act Enforcement Starts Now: What to Know About the FTC and TIDA." May 2026. ftc.gov/...ent-starts-now-what-know-about-ftc-tida
83. ^Noy, Shakked, and Whitney Zhang. "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence." Science 381, no. 6654 (2023): 187-192. science.org/...science.adh2586
84. ^Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. "Generative AI at Work." Quarterly Journal of Economics 140, no. 2 (2025): 889-942. academic.oup.com/...7990658
85. ^Dell'Acqua, Fabrizio, Edward McFowland III, Ethan R. Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, Francois Candelon, and Karim R. Lakhani. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality." Harvard Business School Working Paper 24-013, 2023; published in Organization Science, 2025. papers.ssrn.com/...papers.cfm
86. ^METR. "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity." July 10, 2025. metr.org/...early-2025-ai-experienced-os-dev-study and "We Are Changing Our Developer Productivity Experiment Design." February 24, 2026. metr.org/...2026-02-24-uplift-update
87. ^Bick, Alexander, Adam Blandin, and David J. Deming. "The Rapid Adoption of Generative AI." Management Science, 2025; NBER Working Paper 32966. nber.org/...w32966
88. ^Board of Governors of the Federal Reserve System. "Monitoring AI Adoption in the U.S. Economy." FEDS Notes, April 3, 2026. federalreserve.gov/...-in-the-u-s-economy-20260403
89. ^Humlum, Anders, and Emilie Vestergaard. "Large Language Models, Small Labor Market Effects." Becker Friedman Institute Working Paper 2025-56; NBER Working Paper 33777, 2025. bfi.uchicago.edu/...els-small-labor-market-effects
90. ^Brynjolfsson, Erik, Bharat Chandar, and Ruyu Chen. "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." Stanford Digital Economy Lab, November 2025. digitaleconomy.stanford.edu/...ficial-intelligence
91. ^Challapally, Aditya, Chris Pease, Ramesh Raskar, and Pradyumna Chari. "The GenAI Divide: State of AI in Business 2025." MIT NANDA, July 2025. mlq.ai/..._State_of_AI_in_Business_2025_Report.pdf
92. ^Epoch AI. "LLM Inference Prices Have Fallen Rapidly but Unequally across Tasks." 2025. epoch.ai/...llm-inference-price-trends
93. ^Shehabi, Arman, et al. "2024 United States Data Center Energy Usage Report." Lawrence Berkeley National Laboratory, December 2024. eta-publications.lbl.gov/...rgy-usage-report_1.pdf
94. ^Elsworth, Cooper, et al. "Measuring the Environmental Impact of Delivering AI at Google Scale." Google, August 2025. arxiv.org/...2508.15734
95. ^Mistral AI. "Our Contribution to a Global Environmental Standard for AI." July 2025. mistral.ai/...global-environmental-standard-for-ai
96. ^International Energy Agency. "Data Centre Electricity Use Surged in 2025, Even with Tightening Bottlenecks Driving a Scramble for Solutions." 2026. iea.org/...enecks-driving-a-scramble-for-solutions
97. ^European Parliament and Council of the European Union. "Regulation (EU) 2026/1744 (Digital Omnibus on AI) Amending Regulation (EU) 2024/1689." Official Journal of the European Union, July 24, 2026. digital-strategy.ec.europa.eu/...ibus-enters-force
98. ^Abramson, Josh, et al. "Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3." Nature 630 (2024): 493-500. nature.com/...s41586-024-07487-w
99. ^Zeni, Claudio, et al. "A Generative Model for Inorganic Materials Design." Nature, January 2025. nature.com/...s41586-025-08628-5
100. ^Bennett, Nathaniel R., et al. "Atomically Accurate De Novo Design of Antibodies with RFdiffusion." Nature 649 (2026): 183-193; published online November 5, 2025. nature.com/...s41586-025-09721-5
101. ^Radford, Alec, et al. "Learning Transferable Visual Models from Natural Language Supervision." Proceedings of the 38th International Conference on Machine Learning, 2021. proceedings.mlr.press/...radford21a
102. ^Somepalli, Gowthami, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. "Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. openaccess.thecvf.com/...Diffusion_CVPR_2023_paper
103. ^Federal Bureau of Investigation, Internet Crime Complaint Center. "2025 Internet Crime Report." 2026. ic3.gov/...Reports
104. ^Blattmann, Andreas, et al. "Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets." 2023. arxiv.org/...2311.15127
105. ^Google. "DiffusionGemma: 4x Faster Text Generation." June 10, 2026. blog.google/...fusion-gemma-faster-text-generation
106. ^Anthropic. "Anthropic Economic Index Report: Cadences." June 2026. anthropic.com/...economic-index-june-2026-report
107. ^National Institute of Standards and Technology. "Center for AI Standards and Innovation (CAISI)." Accessed 2026. nist.gov/caisi
108. ^Altman, Sam. "The Gentle Singularity." June 10, 2025. blog.samaltman.com/the-gentle-singularity