Generative AI
Last reviewed
Apr 30, 2026
Sources
34 citations
Review status
Source-backed
Revision
v6 ยท 5,308 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 30, 2026
Sources
34 citations
Review status
Source-backed
Revision
v6 ยท 5,308 words
Add missing citations, update stale details, or suggest a clearer explanation.
Generative AI refers to a category of artificial intelligence systems that can create new content, including text, images, audio, video, code, and 3D models. Unlike discriminative models that classify or label existing data, generative models learn the underlying patterns and statistical distributions within training data, then use that knowledge to produce novel outputs that resemble the original data. The technology has become one of the most significant developments in the history of computing, reshaping industries from entertainment to scientific research.
Since the public release of ChatGPT in November 2022, generative AI has moved from an academic curiosity to a mainstream technology used by hundreds of millions of people worldwide. By early 2026, enterprises have spent tens of billions of dollars adopting generative AI tools, and the technology continues to advance at a rapid pace.
The history of generative AI stretches back decades, though the field accelerated dramatically in the 2010s with the rise of deep learning.
One of the earliest examples of a generative system was ELIZA, created in 1966 by Joseph Weizenbaum at MIT. ELIZA was a rule-based chatbot that simulated a Rogerian psychotherapist, using pattern matching and substitution to generate conversational responses [1]. While ELIZA was not truly "generative" in the modern sense, it demonstrated the concept of a machine producing human-like text output.
Throughout the following decades, research in probabilistic models, neural networks, and statistical natural language processing laid the groundwork for modern generative systems. Recurrent neural networks (RNNs), and later Long Short-Term Memory (LSTM) networks, enabled models to process sequences of data, making them suitable for language and music generation.
A major milestone came in December 2013, when Diederik P. Kingma and Max Welling published "Auto-Encoding Variational Bayes," introducing the variational autoencoder (VAE) [2]. VAEs combine ideas from deep learning and probabilistic inference to learn compact latent representations of data. The key innovation was the "reparameterization trick," which allowed gradient-based optimization of the variational lower bound. VAEs became foundational tools for generating images, molecular structures, and other structured data, and they remain widely used in drug discovery and scientific applications.
In June 2014, Ian Goodfellow and colleagues at the University of Montreal published "Generative Adversarial Nets," introducing the generative adversarial network (GAN) [3]. The idea famously came to Goodfellow during a conversation at a Montreal bar, Les 3 Brasseurs, where friends asked him to help with a project on computer-generated photos. He went home and coded the first prototype that same night; it worked on the first try [4].
GANs consist of two neural networks trained simultaneously in a competitive game: a generator that creates synthetic data and a discriminator that tries to distinguish between real and generated samples. This adversarial training process drives both networks to improve, with the generator eventually producing highly realistic outputs. GANs quickly became the dominant approach for image generation, enabling applications like face synthesis, image super-resolution, and style transfer. Variants such as StyleGAN (2018) produced photorealistic human faces that were indistinguishable from real photographs.
The single most important architectural breakthrough for generative AI came in 2017, when Ashish Vaswani and seven co-authors at Google published "Attention Is All You Need" [5]. The paper introduced the Transformer architecture, which replaced recurrence and convolutions with a self-attention mechanism that allows the model to weigh the importance of different parts of the input sequence simultaneously.
Transformers proved to be far more parallelizable than RNNs and LSTMs, enabling training on much larger datasets with far greater efficiency. The architecture became the foundation for virtually every major large language model (LLM) and many image generation systems that followed.
OpenAI built on the Transformer architecture with its Generative Pre-trained Transformer (GPT) series:
| Model | Release Date | Parameters | Key Milestone |
|---|---|---|---|
| GPT-1 | June 2018 | 117 million | Demonstrated unsupervised pre-training for NLP |
| GPT-2 | February 2019 | 1.5 billion | Initially withheld due to misuse concerns |
| GPT-3 | June 2020 | 175 billion | Showed strong few-shot learning abilities |
| GPT-4 | March 2023 | Undisclosed | Multimodal (text and image input) |
GPT-3, with its 175 billion parameters, was a watershed moment. It showed that scaling up model size and training data could produce emergent capabilities, including the ability to write essays, translate languages, answer questions, and even generate code with minimal task-specific training [6].
In parallel, Google developed BERT (2018) and later the PaLM family, while Meta released the LLaMA series of open-weight models starting in February 2023.
The years 2021 and 2022 saw an explosion in text-to-image generation:
Diffusion models, which generate images by gradually denoising random noise into coherent images through a learned reverse process, became the dominant paradigm for image generation, largely supplanting GANs.
On November 30, 2022, OpenAI released ChatGPT, a conversational interface built on GPT-3.5 that had been fine-tuned using reinforcement learning from human feedback (RLHF). The launch triggered unprecedented adoption: ChatGPT reached 1 million users within five days and 100 million monthly active users within two months, making it the fastest-growing consumer application in history at that time [9].
ChatGPT made generative AI accessible to non-technical users for the first time at scale. People used it for writing, brainstorming, coding assistance, tutoring, and countless other tasks. The launch set off an industry-wide race, with Google, Anthropic, Meta, and others accelerating their own generative AI products.
At a high level, generative AI systems follow a three-stage pipeline: training on data, learning distributions, and sampling new outputs.
Generative models are trained on large datasets relevant to their output domain. A text generation model like GPT might be trained on trillions of tokens drawn from books, websites, academic papers, and code repositories. An image generation model might be trained on billions of image-text pairs scraped from the internet. During training, the model adjusts millions to trillions of internal parameters to minimize a loss function that measures how well it can predict or reconstruct the training data.
Rather than memorizing individual training examples, generative models learn the statistical distribution of the training data. A language model learns the probability distribution over sequences of tokens: given a sequence of words, what is the most likely next word? An image diffusion model learns to reverse a noise-addition process, effectively learning the distribution of natural images. This distributional learning is what allows the model to produce novel outputs that are plausible but not exact copies of training data.
Different model architectures learn these distributions in different ways:
| Architecture | How It Learns | Strengths | Common Use |
|---|---|---|---|
| Autoregressive models (GPT) | Predicts next token in sequence | Long coherent text | Text, code generation |
| Variational Autoencoders | Learns compressed latent space | Smooth interpolation | Molecular design, anomaly detection |
| GANs | Generator vs. discriminator game | Sharp, realistic images | Image synthesis, style transfer |
| Diffusion models | Learns to reverse noise process | High-quality diverse images | Image, video, audio generation |
| Flow-based models | Learns invertible transformations | Exact likelihood computation | Density estimation |
Once trained, generative models produce new content through a sampling process. For a language model, this means starting with a prompt and repeatedly sampling the next token from the learned probability distribution until the output reaches a desired length or a stop condition. Parameters like temperature control the randomness of sampling: low temperature produces more predictable, conservative outputs, while high temperature yields more creative and diverse (but potentially less coherent) results.
For diffusion models, generation starts with pure random noise and iteratively removes noise over many steps, guided by a text prompt or other conditioning signal, until a clean image emerges. Classifier-free guidance is a technique that strengthens the influence of the text prompt on the generated image, trading diversity for prompt adherence.
Generative AI spans many output modalities, each with its own set of models, techniques, and applications.
Text generation is the most widely adopted form of generative AI. Large language models like GPT-4, Claude, Gemini, and LLaMA can produce human-quality text across genres, from conversational dialogue to technical documentation. Applications include chatbots, content writing, summarization, translation, and question answering.
Text-to-image models convert natural language descriptions into visual content. The leading systems as of early 2026 include Midjourney (known for artistic quality), DALL-E 3 (integrated into ChatGPT), Stable Diffusion (open-source), and Google's Imagen. These models use diffusion-based architectures and can produce photorealistic images, illustrations, concept art, and graphic designs.
Video generation has advanced rapidly. OpenAI's Sora can generate one-minute videos with complex scenes and consistent world-building. Google's Veo 3 leads in 4K photorealism with integrated audio generation. Runway's Gen-3 model produces 10-second hyper-realistic clips with strong character consistency. Kling, developed by Kuaishou, excels at high-volume social media content. As of 2025, native audio generation has arrived in consumer video tools, and physics simulation has improved substantially [10].
Generative AI can produce speech, sound effects, and music. Text-to-speech models like ElevenLabs generate highly realistic human voices. Music generation tools like Suno and Udio can compose original songs in various genres from text prompts. Google's MusicLM and Meta's MusicGen represent research-oriented approaches to the same problem.
AI-powered code generation tools have become integral to software development. GitHub Copilot, launched in 2021 and powered by OpenAI's Codex model, reached over 15 million users by early 2025 [11]. Developers using these tools report saving 30 to 60 percent of their time on coding, test generation, and documentation tasks. As of Q1 2025, 82 percent of developers report using AI tools weekly. By 2026, coding agents that can autonomously implement features, write tests, and submit pull requests have become a major focus, with GitHub's Copilot Coding Agent and Anthropic's Claude Code leading the category.
Generative models are increasingly capable of producing 3D assets, including meshes, textures, and full scenes. Tools like OpenAI's Point-E and Shap-E, along with Nvidia's research models, can generate 3D objects from text descriptions. This has applications in gaming, architecture, virtual reality, and product design.
The generative AI field as of early 2026 is shaped by several major players, each with distinct models and strategies.
Founded in 2015, OpenAI has been the most prominent company in generative AI. Its GPT-5 family, released throughout 2025, marked a significant leap over GPT-4, with configurable reasoning effort, native multimodal input, and context windows exceeding one million tokens. GPT-5.2, released in late 2025, features a 400,000-token context window and achieved a perfect 100 percent score on the AIME 2025 math benchmark [12]. OpenAI also operates ChatGPT, which crossed 900 million weekly active users in February 2026, and produces DALL-E for image generation and Sora for video generation.
Anthropic, founded in 2021 by former OpenAI researchers Dario and Daniela Amodei, develops the Claude family of models with a strong emphasis on AI safety. The Claude 4 family (Opus 4 and Sonnet 4.5) integrates multiple reasoning approaches, including an "extended thinking mode" for deliberate self-reflection. Claude Sonnet 4.5 achieved 77.2 percent on SWE-bench Verified, making it the top-ranked coding model, and supports a one-million-token context window with computer use capabilities [13].
Google's Gemini 3.1 Pro, released in February 2026, leads on 12 of 18 tracked benchmarks and more than doubled ARC-AGI-2 performance over its predecessor. The Gemini family features a one-million-token context window and achieved 100 percent on AIME 2025 with code execution [14]. Google also produces Veo 3 for video generation and Imagen for image generation.
Meta has positioned itself as the leader in open-weight AI models. The LLaMA 4 family, released in April 2025, introduced the largest context window of any open or closed model at the time: the Scout variant supports 10 million tokens. LLaMA 4 benchmarks are competitive with GPT-4o and Gemini 2.0 Flash at a fraction of the cost [15]. Meta releases its models under open licenses, enabling widespread adoption and customization by the research community and enterprises.
Founded in 2023 in Paris by former DeepMind and Meta researchers, Mistral AI has become a significant European competitor. In December 2025, Mistral released its Mistral 3 lineup, including a large frontier model with multimodal and multilingual capabilities and nine smaller offline-capable models [16]. Mistral Large 3 is released under the Apache 2.0 license, making it the strongest fully open model developed outside of China.
Stability AI, the company behind Stable Diffusion, has championed the open-source approach to image generation. Stable Diffusion 3.5 was released in multiple sizes, offering deep control and customization. The company's models power a broad ecosystem of third-party applications, fine-tunes, and plugins.
Midjourney, an independent research lab, focuses exclusively on image generation. Midjourney v7, released in April 2025, is widely regarded as the leading model for artistic and aesthetic image quality [17]. The service operates primarily through a Discord-based interface, though a web application has also been developed.
Chinese AI startup DeepSeek stunned the industry in January 2025 with the release of DeepSeek-R1, an open-source reasoning model with 671 billion parameters released under the MIT License. The model achieves performance comparable to OpenAI's o1 on math, code, and reasoning tasks, while its predecessor V3 was reportedly trained for just $6 million, a fraction of the estimated $100 million cost for GPT-4 [18]. The DeepSeek chatbot surpassed ChatGPT as the most downloaded free app on the iOS App Store in the United States on January 27, 2025, triggering an 18 percent drop in Nvidia's share price.
| Company | Model | Release | Context Window | Notable Strength |
|---|---|---|---|---|
| OpenAI | GPT-5.2 | Late 2025 | 400K tokens | Math and reasoning benchmarks |
| Anthropic | Claude Sonnet 4.5 | 2025 | 1M tokens | Coding (SWE-bench leader) |
| Gemini 3.1 Pro | Feb 2026 | 1M tokens | Broad benchmark leadership | |
| Meta | LLaMA 4 Scout | Apr 2025 | 10M tokens | Open-weight, massive context |
| Mistral AI | Mistral Large 3 | Dec 2025 | N/A | Best open-weight non-Chinese model |
| DeepSeek | DeepSeek-R1 | Jan 2025 | 128K tokens | Cost-efficient reasoning |
Generative AI has found applications across virtually every industry.
Writers, marketers, and media professionals use generative AI for drafting articles, creating social media posts, generating marketing copy, and producing visual content. News organizations have experimented with AI-assisted reporting. Advertising agencies use text-to-image tools for rapid prototyping of visual campaigns. The film and entertainment industry uses AI for concept art, storyboarding, and visual effects pre-visualization.
Beyond code completion, generative AI assists with debugging, code review, test generation, documentation, and architecture planning. AI coding agents, which emerged as a major category in 2025, can autonomously implement features by reading codebases, writing code, running tests, and iterating on failures. Companies report that AI tools accelerate development cycles and reduce the barrier to entry for less experienced programmers.
Generative AI is being applied across multiple areas of healthcare. In drug discovery, models generate candidate molecular structures with desired properties, dramatically accelerating the early stages of pharmaceutical research. Clinical documentation tools (sometimes called "ambient scribes") listen to clinician-patient interactions and automatically produce structured notes [19]. AI systems also assist in medical imaging analysis, generating synthetic training data to improve diagnostic models, and in personalized treatment planning.
In science, generative AI accelerates research by proposing hypotheses, designing experiments, analyzing literature, and generating novel molecular and protein structures. Google DeepMind's AlphaFold, while not a generative model in the traditional sense, demonstrated the power of AI in predicting protein structures, and subsequent generative models have extended this to protein design. Weather forecasting models like Google's GraphCast use generative techniques to produce more accurate predictions.
Generative AI is reshaping education at every level. Students use AI tutors for personalized learning, instant feedback, and practice problem generation. Medical schools have begun integrating generative AI into curricula; for example, the Icahn School of Medicine at Mount Sinai provides all students access to ChatGPT Edu [20]. AI enables the creation of virtual patients for clinical training, allowing students to practice diagnoses without risk to real patients. However, educators have raised concerns about academic integrity, critical thinking erosion, and over-reliance on AI tools.
Financial institutions use generative AI for report generation, risk analysis, customer service automation, and fraud detection. Legal professionals use AI for document review, contract drafting, legal research, and summarization of case law. These applications typically involve domain-specific fine-tuning or retrieval-augmented generation to ensure accuracy.
Artists, musicians, and designers use generative AI as a creative tool for brainstorming, prototyping, and producing finished works. AI-generated art has been exhibited in galleries and has won competitions, sparking debate about the nature of creativity and authorship. Musicians use AI tools to compose backing tracks, generate melodies, and experiment with new sounds.
Despite its capabilities, generative AI presents significant risks and limitations that affect individuals, organizations, and society.
Generative AI models sometimes produce outputs that are factually incorrect, fabricated, or nonsensical, yet presented with apparent confidence. This phenomenon, known as hallucination, is an inherent consequence of how these models work: they generate statistically plausible text rather than retrieving verified facts. Hallucinations can range from minor inaccuracies to entirely fabricated citations, legal cases, or scientific claims. In 2023, a New York lawyer was sanctioned after submitting a court brief containing fictitious case citations generated by ChatGPT [21]. Mitigating hallucinations remains one of the most active areas of research, with techniques like retrieval-augmented generation (RAG), chain-of-thought reasoning, and fact-checking pipelines offering partial solutions.
Generative models reflect and can amplify biases present in their training data. These biases may manifest as racial, gender, cultural, or socioeconomic stereotypes in generated text and images. For example, text-to-image models have been shown to over-represent certain demographics in professional roles and under-represent others [22]. Addressing bias requires careful dataset curation, evaluation frameworks, and ongoing auditing of model outputs. Complete elimination of bias remains an unsolved problem, since training data drawn from the internet inherently reflects societal inequalities.
The legal status of AI-generated content and the use of copyrighted material for training are among the most contentious issues in the field. Multiple lawsuits have been filed against AI companies, including a class-action suit by authors against OpenAI and a case brought by The New York Times. Artists have filed suits against Stability AI and Midjourney over the use of their work in training data [23]. As of May 2025, the U.S. Copyright Office has indicated that copyrighted material may be used for training in "limited" ways but that "widespread scraping" is not permitted. The question of whether AI-generated works can themselves be copyrighted remains largely unresolved, with most jurisdictions requiring human authorship for copyright protection.
Generative AI makes it easy to create realistic fake images, audio, and video of real people, commonly known as deepfakes. These can be used for fraud, political manipulation, harassment, and the creation of non-consensual intimate imagery. The technical ecosystem for creating deepfakes is built largely on open-source models like Stable Diffusion [24]. Election cycles in multiple countries have already been affected by AI-generated disinformation. Detecting AI-generated content remains difficult, as generation quality continues to improve faster than detection methods.
Training and running large generative AI models requires enormous computational resources, with significant energy and water consumption. As of May 2025, data centers consume 4.4 percent of all electricity in the United States, and by 2028 more than half of data center electricity is projected to be used for AI, an increase equivalent to the energy consumption of 22 percent of U.S. households annually [25]. Data centers also require large quantities of water for cooling. Training a single large language model can emit hundreds of tons of carbon dioxide. Companies have responded by investing in renewable energy and more efficient hardware, but the environmental impact of scaling AI systems remains a growing concern.
Generative AI introduces new attack surfaces, including prompt injection (where malicious inputs manipulate model behavior), data poisoning (where training data is corrupted to introduce vulnerabilities), and the use of AI to generate sophisticated phishing emails, malware, and social engineering attacks. Securing AI systems against adversarial use is an evolving challenge.
The economic implications of generative AI are substantial and rapidly growing.
Estimates for the global generative AI market vary by research firm, but all project rapid growth:
| Research Firm | 2025 Market Size (USD) | 2026 Projected (USD) | CAGR |
|---|---|---|---|
| Grand View Research | $22.21 billion | $29.63 billion | ~33% |
| Precedence Research | $37.89 billion | $55.51 billion | ~46% |
| Fortune Business Insights | $103.58 billion | $161 billion | ~55% |
| Statista | $59.01 billion | N/A | N/A |
The wide variation in estimates reflects different definitions of what counts as the "generative AI market" (models only vs. the full application stack). Enterprise spending on generative AI reached $37 billion in 2025, a 3.2x year-over-year increase from $11.5 billion in 2024 [26].
McKinsey estimates that generative AI could unlock between $2.6 trillion and $4.4 trillion in annual economic value globally [27]. A Penn Wharton Budget Model analysis projects that AI will increase productivity and GDP by 1.5 percent by 2035, nearly 3 percent by 2055, and 3.7 percent by 2075. Generative AI is expected to augment rather than fully replace most jobs, with the greatest impact on knowledge work, customer service, software development, and creative professions.
The technology has also created new job categories, including prompt engineering, AI training and evaluation, and AI safety research. At the same time, concerns about job displacement are significant, particularly for roles involving routine text production, data entry, basic coding, and customer support.
By 2025, 88 percent of organizations reported using AI in at least one business function, and 71 percent regularly use generative AI specifically [28]. Consumer spending on generative AI apps is expected to exceed $10 billion in 2026. Gartner predicted in August 2025 that enterprise applications integrated with task-specific AI agents will increase from 5 percent to 40 percent by the end of 2026.
Governments worldwide are grappling with how to regulate generative AI, balancing innovation with safety and public interest.
The European Union's AI Act, the world's first comprehensive AI regulation, entered into force on August 1, 2024. It establishes a risk-based framework with different requirements depending on the level of risk an AI system poses [29]. Key milestones include:
The EU is also developing a Code of Practice on marking and labeling of AI-generated content, expected to be finalized in mid-2026 [30]. In November 2025, the European Commission proposed the "Digital Omnibus on AI" to streamline implementation and ease compliance burdens ahead of full application.
U.S. AI regulation has shifted significantly between administrations. On October 30, 2023, President Biden signed Executive Order 14110, "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence," the most comprehensive federal AI governance measure at the time. It required mandatory red-teaming for high-risk AI models, enhanced cybersecurity protocols, and algorithmic bias audits [31].
On January 20, 2025, President Trump revoked Biden's executive order within hours of taking office. Three days later, on January 23, 2025, Trump issued Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence," which shifted the federal approach from oversight and risk mitigation toward deregulation and promotion of AI innovation as a matter of national competitiveness [32]. The order mandated a review and potential rescission of all policies established under the Biden order that could be seen as impediments to innovation. In December 2025, the Trump administration issued a further executive order asserting federal preemption over state-level AI regulations.
China has implemented regulations requiring security assessments for generative AI services and mandating that AI-generated content be clearly labeled. The UK has taken a "pro-innovation" approach, relying on existing regulators rather than creating new AI-specific legislation. Brazil, Canada, and several other countries are developing their own AI governance frameworks.
As of early 2026, generative AI is in a phase of rapid maturation. Several trends define the current moment.
The field has moved away from a "one model fits all" philosophy. Different models now dominate different domains: Claude excels at coding, Gemini leads across broad benchmarks, GPT-5.2 dominates in math and reasoning, and LLaMA 4 offers the best open-weight option for enterprises seeking customization [33]. This specialization reflects both the diversity of use cases and the difficulty of building a single model that leads in every category.
The most significant trend in early 2026 is the rise of agentic AI, systems that can plan, perform multi-step tasks, interact with external tools, and take actions on behalf of users. Rather than simply answering questions, agentic systems can browse the web, execute code, manage files, book appointments, and complete complex workflows. Gartner predicted that 40 percent of enterprise applications will integrate task-specific AI agents by the end of 2026 [34]. GitHub Copilot's coding agent, Anthropic's Claude Code, and OpenAI's Codex agent represent early implementations of this paradigm.
Modern generative AI models increasingly handle multiple modalities (text, image, audio, video, code) within a single architecture. GPT-5, Gemini 3, and Claude 4 all accept and generate across multiple modalities, reducing the need for separate specialized models. This convergence enables new applications like real-time visual question answering, voice-driven image editing, and video generation from text descriptions.
The tension between open-weight and proprietary models continues. Meta's LLaMA, Mistral's models, and DeepSeek's releases have demonstrated that open models can achieve near-frontier performance, putting competitive pressure on closed providers. Open models enable customization, on-premise deployment, and academic research, while closed models typically offer higher peak performance and managed safety features.
DeepSeek's demonstration that competitive models can be trained for a fraction of the cost of leading Western models has intensified focus on training efficiency. Techniques like mixture-of-experts architectures, quantization, distillation, and improved data curation are reducing the compute required to achieve frontier performance. At the same time, the absolute scale of training runs continues to increase, with the largest models requiring billions of dollars in infrastructure investment.
Research into AI alignment and safety has grown substantially. Constitutional AI (developed by Anthropic), RLHF, and other alignment techniques aim to ensure that generative models behave according to human values and intentions. Red-teaming, where researchers systematically probe models for harmful outputs, has become standard practice. The field is increasingly focused on evaluating models not just for capability but for reliability, honesty, and resistance to misuse.