# DALL-E

> Source: https://aiwiki.ai/wiki/dall-e
> Updated: 2026-07-29
> Fact-checked: 2026-07-29
> Categories: Diffusion Models, Generative AI, Image Generation, OpenAI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "DALL-E." aiwiki.ai, 29 Jul 2026. https://aiwiki.ai/wiki/dall-e
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

DALL-E is a family of text-to-image systems developed by [OpenAI](https://aiwiki.ai/wiki/openai) that generates images from natural-language descriptions. OpenAI announced the original DALL-E in January 2021 as a 12-billion-parameter transformer trained on text-image pairs.[1] Its successors changed the generation pipeline substantially: [DALL-E 2](https://aiwiki.ai/wiki/dall_e_2) used a CLIP-latent prior and diffusion decoder, while [DALL-E 3](https://aiwiki.ai/wiki/dall_e_3) emphasized training on detailed synthetic captions and used ChatGPT to turn a user's request into a model-facing prompt.[2][3][4] The name therefore refers to a product and research lineage, not to one unchanged architecture.

DALL-E was an early large-scale demonstration that an autoregressive model could combine concepts, attributes, styles, and spatial instructions in generated images without task-specific training. The later systems improved resolution, photorealism, editing, and prompt adherence, but their published evaluations also documented failures in object-attribute binding, spatial relations, spelling, and detailed scene composition.[2][3][4] Independent studies found additional limitations in DALL-E 2's handling of syntax and in its representation of gender across occupations.[19][20]

By the research cutoff of July 28, 2026, OpenAI's developer catalog described DALL-E 3 as a "previous generation" model and marked both DALL-E 2 and DALL-E 3 as deprecated. The catalog listed GPT Image 2 as its current image-generation model.[5] This status does not retroactively make the GPT Image systems part of the DALL-E series; they are a later, differently named product line.

## Scope and terminology

OpenAI's typography places a centered dot between DALL and E. This article uses the ASCII form "DALL-E" and covers the shared history of the original model, DALL-E 2, and DALL-E 3. Separate articles provide more detail on those two successors.

The three releases should not be treated as successive checkpoints of the same neural network:

| System | Public introduction | Main published design |
|---|---:|---|
| DALL-E | January 5, 2021 | Autoregressive transformer over text and discrete image tokens |
| DALL-E 2 | April 2022 | CLIP-image-embedding prior followed by a diffusion decoder and two diffusion upsamplers |
| DALL-E 3 | October 2023 product rollout | Text-to-image system trained largely on detailed synthetic captions; published appendix describes a VAE latent space and a distilled diffusion decoder |

The term "DALL-E" was also used in 2022 product announcements as shorthand for the DALL-E 2 service. Consequently, contemporary OpenAI posts may say "DALL-E" even when the deployed model was DALL-E 2. Dates and technical descriptions are more reliable identifiers than the abbreviated product wording.

## Original DALL-E

### Architecture and training

The original DALL-E converted a caption and an image into one autoregressive sequence. Images were preprocessed at 256 x 256 pixels, compressed by a discrete [variational autoencoder](https://aiwiki.ai/wiki/autoencoder) into a 32 x 32 grid of 1,024 image tokens, and concatenated with as many as 256 byte-pair-encoded text tokens. A 12-billion-parameter sparse [transformer](https://aiwiki.ai/wiki/transformers) modeled the resulting sequence. Its attention pattern allowed image tokens to attend to the text and to earlier image tokens.[1][2]

The ICML paper reported 250 million text-image pairs collected from the internet for the large model. The dataset included Conceptual Captions and a filtered subset of YFCC100M. It did not include the MS-COCO dataset directly, although the authors found that part of the MS-COCO validation image set overlapped with YFCC100M and reported deduplicated evaluation results to control for that overlap.[2]

Generation involved sampling many candidate token sequences and then using [CLIP](https://aiwiki.ai/wiki/clip) to rank the decoded images for agreement with the caption. In the public demonstration, OpenAI displayed the top 32 among 512 candidates for each interactive caption. This distinction matters: the quality of a selected gallery image reflected both the generative transformer and a large best-of-N reranking stage, rather than a single unfiltered sample.[1]

### Demonstrated capabilities

OpenAI tested the model on prompts involving object attributes, multiple objects, perspective, internal and external views, transformations of existing images, visual analogies, geographic references, and combinations of otherwise unrelated concepts. The same autoregressive objective allowed the model to regenerate a rectangular region extending to the lower-right of an existing image when given a compatible caption.[1]

In the paper's zero-shot MS-COCO comparison with DF-GAN, human evaluators chose DALL-E's sample as more realistic 90.0% of the time and as the better caption match 93.3% of the time. Each DALL-E entry in that comparison was the best of 512 candidates according to the contrastive reranker, so the result should not be read as a comparison of single raw samples. The authors also reported that the model's MS-COCO Frechet Inception Distance was within two points of the best prior system in their test after accounting for overlap in the image data.[2]

### Limitations

The original demonstration showed that success varied with prompt phrasing. OpenAI reported that the system confused associations between objects and their colors as more objects were added, and semantically equivalent rephrasings could produce different results. It also presented some capabilities through selected or reranked samples, which limits what can be inferred about their frequency in ordinary generation.[1]

The model produced 256 x 256 images and its discrete-token pipeline discarded fine visual detail during compression. It was a research result rather than the architecture later used in DALL-E 2: DALL-E 2 did not simply enlarge the first transformer's output. It replaced the main generation method with a system built around CLIP representations and [diffusion models](https://aiwiki.ai/wiki/diffusion_model).[3]

## DALL-E 2

### The unCLIP pipeline

OpenAI's DALL-E 2 paper called its research architecture "unCLIP." The pipeline first encoded the caption with a frozen CLIP model. A prior then predicted a CLIP image embedding from the caption and its CLIP text embedding. Finally, a diffusion decoder generated an image conditioned on that predicted image embedding. The authors tested both autoregressive and diffusion priors and found the diffusion prior more compute-efficient and better in their comparisons.[3]

The base diffusion decoder generated a 64 x 64 image. Two additional diffusion models upsampled it first to 256 x 256 and then to 1,024 x 1,024. This hierarchical design is the source of the "four times greater resolution" comparison in OpenAI's product material: DALL-E 2's 1,024-pixel side length was four times the original DALL-E's 256-pixel side length.[3][6]

The paper's component table, rather than an aggregate marketing figure, is the reliable parameter disclosure for the experimental stack. It lists a 1-billion-parameter prior, a 3.5-billion-parameter 64 x 64 decoder, a 700-million-parameter first upsampler, and a 300-million-parameter second upsampler. The table lists both the autoregressive and diffusion prior alternatives at 1 billion parameters, but only one prior is used in a generation path. The paper did not present "3.5B plus 1.5B upsampler" as the architecture.[3]

For the experiments described in the paper, OpenAI sampled equally from CLIP and DALL-E datasets totaling about 650 million images to train the image encoder. It trained the prior, decoder, and upsamplers only on the approximately 250-million-image DALL-E dataset. The authors also stated that the production version used a different data distribution filtered for aesthetic quality and safety, so the experimental dataset description is not a complete inventory of the deployed service.[3]

### Generation, variations, and editing

Because CLIP maps images and text into a shared representation space, unCLIP could produce variations that retained high-level semantics and style while changing details not preserved in an embedding. The paper also demonstrated interpolation between two image embeddings and language-guided changes to an image.[3] These research capabilities supported several product modes:

- Text-to-image generation created a new image from a caption.
- Variations used an image as the starting semantic and stylistic reference.
- Inpainting regenerated a masked region while taking the surrounding image and prompt into account.
- Outpainting extended an image beyond its original boundaries. OpenAI released the desktop outpainting feature on August 31, 2022 and said it could continue shadows, reflections, and textures while creating a larger canvas.[7]

OpenAI's launch comparison reported that evaluators preferred DALL-E 2 to the original DALL-E 71.7% of the time for caption matching and 88.8% for photorealism.[6] These are pairwise preferences reported by the developer, not universal accuracy rates, and they do not measure every prompt type.

### Deployment

DALL-E 2 began with a limited research preview in spring 2022. OpenAI expanded it into a beta, added editing features, and removed the waitlist on September 28, 2022. At that point, OpenAI reported more than 1.5 million active users creating more than 2 million images per day.[8] Those figures describe a dated service snapshot and should not be treated as current usage.

On November 3, 2022, OpenAI opened the DALL-E API in public beta. It reported that more than 3 million people were then using the service and producing more than 4 million images per day.[9] The API made image generation available inside third-party software and moved DALL-E from a stand-alone creative interface into a developer platform.

### Published weaknesses

The unCLIP paper found that the system was worse than a comparable GLIDE model at binding attributes to the correct objects. Its example involving a red cube on top of a blue cube often mixed the colors. The authors attributed part of this problem to CLIP image embeddings not explicitly preserving object-attribute bindings.[3]

The paper also said DALL-E 2 struggled to render coherent text and to preserve details in complex scenes. The authors linked the detail problem to generating the base image at 64 x 64 before upsampling. These observations are more specific than broad claims that DALL-E 2 "understood" natural language or could accurately write arbitrary text into an image.[3]

## DALL-E 3

### Training with generated captions

The central published contribution behind DALL-E 3 was a change to training captions. OpenAI argued that short, noisy, or inaccurate image descriptions prevent a text-to-image model from learning every relation expressed in a detailed prompt. Its researchers trained an image captioner, fine-tuned versions of it to produce short and highly descriptive synthetic captions, and used those generated descriptions to recaption the image dataset.[4]

The DALL-E 3 paper states that the model was trained on a mixture of 95% synthetic captions and 5% ground-truth captions. The percentage describes the caption mixture, not an accuracy score. The paper did not report "95% text-rendering accuracy." Instead, it explicitly said DALL-E 3's text rendering remained unreliable and could add or omit characters.[4]

OpenAI's system card said DALL-E 3's image-caption training pairs came from a combination of publicly available and licensed sources. It did not publish a complete dataset list, dataset size, model parameter count, or a full training recipe.[13] The research paper likewise says its focus is captioning and prompt following, not complete DALL-E 3 implementation details.[4]

### Disclosed image architecture

The paper's appendix provides a partial architecture description. It says DALL-E 3 used the variational autoencoder from the latent-diffusion work of Rombach and colleagues and that OpenAI trained a convolutional U-Net diffusion decoder on that latent space. The decoder was then reduced to two denoising steps through consistency distillation. The appendix does not disclose a total parameter count, and the main paper cautions that it does not cover the system's full implementation.[4]

This evidence supports describing DALL-E 3 as using a latent representation and a diffusion decoder. It does not support assigning it DALL-E 2's CLIP-prior architecture, claiming that its full architecture was public, or inventing a parameter total.

### ChatGPT prompt synthesis

DALL-E 3 was deployed as an image component inside [ChatGPT](https://aiwiki.ai/wiki/chatgpt). In the launch system, GPT-4 interacted with the user and synthesized the prompt sent to the image model. OpenAI's launch system card also described prompt transformations used for both effectiveness and policy compliance, including removing public-figure names, grounding underspecified descriptions of people, and replacing some branded-object wording with generic descriptions.[13]

This two-model interaction changed [prompting](https://aiwiki.ai/wiki/prompt_engineering) in two ways. A user could iteratively describe and revise an image in conversation, but the text received by DALL-E 3 could differ from the user's original words. The system card acknowledged that automatic transformations can alter meaning, introduce their own biases, and conflict with user preferences.[13]

OpenAI made DALL-E 3 available to ChatGPT Plus and Enterprise users on October 19, 2023.[14] It announced API access at its November 6 DevDay event.[15] Microsoft had already made DALL-E 3 generally available through Bing Chat and Bing Image Creator on October 3, 2023, giving the model a separate consumer distribution channel with Microsoft's own safeguards and Content Credentials implementation.[16]

### Developer-reported evaluation

OpenAI compared DALL-E 3 with DALL-E 2, Stable Diffusion XL 1.0 with its refiner, and Midjourney 5.2. For a developer-created set of 170 captions, raters made pairwise judgments about prompt following and style. A separate sample of 250 MS-COCO captions was used for coherence. The paper reports that DALL-E 3 received higher Elo-style scores in these comparisons, particularly for prompt following.[4]

The result is useful evidence about the authors' target improvement, but its boundaries matter. The prompt set was small and assembled by the model developer, the comparison was not a blinded third-party benchmark covering all uses, and the paper reports relative preferences rather than a general percentage accuracy. The paper released evaluation samples and code, enabling later scrutiny, but it does not justify converting those preferences into a claim such as "95% accurate."

### Limitations

The DALL-E 3 paper says the model still struggled with object placement and spatial language such as "to the left of," "underneath," and "behind." It also reported missing or extra characters in generated words. The captioner sometimes hallucinated fine-grained labels such as plant or bird species, which made the downstream image model unreliable for those specific terms.[4]

These failures distinguish prompt adherence from factual understanding. An image may contain more requested objects or text than a predecessor while still placing them incorrectly, misspelling words, or depicting a falsely labeled species. Generated diagrams, historical scenes, maps, and other claim-bearing visuals therefore require verification rather than being treated as factual illustrations.

## Safety and bias

### DALL-E 2 mitigations

OpenAI described safety as a combination of training-data interventions, model behavior, access controls, prompt and image filters, and monitoring. For DALL-E 2, it filtered graphic sexual and violent images from training data, reweighted the remaining data to address a gender imbalance introduced by filtering, and deduplicated visually similar images to reduce memorized reproduction.[10] These steps changed risks but did not establish that the dataset was unbiased or that memorization was impossible.

The April 2022 preview system card warned that an unguarded system could generate deceptive or harmful content and that DALL-E 2 inherited biases from its training data. It explicitly described that card as an early assessment of the April 2022 preview, not a permanent description of later service policy.[11] In July 2022, OpenAI said it had added a system-level technique that diversified outputs when a prompt described a person without specifying race or gender.[12]

Restrictions also changed during deployment. Early preview controls rejected uploaded realistic faces and attempts to generate public figures, while later products and successor models adopted different rules. Statements about what "DALL-E blocks" must therefore name the model, interface, and date; a launch-era rule should not be presented as a timeless property of the neural network.

### DALL-E 3 mitigations

The DALL-E 3 launch stack included ChatGPT refusals, input classifiers, blocklists, prompt transformations, and output-image classifiers. OpenAI also filtered explicit training images and used internal and external red teaming before launch.[13] These system components surround the image model. A result observed through ChatGPT may therefore differ from direct API behavior or from an unmitigated research model.

OpenAI's system card reported that, before mitigation, underspecified prompts tended disproportionately to produce people who appeared White, female, and young and to take a Western perspective. It used prompt transformations to add demographic detail to some underspecified requests, while acknowledging that such transformations could over-correct or alter a user's intent.[13]

At launch, OpenAI also configured the system to refuse requests for the style of a living artist and maintained a blocklist of living-artist names. It warned that generated concepts, characters, or designs could still implicate third-party copyright or trademarks and that no mitigation could anticipate every output.[13] This was a product policy and risk-control choice, not proof that the model had not learned from artists' works.

### Provenance and deceptive media

In 2024 OpenAI began adding C2PA metadata to images created or edited with DALL-E 3 in ChatGPT and the [OpenAI API](https://aiwiki.ai/wiki/openai_api). C2PA Content Credentials can cryptographically bind statements about an image's provenance to the file.[17] They do not prove that the depicted event occurred.

Metadata is also not durable in every workflow. OpenAI's documentation notes that C2PA information can be removed intentionally or accidentally, including by taking a screenshot. The absence of a credential therefore does not show that an image is human-made, and the presence of a valid OpenAI credential indicates origin rather than factual truth.[18]

DALL-E's ability to produce photorealistic or official-looking scenes creates [deepfake](https://aiwiki.ai/wiki/deepfake) and misinformation risks even when the image contains errors. Conversely, detection based only on visual artifacts becomes less reliable as image models improve. Provenance, moderation, source checking, and context are complementary controls rather than substitutes for one another.

## Independent evaluation

Most headline comparisons for DALL-E came from OpenAI's own release papers and product pages. They measured different things: Frechet Inception Distance, CLIP-based alignment, human preference, or small prompt-following sets. None is a universal measure of correctness. Best-of-N selection, prompt rewriting, model version, content filtering, and the evaluation prompt distribution can all change the observed result.

A 2023 open-access study in *Social Sciences & Humanities Open* tested DALL-E 2 on ten groups of syntactic phenomena, including negation, word order, structural ambiguity, passives, quantification, coordination, ellipsis, and comparatives. The authors reported at-or-near-chance performance on many of the tested mappings from syntax to image meaning.[19] This study evaluated particular prompts and a particular service version; it does not show that every prompt fails, but it supports caution about equating fluent prompt handling with stable grammatical interpretation.

A 2024 peer-reviewed audit in the *Journal of Computer-Mediated Communication* examined 15,300 DALL-E 2 images across 153 occupations. Compared with 2021 United States labor statistics and Google Images, it found that DALL-E 2 risked amplifying occupational gender representation and was more likely to depict women smiling or with a downward head pitch in female-dominated occupations.[20] Automated classifications and a U.S.-occupation benchmark limit the study's cultural scope, but its scale provides direct evidence that neutral occupation prompts could yield patterned representations.

OpenAI's own DALL-E 2 and DALL-E 3 documents reached a compatible high-level conclusion: training data and safety filters can both create or amplify representation problems, and mitigation changes the distribution without eliminating bias.[10][13] Users evaluating a current service should therefore record the model name, date, exact prompt, any prompt rewriting, sample count, and selection method.

## Data, output rights, and copyright

The public record differs by generation. The original DALL-E paper described 250 million internet-collected text-image pairs.[2] The DALL-E 2 paper described the datasets used for its research stack but said the production version used a changed, filtered distribution.[3] The DALL-E 3 system card said its pairs came from public and licensed sources without publishing a full inventory.[13] These disclosures do not support claims that every training image was licensed, nor do they establish the legal status of any particular item.

OpenAI's January 2026 consumer terms say that, as between the user and OpenAI and to the extent permitted by law, the user owns output and OpenAI assigns any right it has in that output. The terms also say outputs may not be unique and require users to ensure that their content does not violate law or other people's rights.[21] Contractual allocation between a service and a user is not the same question as whether an output qualifies for copyright in a jurisdiction.

In January 2025, the U.S. Copyright Office concluded that generative-AI output can receive copyright protection when a human author determines sufficient expressive elements. It identified human-authored material retained in an output, creative arrangement, or creative modification as possible grounds, while saying that prompts alone do not provide sufficient control under the systems it examined.[22] Other countries may apply different law, and the conclusion does not decide whether a generated image infringes an existing work.

Accordingly, statements that DALL-E users receive "full commercial rights" are too broad. Service terms, copyrightability, rights in input material, trademark and publicity rights, and the similarity of an output to protected material are separate issues. The applicable terms and law should be checked for the specific account, territory, and intended use.

## Place in text-to-image research

DALL-E was influential, but it was not the first text-to-image system. Its paper discussed earlier work based on text-conditioned generative adversarial networks, including StackGAN and AttnGAN, as well as methods that optimized images against pretrained vision-language models. The distinctive combination in the first DALL-E was scale, an autoregressive transformer shared across text and image tokens, and CLIP-based selection from many candidates.[1][2]

DALL-E 2 moved the family into the diffusion era while retaining CLIP as an organizing representation. It also differed from text-conditioned diffusion systems such as OpenAI's GLIDE: unCLIP predicted a CLIP image embedding and then decoded it, which encouraged variations that retained semantic and stylistic features. The paper reported a tradeoff, however, because the compact representation could lose exact bindings between an object and its attributes.[3]

DALL-E 3's main published idea was data description rather than a wholly disclosed new architecture. Its captioning experiments argued that a model trained on detailed descriptions learns to heed detailed prompts more reliably. The paper compared DALL-E 3 with [Stable Diffusion](https://aiwiki.ai/wiki/stable_diffusion) XL and [Midjourney](https://aiwiki.ai/wiki/midjourney) 5.2, but those systems differed in training data, interfaces, hidden prompt processing, sampling options, and safety layers.[4] A comparison of screenshots or default outputs is therefore a comparison of complete products, not just their underlying neural networks.

The lineage also illustrates three separate ways to improve text-to-image generation: increase autoregressive model and dataset scale, generate through a learned multimodal latent space with diffusion, and improve the correspondence between images and their training descriptions. Later systems combined or replaced these ideas rather than following one simple scaling curve.

## Transition to GPT Image

OpenAI changed its main image-generation product line after DALL-E 3. On March 25, 2025, it introduced native GPT-4o image generation as the default image generator in ChatGPT and said DALL-E remained accessible through a dedicated GPT. Unlike DALL-E 3's tool-like integration, the new capability was described as native to a multimodal model and could use conversation context and input images directly.[23]

On April 23, 2025, OpenAI released the same generation capability to developers as `gpt-image-1`.[24] GPT Image 1.5 followed on December 16, 2025 with a new ChatGPT Images experience and an API model of that name.[25] On April 21, 2026, OpenAI introduced ChatGPT Images 2.0 and a corresponding GPT Image 2 model, including a mode in which a reasoning model could plan before image generation.[26]

These releases explain the DALL-E models' later catalog status. As of July 28, 2026, OpenAI's model catalog listed GPT Image 2 as the current image-generation model, labeled GPT Image 1.5 and earlier GPT Image models deprecated, and also labeled DALL-E 2 and DALL-E 3 deprecated.[5] DALL-E remains historically important as the earlier family, but current availability, pricing, endpoints, and retirement schedules should be checked in the live developer documentation rather than inferred from old launch posts.

## See also

- [AI-generated content](https://aiwiki.ai/wiki/ai_generated_content)
- [Artificial Intelligence](https://aiwiki.ai/wiki/artificial_intelligence)
- [Latent diffusion model](https://aiwiki.ai/wiki/latent_diffusion)
- [Imagen](https://aiwiki.ai/wiki/imagen)
- [Flux (text-to-image model)](https://aiwiki.ai/wiki/flux)

## References

1. OpenAI. "DALL-E: Creating images from text." January 5, 2021. https://openai.com/index/dall-e/
2. Ramesh, Aditya, et al. "Zero-Shot Text-to-Image Generation." *Proceedings of the 38th International Conference on Machine Learning*, 2021. https://proceedings.mlr.press/v139/ramesh21a/ramesh21a.pdf
3. Ramesh, Aditya, et al. "Hierarchical Text-Conditional Image Generation with CLIP Latents." 2022. https://cdn.openai.com/papers/dall-e-2.pdf
4. Betker, James, et al. "Improving Image Generation with Better Captions." 2023. https://cdn.openai.com/papers/dall-e-3.pdf
5. OpenAI. "All models." OpenAI API documentation. Accessed July 28, 2026. https://developers.openai.com/api/docs/models/all
6. OpenAI. "DALL-E 2." 2022. https://openai.com/index/dall-e-2/
7. OpenAI. "DALL-E: Introducing outpainting." August 31, 2022. https://openai.com/index/dall-e-introducing-outpainting/
8. OpenAI. "DALL-E now available without waitlist." September 28, 2022. https://openai.com/index/dall-e-now-available-without-waitlist/
9. OpenAI. "DALL-E API now available in public beta." November 3, 2022. https://openai.com/index/dall-e-api-now-available-in-public-beta/
10. OpenAI. "DALL-E 2 pre-training mitigations." June 28, 2022. https://openai.com/index/dall-e-2-pre-training-mitigations/
11. OpenAI. "DALL-E 2 Preview - Risks and Limitations." April 11, 2022. https://github.com/openai/dalle-2-preview/blob/main/system-card.md
12. OpenAI. "Reducing bias and improving safety in DALL-E 2." July 18, 2022. https://openai.com/index/reducing-bias-and-improving-safety-in-dall-e-2/
13. OpenAI. "DALL-E 3 System Card." October 2023. https://cdn.openai.com/papers/DALL_E_3_System_Card.pdf
14. OpenAI. "DALL-E 3 is now available in ChatGPT Plus and Enterprise." October 19, 2023. https://openai.com/index/dall-e-3-is-now-available-in-chatgpt-plus-and-enterprise/
15. OpenAI. "New models and developer products announced at DevDay." November 6, 2023. https://openai.com/index/new-models-and-developer-products-announced-at-devday/
16. Microsoft Bing Blogs. "DALL-E 3 now available in Bing Chat and Bing.com/create, for free!" October 3, 2023. https://blogs.bing.com/search/october-2023/DALL-E-3-now-available-in-Bing-Chat-and-Bing-com-create-for-free
17. OpenAI. "Understanding the source of what we see and hear online." May 7, 2024. https://openai.com/index/understanding-the-source-of-what-we-see-and-hear-online/
18. OpenAI Help Center. "C2PA in ChatGPT Images." Accessed July 28, 2026. https://help.openai.com/en/articles/8912793-c2pa-in-dall-e-3
19. Leivada, Evelina, Elliot Murphy, and Gary Marcus. "DALL-E 2 fails to reliably capture common syntactic processes." *Social Sciences & Humanities Open* 8, no. 1 (2023): 100648. https://www.sciencedirect.com/science/article/pii/S259029112300253X
20. Sun, Luhang, et al. "Smiling women pitching down: auditing representational and presentational gender biases in image-generative AI." *Journal of Computer-Mediated Communication* 29, no. 1 (2024): zmad045. https://academic.oup.com/jcmc/article/29/1/zmad045/7596749
21. OpenAI. "Terms of Use." Effective January 1, 2026. https://openai.com/policies/row-terms-of-use/
22. U.S. Copyright Office. "Copyright Office Releases Part 2 of Artificial Intelligence Report." January 29, 2025. https://www.copyright.gov/newsnet/2025/1060.html
23. OpenAI. "Introducing 4o Image Generation." March 25, 2025. https://openai.com/index/introducing-4o-image-generation/
24. OpenAI. "Introducing our latest image generation model in the API." April 23, 2025. https://openai.com/index/image-generation-api/
25. OpenAI. "The new ChatGPT Images is here." December 16, 2025. https://openai.com/index/new-chatgpt-images-is-here/
26. OpenAI. "Introducing ChatGPT Images 2.0." April 21, 2026. https://openai.com/index/introducing-chatgpt-images-2-0/