# AI in Media

> Source: https://aiwiki.ai/wiki/media
> Updated: 2026-07-28
> Categories: Artificial Intelligence, Generative AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "AI in Media." aiwiki.ai, 28 Jul 2026. https://aiwiki.ai/wiki/media
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**AI in media** is the use of [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) to help create, transform, organize, distribute, recommend, moderate, measure, or authenticate media. The term covers many different systems rather than one production method. Examples range from software that turns structured financial data into a short news report, to ranking models that select videos for a feed, to [generative AI](https://aiwiki.ai/wiki/generative_ai) that produces text, images, speech, music, or video. AI may support a human-led workflow, automate a narrowly specified task, or generate material that reaches an audience directly.

The relevant media sectors include journalism, film and television, games, [music](https://aiwiki.ai/wiki/music), [advertising](https://aiwiki.ai/wiki/advertising), and online platforms. These sectors share technical tools but differ in their professional standards, rights arrangements, audience expectations, and regulatory duties. A generated image used for storyboarding, for example, raises different questions from a photorealistic image presented as documentary evidence. For that reason, analysis of AI in media must identify the task, the responsible organization, the source material, the way the output is reviewed, and how the audience encounters it.

## Functions in the media process

AI systems can operate at several stages of a media workflow.

| Stage | Representative uses | Main questions |
| --- | --- | --- |
| Research and acquisition | Transcription, translation, document search, data extraction, and detection of patterns in large collections | Whether sources are complete, lawfully obtained, accurately represented, and protected where necessary |
| Production | Drafting, image and video generation, speech synthesis, editing assistance, visual effects, and automated reports | Authorship, factual accuracy, consent, provenance, working conditions, and rights in training data and outputs |
| Localization and access | Captioning, speech recognition, translation, dubbing, audio description support, and searchable archives | Error rates across languages and speakers, preservation of meaning and performance, and human quality control |
| Distribution and discovery | Search, recommendation, feed ranking, notification, and audience segmentation | Transparency, user control, diversity of exposure, feedback loops, and incentives created by ranking objectives |
| Safety and integrity | Spam filtering, content moderation, duplicate detection, synthetic-media labeling, and provenance checks | Context-sensitive errors, appeals, evasion, interoperability, and the limits of automated detection |
| Monetization and measurement | Ad targeting, bidding, creative variation, attribution, forecasting, and campaign optimization | Privacy, discrimination, deceptive content, disclosure, and the validity of performance measurements |

These functions are often combined. A video platform may transcribe an upload, classify it for safety, translate its captions, rank it for viewers, insert advertising, and attach a synthetic-content label. The models, evidence, and accountability arrangements for those steps need not be the same.

## Development

Automation in media predates modern generative models. Earlier systems commonly relied on rules, templates, metadata, and statistical prediction. In 2014, the Associated Press announced that it would use software from Automated Insights and data from Zacks Investment Research to expand automated corporate earnings coverage from about 300 stories to as many as 4,400 per quarter. The reports were generated from structured data, labeled as automated, and initially checked before the workflow moved to spot checks.[1] A study using the staggered deployment found that the added articles increased trading volume and liquidity for covered firms, while finding no evidence that they changed the speed of price discovery.[2] This example illustrates a narrow form of automation: the subject, input fields, output structure, and publication conditions were specified in advance.

Machine learning also became central to media discovery. A 2016 description of YouTube's recommender architecture separated candidate generation, which reduces a very large video corpus to a manageable set, from ranking, which scores those candidates for a particular request.[3] This two-stage pattern is influential, but a deployed recommendation service is not defined by its model architecture alone. Training data, objectives, product rules, inventory, user controls, and continuous experimentation all affect what an audience sees.

Generative systems broadened the range of outputs that could be synthesized from prompts or other inputs. They can be incorporated into established tools or offered as stand-alone services. Their output is probabilistic, so fluent or realistic material does not by itself establish accuracy, originality, consent, or authenticity. The distinction between assistance and substitution is therefore operational rather than merely technical: it depends on who sets the purpose, verifies the result, accepts responsibility, and can stop publication.

## News and journalism

News organizations use or test AI for transcription, translation, document analysis, data extraction, headline suggestions, search, personalization, and template-based reporting. They may also use generative systems to summarize background material or propose text that a journalist then verifies and rewrites. Current Associated Press guidance permits several of these assistive uses but requires journalistic review, keeps editorial judgment and accountability with journalists, and prohibits generative AI from creating, altering, or enhancing news photography. It also calls for disclosure when AI has played a material role in published work.[4]

The reliability requirement is unusually strict in journalism because an output can make a factual assertion about a person or event. A model summary can omit a qualification, merge two sources, or supply a plausible statement that was absent from the evidence. Translation and transcription can change names, numbers, or attribution. Verification therefore has to return to recordings, documents, direct observations, and attributable sources rather than treating model output as a source. The Paris Charter on AI and Journalism, a voluntary set of principles published in 2023, similarly places editorial responsibility with news organizations and calls for human agency, evaluation, and a clear distinction between authentic and synthetic content.[5]

Audience acceptance varies with the task and the degree of human involvement. A Reuters Institute survey conducted in six countries in 2025 found average comfort levels of 12 percent for news produced entirely by AI, 21 percent for AI production with human oversight, 43 percent for mainly human production with some AI help, and 62 percent for entirely human production. Respondents were more comfortable with spelling and grammar correction or translation than with an artificial presenter or author. These are self-reported responses from Argentina, Denmark, France, Japan, the United Kingdom, and the United States, not a measurement of every audience or of actual behavior.[6]

AI chatbots are also becoming an additional route to news. In the Reuters Institute's 2026 cross-market survey, weekly use of AI chatbots for news rose from 7 to 10 percent, while 1 percent of respondents named a chatbot as their main news source. The report also found substantially higher stated trust among chatbot news users than in the population overall and low self-reported click-through from chatbot answers to publishers. The authors caution that use and trust vary by market and user group.[7] For publishers, this distribution channel creates questions about attribution, referral traffic, corrections, source visibility, and whether a generated answer faithfully represents the underlying report. Detailed coverage appears in [AI in journalism](https://aiwiki.ai/wiki/news).

## Film, television, games, and visual production

AI-assisted visual production can include search across footage, rotoscoping, object removal, cleanup, upscaling, storyboard or concept generation, visual effects, and synthetic background or character elements. Text and image systems may also be used during development to explore alternatives. These uses differ from releasing an unreviewed generated scene: a production can constrain a tool to a specific stage, retain human creative decisions, and document the origin and permission status of its inputs.

Commercial announcements show how studios and tool vendors are exploring customized models. In 2024, Lionsgate and Runway announced an agreement to develop a model customized on Lionsgate's proprietary catalog for pre-production and post-production work. The announcement establishes the scope of the partnership, but it does not by itself demonstrate the quality, cost, labor effect, or final-screen use of any output.[8]

Awards eligibility and copyright ownership apply different tests. For the 98th Academy Awards, the Academy stated that generative AI and other digital tools would neither help nor harm a film's nomination chances, while branches would consider the degree to which a human was at the center of the creative authorship.[9] This rule concerns award consideration, not whether a generated element is copyrightable or contractually permitted.

Collective bargaining has produced more specific workplace rules. Under the Writers Guild of America's 2023 agreement, AI is not a writer and AI-generated material is not treated as literary or source material; a company cannot require a writer to use AI, and it must disclose AI-generated material given to a writer.[10] The guild's 2026 agreement kept those protections and added notice provisions when a company licenses covered literary material to train a commercial generative system.[11] SAG-AFTRA's 2023 television and theatrical agreements address informed consent and compensation for the creation and use of digital replicas, along with notice and bargaining over synthetic performers.[12] These agreements apply to covered work and parties, so they should not be generalized into universal rules for every production or jurisdiction.

Games combine authored assets, software behavior, player input, and continuing updates. AI may be used for development tools, character behavior, localization, moderation, or generated dialogue. Whether a system is appropriate depends on the game's design and on contracts, platform policies, safety testing, and the rights attached to voices, performances, code, and art. Broader treatment of screen media, interactive works, and production practice appears in [AI in entertainment](https://aiwiki.ai/wiki/entertainment).

## Music, speech, and audio

AI in audio includes source separation, noise reduction, mastering assistance, speech recognition, speech synthesis, music recommendation, and systems that generate a composition or recording. A [voice cloning](https://aiwiki.ai/wiki/voice_cloning) system can reproduce features associated with a speaker from reference recordings. The same capability can support authorized localization or accessibility and can also be used for impersonation. The relevant questions include whose voice or recordings were used, what consent covered, whether the output could mislead listeners, and how it is labeled.

Music generation has become a focus of copyright litigation. In 2024, record companies sued Suno and Udio in separate United States federal cases, alleging that copyrighted sound recordings were copied without authorization to train the companies' music-generation systems.[13] Those allegations and the defendants' responses are matters for the courts; the filing of a complaint is not a finding of infringement. The disputes exemplify separate questions that are sometimes collapsed: whether training copies are authorized or protected by an exception, whether an output is substantially similar to protected expression, who owns an eligible output, and whether a performer has consented to a simulated voice or likeness.

California enacted two performer-likeness measures in 2024. One requires covered contracts to specify the use of an AI-generated digital replica of a performer's voice or likeness and requires professional representation in negotiating the contract; the other addresses specified commercial uses of a deceased performer's digital replica without consent from the rights holder.[14] Their precise application depends on the statutory definitions and facts of a use, and other jurisdictions have different publicity, privacy, contract, and unfair-practice rules.

## Localization and accessibility

Speech recognition, machine translation, text-to-speech, and image analysis can make media searchable or help produce captions, translated audio, and descriptions. Automatic dubbing is usually a pipeline rather than a single model. A published system such as VideoDubber combines speech recognition, machine translation, and speech synthesis, then adds duration control so translated speech better matches the source video.[15] Research on 319.57 hours of professionally dubbed material from 54 titles found that translation quality and vocal naturalness were important, while simple character-length or lip-sync constraints did not fully describe professional practice. The researchers also found that emphasis, emotion, and other properties of the source performance matter.[16]

Platform deployments show both the scale and the changing status of these tools. YouTube announced in December 2024 that automatic dubbing was available to hundreds of thousands of channels in its Partner Program that focused on knowledge and information, with a limited set of source and target languages at launch.[17] In February 2026, it announced broader channel access and an "Expressive Speech" feature for eight languages.[18] These announcements describe YouTube's service at those dates; they do not establish that every language pair, speaker, or program receives equivalent quality.

Accessibility gains should be evaluated against error patterns. A 2020 study of five commercial speech-recognition systems reported an average word error rate of 0.35 for Black speakers and 0.19 for white speakers in the studied recordings.[19] The result concerns the systems and speech samples tested at that time, not every present-day service. It nevertheless demonstrates why aggregate accuracy can hide unequal performance. Media organizations need evaluation sets that reflect their languages, accents, acoustic conditions, program genres, and accessibility requirements, along with a correction path for consequential errors.

## Advertising and marketing

AI supports audience modeling, prediction, auction bidding, budget allocation, creative testing, measurement, and generation of copy or images. Real-time bidding systems make rapid decisions about whether and how much to bid for an advertising opportunity, using campaign goals and available context. A technical survey of display advertising describes the interaction among auctions, behavioral targeting, prediction, and optimization in this market.[20]

Generative tools add another layer by producing or modifying creative assets. Google Ads, for example, documents tools that let eligible advertisers generate images and emphasizes that the advertiser remains responsible for reviewing assets and complying with advertising policies and applicable law.[21] This allocation of responsibility is important: vendor safeguards do not verify every factual claim, license, endorsement, or depiction in an advertisement.

Disclosure practices are still developing. Google describes both automatically generated and advertiser-supplied disclosures for some ads containing synthetic or altered content, with presentation varying by region and format.[22] It stated that, beginning July 13, 2026, AI labels could appear on creatives served to publisher properties through Google demand, while its sell-side platforms did not offer labeling features.[23] Such labels can inform viewers but do not resolve whether an ad is deceptive, whether personal data was used lawfully, or whether targeting has discriminatory effects. Detailed treatment of targeting, campaign optimization, and measurement appears in [AI in marketing](https://aiwiki.ai/wiki/marketing).

## Recommendation, discovery, and moderation

[Recommender systems](https://aiwiki.ai/wiki/recommender_system) select and order media from a much larger inventory. Current YouTube documentation lists watch and search histories, subscriptions, likes, dislikes, "not interested" feedback, and satisfaction surveys among its recommendation signals, with different signals used on different surfaces.[24] TikTok likewise describes the For You feed as personalized by interests and engagement and provides controls such as "not interested," feed refresh, keyword filtering, and topic management.[25] These product descriptions can change, and neither implies that a user can observe or control every ranking input.

Claims about recommendation effects require care. A study of news recommenders argues that selective exposure depends on system design and user behavior, including how customization and content diversity are configured.[26] It is therefore misleading to assume that every ranking model necessarily creates the same political or cultural effect. Outcomes can differ with the platform, objective, available content, population, and measurement period.

Regulation can require transparency or alternatives without prescribing one ranking algorithm. Under the European Union's Digital Services Act, very large online platforms and very large online search engines must offer at least one recommender option that is not based on profiling. The Act also requires information about the main parameters of recommender systems and gives users mechanisms to challenge certain moderation decisions.[27]

Content moderation uses related techniques for a different purpose. Systems may compare uploads with known-content hashes or predict whether text, images, audio, or video fit a policy category. Research on algorithmic content moderation emphasizes that scale is not the same as neutrality: category definitions, thresholds, training examples, review resources, and appeals are governance choices, and automated tools can lose context or distribute errors unevenly.[28] Social-platform ranking, moderation, generated influencers, and platform governance are covered further in [Social Media](https://aiwiki.ai/wiki/social_media).

## Synthetic media and authenticity

[Synthetic media](https://aiwiki.ai/wiki/synthetic_media) includes content generated or materially altered by computational systems. A [deepfake](https://aiwiki.ai/wiki/deepfake) is a prominent subset, often involving a realistic depiction of a person saying or doing something that did not occur. The risk is not limited to whether a viewer is fully deceived. An experiment involving a political deepfake found that exposure was more likely to create uncertainty than outright false belief and that this uncertainty could reduce trust in news on social media.[29] Another study found that participants performed at chance when distinguishing the tested AI-synthesized faces from real faces and rated the synthetic faces as more trustworthy on average.[30] These bounded experimental findings should not be read as proof that all synthetic media is undetectable or believed.

Technical responses fall into several categories. Detection estimates whether content has properties associated with a generation or manipulation process. Watermarking embeds a signal intended to survive specified transformations. Provenance records information about origin and editing history. Prevention and access controls seek to limit unauthorized generation or distribution. NIST's overview of synthetic-content risk treats these as complementary approaches and stresses measurement, testing, and auditing rather than a single universal detector.[31]

The Coalition for Content Provenance and Authenticity, or C2PA, publishes a technical standard for cryptographically signed assertions about a digital asset's origin and history.[32] Valid provenance can help show what a participating device or editor asserted and whether the record was altered after signing. It is not a general proof that the depicted event is true, and the absence of credentials is not proof that an asset is synthetic.

Platforms also use audience-facing labels. YouTube's 2026 policy update describes more prominent labels for realistic and meaningful AI alterations, continued creator disclosure, and automatic detection for some content.[33] A label's usefulness depends on coverage, accuracy, placement, comprehensibility, and whether it remains attached when media is copied between services. Detection and labeling should therefore complement source verification and editorial judgment rather than replace them.

## Copyright, authorship, likeness, and consent

Copyright questions differ across countries and across stages of a system. Relevant acts can include acquiring a training copy, adapting or reproducing protected material, generating an output, and distributing that output. The legal analysis may depend on the kind of work, source of the copy, purpose of the use, similarity of the output, market effect, license terms, and applicable exceptions.

The United States Copyright Office's 2025 report on copyrightability states that using AI as an assistive tool does not by itself prevent copyright protection. Protection can cover perceptible human-authored expression, a human selection or arrangement, or sufficiently creative human modification, but not purely AI-generated material; the Office also concluded that prompts alone did not provide sufficient human control under the systems it examined.[34] This is a statement of the Office's approach under United States law, not a worldwide rule.

In its pre-publication report on generative-AI training, the Office concluded that several stages of development can implicate copyright owners' exclusive rights and that fair use requires case-specific analysis. It described research or analytical uses as more likely to qualify than commercial uses that produce expressive substitutes, particularly where copyrighted works were obtained unlawfully, while recommending that licensing markets continue to develop before broad government intervention.[35] Courts decide individual disputes, and the Office noted that its final report could include clarifications without changing the substantive conclusions.

Identity rights are related but distinct. A media output may imitate a person's face, body, or voice without copying a copyrightable work. The Copyright Office's digital-replica report identified gaps in existing United States protections and recommended a federal law covering unauthorized digital replicas of all individuals, subject to safeguards for speech.[36] Contract, publicity, privacy, consumer-protection, election, and criminal laws may also apply, depending on the jurisdiction and use.

The European Union's AI Act establishes transparency duties for certain synthetic content. Article 50 requires providers of relevant generative systems to mark outputs in a machine-readable format and requires deployers to disclose certain deepfakes. It also addresses AI-generated or manipulated text published to inform the public on matters of public interest, with an exception where human review or editorial control and responsibility are present.[37] As of July 28, 2026, these obligations were scheduled to apply on August 2, 2026. European Commission guidance says that artistic, satirical, fictional, and comparable works receive a more limited disclosure treatment and that deployers cannot rely only on invisible metadata when a visible or audible disclosure is required.[38]

## Editorial and operational controls

No single control makes AI-assisted media reliable. Organizations can instead match controls to the consequence of the task. Common measures include:

- defining which uses are allowed, prohibited, or require approval;
- documenting the model, version, settings, input sources, licenses, and material human edits;
- separating brainstorming or low-risk internal assistance from claims intended for publication;
- checking factual outputs against primary evidence and preserving source attribution;
- obtaining appropriate consent for voices, likenesses, performances, and confidential material;
- testing on representative languages, speakers, genres, and failure conditions;
- screening for privacy, security, discrimination, and policy violations;
- providing audience disclosure when synthesis or alteration is material;
- maintaining correction, takedown, appeal, and incident-response processes; and
- reviewing whether a tool remains fit for purpose after model or product changes.

Human review is useful only when the reviewer has enough time, evidence, authority, and expertise to identify an error. A nominal approval step can otherwise become automation bias. The responsible publisher or platform still needs to decide what evidence supports publication, how contributors and subjects are treated, what the audience should be told, and how errors will be corrected.

## See also

- [AI in journalism](https://aiwiki.ai/wiki/news)
- [AI in entertainment](https://aiwiki.ai/wiki/entertainment)
- [Social Media](https://aiwiki.ai/wiki/social_media)
- [AI in marketing](https://aiwiki.ai/wiki/marketing)
- [Recommendation system](https://aiwiki.ai/wiki/recommendation_system)
- [Synthetic media](https://aiwiki.ai/wiki/synthetic_media)

## References

1. Associated Press. "A leap forward in quarterly earnings stories." https://www.ap.org/the-definitive-source/announcements/a-leap-forward-in-quarterly-earnings-stories/
2. Elizabeth Blankespoor, Ed deHaan, and Christina Zhu. "Capital market effects of media synthesis and dissemination: evidence from robo-journalism." Review of Accounting Studies 23, 2018. https://doi.org/10.1007/s11142-017-9422-2
3. Paul Covington, Jay Adams, and Emre Sargin. "Deep Neural Networks for YouTube Recommendations." Proceedings of the 10th ACM Conference on Recommender Systems, 2016. https://doi.org/10.1145/2959100.2959190
4. Associated Press. "AP updates newsroom standards for artificial intelligence." 23 July 2026. https://www.ap.org/the-definitive-source/announcements/ap-updates-newsroom-standards-for-artificial-intelligence/
5. Reporters Without Borders. "Paris Charter on AI and Journalism." 10 November 2023. https://rsf.org/en/rsf-and-16-partners-unveil-paris-charter-ai-and-journalism
6. Richard Fletcher and Rasmus Kleis Nielsen. "Generative AI and News Report 2025: How People Think about AI's Role in Journalism and Society." Reuters Institute for the Study of Journalism, 2025. https://doi.org/10.60625/risj-5bjv-yt69
7. Nic Newman. "Emerging uses of AI chatbots for news, and what it means for journalism." Reuters Institute Digital News Report 2026, 16 June 2026. https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2026/emerging-uses-ai-chatbots-news-and-what-it-means-journalism
8. Runway. "Runway Partners with Lionsgate." 18 September 2024. https://runwayml.com/news/runway-partners-with-lionsgate
9. Academy of Motion Picture Arts and Sciences. "Awards Rules and Campaign Promotional Regulations Approved for 98th Oscars." 21 April 2025. https://press.oscars.org/news/awards-rules-and-campaign-promotional-regulations-approved-98th-oscarsr
10. Writers Guild of America. "Artificial Intelligence." 2023 MBA guidance. https://www.wga.org/contracts/know-your-rights/artificial-intelligence
11. Writers Guild of America. "2026 MBA Contract Changes FAQ." 2026. https://origin.www.wga.org/contracts/contracts/mba/2026-mba-contract-changes-faq
12. SAG-AFTRA. "2023 TV/Theatrical Contracts." 2023. https://www.sagaftra.org/contracts-industry-resources/contracts/2023-tvtheatrical-contracts
13. Recording Industry Association of America. "Record Companies Bring Landmark Cases for Responsible AI Against Suno and Udio." 24 June 2024. https://www.riaa.com/record-companies-bring-landmark-cases-for-responsible-ai-againstsuno-and-udio-in-boston-and-new-york-federal-courts-respectively/
14. Office of Governor Gavin Newsom. "Governor Newsom signs bills to protect digital likeness of performers." State of California, 17 September 2024. https://www.gov.ca.gov/2024/09/17/governor-newsom-signs-bills-to-protect-digital-likeness-of-performers/
15. Yihan Wu et al. "VideoDubber: Machine Translation with Speech-Aware Length Control for Video Dubbing." Proceedings of the AAAI Conference on Artificial Intelligence 37(11), 2023. https://doi.org/10.1609/aaai.v37i11.26613
16. William Brannon, Yogesh Virkar, and Brian Thompson. "Dubbing in Practice: A Large Scale Study of Human Localization With Insights for Automatic Dubbing." Transactions of the Association for Computational Linguistics 11, 2023. https://doi.org/10.1162/tacl_a_00551
17. YouTube. "Break down language barriers with auto dubbing on YouTube." 10 December 2024. https://blog.youtube/news-and-events/auto-dubbing-on-youtube/
18. YouTube. "Unlocking a global audience with auto dubbing." 4 February 2026. https://blog.youtube/news-and-events/youtube-auto-dubbing-expressive-speech/
19. Allison Koenecke et al. "Racial disparities in automated speech recognition." Proceedings of the National Academy of Sciences 117(14), 2020. https://doi.org/10.1073/pnas.1915768117
20. Jun Wang, Weinan Zhang, and Shuai Yuan. "Display Advertising with Real-Time Bidding (RTB) and Behavioural Targeting." Foundations and Trends in Information Retrieval 11(4-5), 2017. https://doi.org/10.1561/1500000049
21. Google Ads Help. "Create images with generative AI." Accessed 28 July 2026. https://support.google.com/google-ads/answer/14150986?hl=en-gb
22. Google My Ad Center Help. "AI ad transparency." Accessed 28 July 2026. https://support.google.com/My-Ad-Center-Help/answer/17196133?hl=en
23. Google AdSense Help. "AI labeling regulations." 13 July 2026. https://support.google.com/adsense/answer/17258538?hl=en
24. YouTube Help. "How YouTube recommendations work." Accessed 28 July 2026. https://support.google.com/youtube/answer/16089387?hl=en
25. TikTok Help Center. "For You." Accessed 28 July 2026. https://support.tiktok.com/en/getting-started/for-you/test-for-you
26. Erik Knudsen. "Modeling news recommender systems' conditional effects on selective exposure." Journal of Communication 73(2), 2023. https://doi.org/10.1093/joc/jqac047
27. European Union. Regulation (EU) 2022/2065, Digital Services Act, Articles 27 and 38. 19 October 2022. https://eur-lex.europa.eu/eli/reg/2022/2065/oj
28. Robert Gorwa, Reuben Binns, and Christian Katzenbach. "Algorithmic content moderation: Technical and political challenges in the automation of platform governance." Big Data & Society 7(1), 2020. https://doi.org/10.1177/2053951719897945
29. Cristian Vaccari and Andrew Chadwick. "Deepfakes and Disinformation: Exploring the Impact of Synthetic Political Video on Deception, Uncertainty, and Trust in News." Social Media + Society 6(1), 2020. https://doi.org/10.1177/2056305120903408
30. Sophie J. Nightingale and Hany Farid. "AI-synthesized faces are indistinguishable from real faces and more trustworthy." Proceedings of the National Academy of Sciences 119(8), 2022. https://doi.org/10.1073/pnas.2120481119
31. 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, 2024, updated 8 April 2026. https://doi.org/10.6028/NIST.AI.100-4
32. Coalition for Content Provenance and Authenticity. "C2PA Specifications." Accessed 28 July 2026. https://spec.c2pa.org/specifications/
33. YouTube. "Improving AI labels on YouTube to help viewers better understand how content was made." 27 May 2026. https://blog.youtube/news-and-events/improving-ai-labels-viewers-creators/
34. United States Copyright Office. "Copyright and Artificial Intelligence, Part 2: Copyrightability." 29 January 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf
35. United States Copyright Office. "Copyright and Artificial Intelligence, Part 3: Generative AI Training." Pre-publication version, May 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf
36. United States Copyright Office. "Copyright and Artificial Intelligence, Part 1: Digital Replicas." July 2024. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-1-Digital-Replicas-Report.pdf
37. European Union. Regulation (EU) 2024/1689, Artificial Intelligence Act, Article 50. 13 June 2024. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
38. European Commission. "Transparency obligations under Article 50 of the AI Act." 24 July 2026. https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act