Digital Media
Last reviewed
Sources
40 citations
Review status
Source-backed
Revision
v3 ยท 6,671 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
40 citations
Review status
Source-backed
Revision
v3 ยท 6,671 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Digital Media ChatGPT Plugins
Artificial intelligence in digital media covers a sprawling set of practices that touch nearly every part of how news, video, audio, and online content are produced, distributed, moderated, and consumed. The category includes automated news writing systems that have run inside wire services since the mid 2010s, generative tools such as ChatGPT, Sora, and Veo that produce text, image, video, and audio outputs, recommender system pipelines that decide which videos surface on YouTube or TikTok, and content moderation classifiers that flag posts on most large platforms.
Since 2023, the field has also been defined by a wave of high profile lawsuits and licensing deals between AI labs and news publishers, most prominently The New York Times Company v. Microsoft Corporation et al., filed in December 2023 in the Southern District of New York. By late 2024 the news watchdog NewsGuard had identified more than 1,100 sites that publish AI generated content with little or no human oversight, and by early 2026 that count exceeded 3,000. Regulators in the European Union, China, and several US states have begun to require disclosure of synthetic content, and content provenance work led by the Coalition for Content Provenance and Authenticity (C2PA) has produced a cryptographic signing standard that is now being deployed by camera makers, AI labs, and social platforms.
Digital media is used here in the broad commercial sense that publishers, agencies, and regulators use it: text journalism on the web, online video and audio, social media platforms, and the recommendation, search, and moderation systems that sit between content and audiences. AI shows up in this stack in four main ways.
First, AI helps produce content. That can be straight automated writing, as in the Associated Press quarterly earnings reports, or it can be assistive generation in tools such as Forbes' Bertie, Bloomberg's Cyborg, or general purpose chatbots used by reporters and editors.
Second, AI decides what people see. Recommender systems at YouTube, TikTok, Netflix, and Spotify use deep learning to rank candidate items against models of each user's behavior.
Third, AI filters out unwanted content. Classifiers such as Google's Perspective API and Meta's Llama Guard score posts, comments, and chat messages for toxicity, sexual content, hate speech, or policy violations.
Fourth, AI has become a subject for digital media to cover and a source of legal and commercial conflict. The same AI systems that publishers use as tools also scrape, summarize, and sometimes reproduce their content, which has driven both new licensing markets and copyright litigation.
Automated story writing in newsrooms predates the generative AI wave by about a decade. In June 2014 the Associated Press announced a partnership with the natural language generation firm Automated Insights to produce most of its US corporate quarterly earnings stories using a system called Wordsmith. AP went from publishing roughly 300 earnings stories per quarter to about 3,000 in the last three months of 2014, an order of magnitude increase produced without adding staff. Automated Insights had earlier signed deals with Yahoo for fantasy football recaps and with the NFL, Edmunds, and Allstate.
Other wire services and major papers followed. The Washington Post launched Heliograf in 2016 to cover the Rio Olympics and the US election, generating short bulletins from structured data. The same system was used during the 2017 high school sports season in the Washington area. Bloomberg had begun deploying its own Cyborg automation around the same time, and the company has said roughly a third of its published content is automated in some way, primarily structured financial filings and earnings stories. Reuters introduced Lynx Insight in 2018, a tool that scanned market data for anomalies and suggested story ideas. Forbes launched the Bertie content management system in 2018, designed to recommend topics, headlines, and related links to its contributor network. The BBC's research and development unit experimented with semi automated story generation through projects such as Juicer and the Salco prototype.
The November 2022 release of ChatGPT by OpenAI reset the conversation about AI in newsrooms. The earlier wave of automation systems were template based and limited to data heavy beats. Large language models, by contrast, could draft prose on any topic, summarize documents, suggest headlines, and translate copy across languages. Within months, several publishers ran public experiments that ended badly, and several others quietly built internal tools that they did not advertise. By 2024 most major publishers had at least pilot programs that used generative models for translation, transcription, search, and summarization, and many had drafted internal AI usage policies.
The legal landscape also shifted starting in 2023. Authors and publishers began filing copyright suits, and a parallel market emerged in which AI companies licensed news archives for training and for use in chatbot answers. Several of these deals are detailed below.
The table below lists representative uses of AI inside newsrooms. It is not exhaustive; most large publishers have rolled out additional internal projects since 2023 that are not publicly named.
| Outlet | AI use | Year |
|---|---|---|
| Associated Press | Wordsmith generated quarterly earnings stories (Automated Insights) | 2014 |
| Washington Post | Heliograf, automated bulletins for Olympics, elections, prep sports | 2016 |
| Bloomberg | Cyborg, automated earnings and filings coverage | 2016 |
| Reuters | Lynx Insight, anomaly detection and story suggestion | 2018 |
| Forbes | Bertie, AI assisted content management system for contributors | 2018 |
| BBC R&D | Juicer, Salco, semi automated local story generation | 2018 |
| The Guardian | Reporters Lab and ReporterMate experiments for data driven stories | 2019 |
| Wall Street Journal | Internal chatbot pilot, AI assisted research tool for newsroom | 2024 |
| New York Times | Internal generative tools for headlines, summaries, and translations | 2024 |
| Le Monde | AI translations into English under the OpenAI partnership | 2024 |
| Financial Times | Internal generative AI assistant Ask FT, trained on FT archive | 2024 |
| The Washington Post | Climate Answers and Ask the Post AI tools | 2024 |
| Time | Person of the Year AI chatbot, generated answers about archived issues | 2024 |
| LA Times | Insights tool, generates political bias rating and alternate views | 2025 |
Industry surveys from the Reuters Institute for the Study of Journalism and the World Association of News Publishers in 2024 found that roughly three quarters of news leaders were experimenting with generative AI, mostly for back office tasks such as transcription and translation rather than for direct text generation in published articles. The Associated Press, BBC, Bloomberg, Guardian, Reuters, Washington Post, and Wall Street Journal have all published internal AI use policies that prohibit using generative tools to produce final published copy without explicit editorial review and disclosure.
Starting in mid 2023, large publishers began signing licensing deals with OpenAI, Google, Microsoft, and Perplexity covering use of news archives for training and current articles for chatbot responses. The first publicly announced deal between a US news organization and an AI company was an Associated Press agreement with OpenAI in July 2023, which gave OpenAI access to AP archives going back to 1985 in exchange for access to OpenAI technology and expertise.
The second landmark deal came on 13 December 2023, when Axel Springer, the German publisher that owns Bild, Welt, Politico, and Business Insider, signed a multi year partnership with OpenAI. ChatGPT could now include summaries of Axel Springer content, with links and attribution, in responses to user queries. Trade press reports put the deal at roughly thirteen million dollars per year for three years.
The table below lists the major publicly announced deals between AI companies and news publishers.
| AI company | Publisher | Date announced | Notes |
|---|---|---|---|
| OpenAI | Associated Press | July 2023 | Two year deal for archive access |
| OpenAI | Axel Springer | December 2023 | Bild, Welt, Politico, Business Insider |
| OpenAI | Le Monde | March 2024 | French content, translations |
| OpenAI | Prisa Media | March 2024 | El Pais, Cinco Dias, As, Cadena Ser |
| OpenAI | Financial Times | April 2024 | Reported five to ten million per year |
| OpenAI | Dotdash Meredith | May 2024 | InStyle, People, Better Homes and Gardens |
| OpenAI | News Corp | May 2024 | Reported up to 250 million over five years |
| OpenAI | Vox Media | May 2024 | Vox, The Verge, Eater, New York Magazine |
| OpenAI | The Atlantic | May 2024 | Five year deal with product integration |
| OpenAI | Time | June 2024 | 101 years of archive access |
| OpenAI | PMC | June 2024 | Variety, Rolling Stone, Billboard |
| OpenAI | Hearst | October 2024 | Houston Chronicle, San Francisco Chronicle, Esquire |
| OpenAI | Conde Nast | August 2024 | Vogue, New Yorker, Wired, GQ |
| OpenAI | May 2024 | Real time forum data access | |
| News Corp | February 2024 | Reported five to six million per year | |
| Associated Press | January 2024 | Gemini news feed access | |
| February 2024 | Reported 60 million per year for API | ||
| Microsoft | Semafor | February 2024 | Funded Signals news service |
| Perplexity | Time, Fortune, Texas Tribune, Der Spiegel, Entrepreneur | July 2024 | Revenue share Publishers Program |
| Meta | Reuters | October 2024 | Multi year content for Meta AI |
Financial terms in most cases were not disclosed publicly. The News Corp deal in May 2024 reported by trade press at up to 250 million over five years was the largest single publisher licensing arrangement reported through 2024. The deal covers The Wall Street Journal, Barron's, MarketWatch, the New York Post, the UK Times and Sun, and News Corp's Australian mastheads. The next month, OpenAI signed deals with Vox Media and The Atlantic that paired licensing with product development cooperation. A short list of holdouts, including the New York Times, Daily News, and the Center for Investigative Reporting, took the litigation route instead.
In July 2024 Perplexity took a different approach with its Publishers Program, which committed a percentage of advertising revenue back to participating publishers when their articles were cited in answers. Time, Fortune, Entrepreneur, the Texas Tribune, and Der Spiegel were the launch partners. The move followed widely reported plagiarism accusations earlier that year, including a Wired investigation alleging that Perplexity had ignored robots.txt rules to scrape sites it had not licensed.
The most consequential lawsuit in this area is The New York Times Company v. Microsoft Corporation et al., case number 1:23-cv-11195 in the Southern District of New York, filed 27 December 2023. The complaint alleges direct copyright infringement, contributory infringement, vicarious infringement, common law unfair competition, trademark dilution, and Digital Millennium Copyright Act violations. The Times included one hundred exhibits showing GPT-4 producing near verbatim excerpts of Times articles in response to specific prompts. The Times has asked for billions of dollars in damages and an order to destroy the offending model weights and training datasets.
The case has been assigned to Judge Sidney Stein. On 4 April 2025, Judge Stein issued a long awaited ruling on the defendants' motions to dismiss. He denied the bulk of the motions, including OpenAI's argument that infringement claims for conduct more than three years before filing were time barred, and Microsoft's request to dismiss the contributory infringement and DMCA claims. The Daily News trademark dilution claims also survived. The court did dismiss the common law unfair competition by misappropriation claims and certain abridgment claims with prejudice. The substantive copyright and DMCA claims now proceed to discovery, and the case is widely seen as the most important AI copyright suit yet filed.
Other newsroom suits followed.
Raw Story Media, Inc. and AlterNet Media v. OpenAI, Inc. was filed in the Southern District of New York on 28 February 2024 by The Intercept's former parent and the progressive site AlterNet. Unlike the Times case, the complaint focused narrowly on alleged violations of section 1202(b) of the DMCA, which prohibits removal of copyright management information. In November 2024, Judge Colleen McMahon dismissed the case for lack of concrete injury, but gave the plaintiffs leave to amend. The Intercept's separate suit was similarly dismissed.
In April 2024 the Center for Investigative Reporting, publisher of Mother Jones and Reveal, filed its own copyright suit against OpenAI and Microsoft. On 30 April 2024 the Daily News and seven other Alden Global Capital owned papers, including the Chicago Tribune, Denver Post, Orlando Sentinel, Sun Sentinel, San Jose Mercury News, and the Orange County Register, filed a similar suit. Both were consolidated with the Times case for pretrial purposes.
Authors and book publishers have also sued. The Authors Guild, joined by named plaintiffs including John Grisham, George R. R. Martin, Jodi Picoult, Jonathan Franzen, David Baldacci, Scott Turow, and George Saunders, filed a class action against OpenAI on 19 September 2023 in the Southern District of New York. The complaint alleged that OpenAI used pirated copies of plaintiffs' books, downloaded from shadow libraries such as Library Genesis and Z Library, to train GPT-3.5 and GPT-4. The case was assigned docket number 1:23-cv-08292.
Non news plaintiffs have joined the litigation pipeline as well. Getty Images sued Stability AI in January 2023 in the United Kingdom and the District of Delaware. Universal Music Group, ABKCO, and Concord sued Anthropic in October 2023 over song lyrics. Sarah Silverman, Christopher Golden, and Richard Kadrey sued OpenAI and Meta in July 2023 over book training data. A long running consolidated visual artists' suit, Andersen et al. v. Stability AI Ltd. et al., has been working through the Northern District of California since 2023.
A cluster of high profile incidents from 2023 and 2024 illustrates what happens when generative AI is deployed in newsrooms without sufficient editorial oversight. These cases helped harden many publishers' policies and shaped public skepticism about AI in journalism.
In January 2023 Futurism reported that the technology site CNET had quietly published 77 articles, mostly personal finance explainers, under the byline CNET Money Staff using an internally developed AI system. Readers could see that the articles were AI assisted only by hovering over the byline. The investigation found that the AI had borrowed structural and phrasing patterns from previously published articles on other sites in ways that several plagiarism experts described as plagiarism. A separate investigation found factual errors in more than half the published pieces. One article about compound interest incorrectly stated that a 10,000 dollar deposit at 3 percent would yield 10,300 dollars in interest after a year, when the correct figure is 300 dollars.
CNET paused the program in late January 2023 and ran corrections on 41 of the 77 stories. The Wikipedia community downgraded CNET to generally unreliable for the period in question. The damage to the brand continued through 2023 as parent Red Ventures explored a sale.
On 27 November 2023 Futurism published an investigation alleging that Sports Illustrated had been publishing product reviews authored by fake AI generated writers. The investigation focused on Drew Ortiz, whose author photograph appeared on a stock site selling AI generated headshots, and whose biography described improbably specific weekend hobbies. After Futurism reached out for comment, the bylines disappeared from the site without explanation. SI's parent company The Arena Group attributed the content to a third party contractor, AdVon Commerce, and said writers had used pseudonyms to protect their privacy. The SI editorial union and the NewsGuild publicly demanded answers, and Arena's CEO Ross Levinsohn was later fired. The Arena Group lost its license to publish Sports Illustrated from Authentic Brands Group in early 2024.
In late August 2023 Gannett, the largest US newspaper chain, paused an experiment with the AI vendor LedeAI that had been generating high school sports recaps for the Columbus Dispatch and other titles, including the Louisville Courier Journal, AZ Central, Florida Today, and the Milwaukee Journal Sentinel. The recaps went viral for clearly templated prose, including the phrase "close encounter of the athletic kind," missing team names, repeated dates, and placeholder text such as "WINNING TEAM MASCOT." Gannett pulled the program from all of its local markets within days.
Microsoft's MSN and Microsoft Start news aggregation portal had a string of AI related embarrassments in 2023. In August 2023 a travel guide written using AI tools recommended the Ottawa Food Bank as a tourist attraction, advising readers to visit "on an empty stomach." In September 2023, a Microsoft hosted republication of a TMZ Sports obituary for the former NBA player Brandon Hunter, who died at 42, ran under the headline "Brandon Hunter useless at 42" and described him as having "performed in 67 video games over two seasons." In October 2023 an AI driven poll alongside a Guardian story about a young woman's death asked readers to vote on whether the cause was murder, accident, or suicide. Microsoft retracted the items and Guardian editor in chief Katharine Viner sent a public letter of complaint to Microsoft president Brad Smith.
In March 2025 the Los Angeles Times launched Insights, a tool built on its in house Graphene system that attaches an AI generated political rating and a list of alternate viewpoints to opinion pieces. Within hours of launch, the tool generated counterpoints to a column by Gustavo Arellano about the Ku Klux Klan's history in Orange County. The alternate view, presented in the Times's voice, described the Klan as a response by "white Protestant culture" to demographic change rather than a hate group. The LA Times Guild publicly criticized the tool, and the Klan related output was pulled. Owner Patrick Soon-Shiong, who had introduced the tool over the objections of the editorial board, defended it as a transparency feature.
The ratings firm NewsGuard launched an AI tracking center in May 2023 to catalogue sites that publish AI generated content with minimal or no human editorial review. NewsGuard refers to these as Unreliable AI Generated News and Information sites, or UAINs. Inclusion criteria include clear evidence that the bulk of the content is AI generated, an absence of meaningful human oversight, and non disclosure of the AI authorship.
The number tracked has grown sharply. NewsGuard reported 49 UAIN sites in May 2023, more than 700 in February 2024, 1,121 in November 2024, and more than 3,000 across 16 languages by March 2026. Languages now covered include Arabic, Chinese, Czech, Dutch, English, French, German, Indonesian, Italian, Korean, Portuguese, Russian, Spanish, Tagalog, Thai, and Turkish. The catalogue includes sites with generic names such as iBusiness Day, Ireland Top News, and Daily Time Update that mix repackaged wire copy with low quality AI generated articles, often filled with characteristic chatbot language such as "as of my last update." A subset of these sites is funded through Google AdSense and similar programmatic advertising networks, a fact that researchers at NewsGuard, ProPublica, and the Reuters Institute have flagged as a structural problem.
NewsGuard, the Stanford Internet Observatory, and Graphika have also tracked state aligned operations using AI to generate disinformation. The Russian Pravda network, sometimes called Portal Kombat, produces millions of articles annually using generative models, with the apparent aim of seeding LLM training data with pro Kremlin claims. NewsGuard found in early 2025 that several leading chatbots, including ChatGPT, repeated the network's false claims about Ukraine in roughly a third of test prompts. Doppelganger, another Russian influence operation tracked since 2022, has used AI generated images and text to impersonate Western news outlets.
Digital media recommendation is one of the oldest and largest applied uses of machine learning. The recommender system on a large platform decides which video, song, post, or article appears next, and changes to these systems can change the consumption habits of hundreds of millions of users overnight.
Netflix ran Cinematch from the early 2000s and famously offered the Netflix Prize in October 2006, a one million dollar prize for any team that could improve Cinematch root mean squared error by ten percent. The competition ended in September 2009 when BellKor's Pragmatic Chaos crossed the threshold. Netflix has since moved away from the user item matrix factorization approach that dominated the Prize era toward deep learning models that combine viewing history, time of day, device, country, and a learned representation of each title.
YouTube's recommender, which selects videos for the home feed and the "Up next" sidebar, was redesigned around deep neural networks in the mid 2010s. Paul Covington, Jay Adams, and Emre Sargin published the canonical paper, "Deep Neural Networks for YouTube Recommendations," at the ACM Conference on Recommender Systems (RecSys) in September 2016. The paper described a two stage system: a candidate generation network that narrows the entire YouTube corpus down to hundreds of candidates per user, and a ranking network that scores each candidate. The candidate generation tower learns user embeddings and video embeddings in a shared space, an approach later named the two tower model. YouTube has since iterated heavily on this architecture, including reinforcement learning components to optimize for long term user satisfaction rather than immediate clicks.
TikTok operates the most influential recommender system in current digital media. The For You feed is generated by a recommendation pipeline that ranks short videos by a learned proxy for engagement. In December 2021 The New York Times published an internal document titled "TikTok Algo 101" that had been leaked from inside the company. The document explained, in plain language for non technical staff, that the algorithm scores videos on a combination of likes, comments, playtime, and watch completion, and that it optimizes for retention and daily active use. The leak also confirmed that the system uses content embeddings and user embeddings rather than collaborative filtering, allowing rapid cold start on new accounts and videos.
Spotify introduced Discover Weekly in July 2015, a personalized 30 song playlist generated for each user every Monday. The playlist combined collaborative filtering, natural language processing on artist descriptions and music journalism, and audio model features. Over the next decade Spotify rolled out related playlists including Release Radar, Daily Mix, and the time of day daylist. In February 2023 Spotify launched AI DJ, a feature that uses generative voice synthesis (originally based on technology from Sonantic, which Spotify acquired in 2022) to introduce songs in a curated mix with brief spoken commentary. In April 2024 Spotify added AI Playlist, a feature that lets users type natural language prompts such as "songs for a rainy Sunday morning" and generates a playlist that matches. Discover Weekly passed 100 billion total streams by 2025.
Academic and regulatory work on recommender systems has focused on three related concerns: filter bubbles, runaway engagement optimization (particularly on YouTube, where research by Guillaume Chaslot and others in the late 2010s argued the system favored extreme content), and effects on adolescent mental health. The 2024 Surgeon General's advisory and several state Attorney General investigations of TikTok and Meta drew explicitly on internal documents about recommender behavior. The EU Digital Services Act, in force since 2024, requires very large online platforms to allow users to opt out of profiling based recommendations.
Large social platforms rely on AI to triage user content for human review and to take immediate action against the highest confidence violations.
Google's Jigsaw subsidiary released Perspective API in 2017 as a free tool that scores text for toxicity, identity attack, insult, profanity, and threat. Early customers included The New York Times, the Guardian's comments section, and Wikipedia's editing tools. Perspective is now used by more than 1,000 partners and processed roughly 500 million API calls per day by 2021, according to Jigsaw figures. Academic work has documented that the original models had higher false positive rates for African American English and for posts mentioning protected identities, and Jigsaw has issued several model updates intended to reduce these gaps.
A cluster of dedicated content moderation vendors emerged in the late 2010s and early 2020s. Hive offers a moderation classifier that detects more than forty content categories, including sexual content, violence, hateful symbols, and a growing list of AI generated content classes. Sightengine focuses on visual moderation with finer grained categories. Spectrum Labs, acquired by the dating app maker Bumble in March 2024, built behavior based moderation focused on dating and social platforms. Modulate ToxMod analyzes voice in real time multiplayer games and is used by Call of Duty publisher Activision, by Riot, and by Microsoft's Xbox platform.
The major AI labs also offer moderation tools. OpenAI's Moderation API, available since 2022 and updated with a multimodal version in 2024, classifies text and images against an OpenAI specific policy taxonomy. Microsoft's Azure AI Content Safety, launched in 2023, scores text and images on hate, sexual, violence, and self harm dimensions and is integrated with Azure OpenAI. Meta released Llama Guard in December 2023 as an open weights LLM specifically fine tuned for input output moderation; later versions Llama Guard 2 and 3 were released through 2024. Anthropic offers content classification through Claude's API, and Google's Gemini has integrated safety filters that customers can configure.
At the platform level, Meta has said that the vast majority of hate speech, nudity, and graphic violence taken down on Facebook and Instagram is removed by automated systems before any user reports it. YouTube, TikTok, and X all maintain similar automated front ends, supplemented by tens of thousands of contracted human reviewers globally.
Generative video has progressed rapidly since early 2024. OpenAI previewed Sora on 15 February 2024 with a release of short clips up to one minute long that quickly went viral. The system was released to ChatGPT Plus and Pro subscribers on 9 December 2024 as Sora Turbo, with Plus users limited to 480p output and Pro users at 1080p. Google's Veo followed: Veo 2 launched in December 2024 with 4K output and Veo 3 launched in May 2025 with synchronized native audio generation. Runway released Gen-3 Alpha in mid 2024. Chinese companies Kuaishou, with Kling, and ByteDance, with Dreamina and Seaweed, released competitive systems by late 2024.
For corporate and explainer video, Synthesia, founded in London in 2017, builds AI avatars from a short recording of a real person and synthesizes speech and lip movements in more than 140 languages. HeyGen, founded in Shenzhen and Los Angeles in 2020, offers similar avatar based video generation and gained particular attention in late 2023 for a viral lip synced translation feature.
Generative audio has produced its own wave of products. Suno and Udio generate full vocal tracks from short text prompts and are covered in detail at music. ElevenLabs and OpenAI's Voice Engine produce voice clones from short audio samples. Descript and Adobe Podcast offer AI driven editing tools that let editors rewrite a podcast transcript and have the audio re generated to match. Major labels Universal, Sony, and Warner sued Suno and Udio in June 2024 for using copyrighted recordings to train their models. Hollywood unions SAG-AFTRA, the Writers Guild of America, and the Directors Guild secured AI specific contract language in their 2023 negotiations, including requirements for consent and compensation when synthetic voices or likenesses are used.
A running list of incidents has helped move synthetic media from a research curiosity to a regulated category. The deepfake article covers the technical lineage. The selection below focuses on incidents that directly affected digital media norms and policy.
On 21 January 2024 thousands of New Hampshire voters received a robocall in which a synthetic voice impersonating President Joe Biden urged them to "save your vote" and not participate in the state's first in the nation primary two days later. New Hampshire Attorney General John Formella announced an investigation the next day. The political consultant Steve Kramer admitted to commissioning the call and was indicted on 13 felony counts of voter suppression and 13 misdemeanor counts of candidate impersonation under New Hampshire law in May 2024. The Federal Communications Commission proposed a six million dollar fine. The incident was widely reported as the first nationally consequential use of a generative AI voice clone in US politics.
In the last week of January 2024 sexually explicit deepfake images of Taylor Swift spread rapidly on X. A single post was viewed more than 47 million times before the account was suspended. X temporarily disabled searches for "Taylor Swift" on 27 January 2024 and restored them two days later. The case prompted statements from White House Press Secretary Karine Jean-Pierre, Microsoft CEO Satya Nadella, and a host of US senators. The DEFIANCE Act, a federal civil remedy for victims of non consensual intimate imagery created by AI, passed the Senate unanimously in July 2024 but was not enacted into law in that Congress. The TAKE IT DOWN Act was signed into law in May 2025.
Throughout 2024 and 2025 fraudulent advertisements using AI voice and face clones of celebrities including Joe Rogan, Kelly Clarkson, Tom Hanks, Oprah Winfrey, and Taylor Swift circulated widely on Facebook, Instagram, YouTube, and TikTok. The ads typically promoted dietary supplements, dental clinics, or financial scams. Hanks publicly disowned a Hanks lookalike dental ad in October 2023; Clarkson disowned a Clarkson lookalike weight loss product the same month. By 2025 ad transparency tools from Meta and Google had begun to flag and remove the campaigns more aggressively, but enforcement has remained imperfect.
State aligned and politically motivated synthetic media has shown up in elections including Slovakia in September 2023, where a fake audio of opposition leader Michal Simecka allegedly discussing election fraud spread shortly before the vote, Indonesia in February 2024, where a digital revival of late dictator Suharto appeared in campaign material, and Argentina in 2023, where both major candidates used generative imagery in attack ads.
Industry and government efforts to label synthetic media have converged on two strands: cryptographic provenance, led by C2PA, and regulatory labeling regimes, led by the European Union and China.
The Coalition for Content Provenance and Authenticity (C2PA) is a Joint Development Foundation project founded in February 2021 by Adobe, Arm, the BBC, Intel, Microsoft, and Truepic. It merged the earlier Content Authenticity Initiative led by Adobe with Microsoft and the BBC's Project Origin. C2PA defines Content Credentials, a manifest format that cryptographically signs the source, edit history, and AI involvement for an image, audio file, or video. The 2.0 specification was published in January 2024 and the standard was submitted to the International Organization for Standardization. Through 2024 Google, OpenAI, Meta, and Amazon joined the C2PA steering committee. Camera makers Leica, Nikon, and Sony added native Content Credentials capture to flagship cameras. Adobe Photoshop, Adobe Firefly, OpenAI's DALL-E 3 and Sora, and Microsoft Designer write Content Credentials by default. The standard does not prevent stripping or forgery of metadata, but it gives a verifiable chain when one is present.
The EU Artificial Intelligence Act, formally adopted in May 2024 and entering into force in stages through 2026, requires labeling of synthetic content under Article 50. Providers of generative AI systems must mark outputs as artificially generated in a machine readable format, and deployers must disclose to users when content has been generated or manipulated by AI, including deepfakes and AI generated news on matters of public interest. The Article 50 obligations begin to apply in August 2026. The European Commission published a draft Code of Practice on Transparency of AI Generated Content in December 2025 with the final version expected in mid 2026.
China's Cyberspace Administration deep synthesis regulations took effect 10 January 2023, well before any comparable Western rule. The regulations require providers of deepfake and synthetic media services to obtain consent from subjects, label synthetic content visibly or with metadata, and verify user identities. A parallel generative AI regulation took effect in August 2023, requiring filtering and content moderation.
In the United States, federal legislation has lagged. California Governor Gavin Newsom signed AB 2655, the Defending Democracy from Deepfake Deception Act of 2024, in September 2024. The law required large platforms to block materially deceptive election content during specified pre election windows. In October 2024 Judge John Mendez of the Eastern District of California enjoined the law in Kohls v. Bonta, finding it preempted by Section 230 of the Communications Decency Act and likely unconstitutional under the First Amendment; the court issued a final order striking down the law in 2025. Texas, Minnesota, and several other states have enacted narrower deepfake election laws, mostly focused on disclosure and civil remedies.
The FTC issued a final rule in February 2024 prohibiting AI generated impersonation of government and businesses, and proposed extending it to impersonation of individuals. Federal communications regulators ruled in February 2024 that AI generated voices in robocalls fall under the Telephone Consumer Protection Act, requiring prior consent.
AI in digital media sits at the intersection of intellectual property law, labor economics, and the social politics of speech. The disputes are continuous, but a few themes recur.
Most of the news industry believes the AI training pipeline has consumed enormous amounts of copyrighted reporting without compensation. The licensing deals struck since mid 2023 are an attempt by some publishers to monetize that use; the lawsuits filed in the same period are an attempt by others to prevent it. The trade off is sharper than usual because the products that depend on news training data, including ChatGPT and Google's Gemini, also compete with publishers for the same reader attention. A working paper from researchers at the Tow Center for Digital Journalism in late 2024 documented persistent referral traffic declines from Google search to news sites as AI Overviews expanded.
A second strand of the debate concerns labor. Newsrooms have shed staff since the 2008 financial crisis, and generative tools raise both productivity and substitution. Industry surveys in 2024 reported that journalists overwhelmingly worry about AI replacing entry level roles in copy editing, summarization, translation, and SEO writing.
A third strand is trust. The CNET, Sports Illustrated, Gannett, and Microsoft Start incidents showed how quickly published AI content can damage a brand. Reader surveys by the Reuters Institute and the Knight Foundation in 2024 found majorities of respondents in the US, UK, France, and Germany said they did not trust AI generated news and would be less likely to read an article they knew had been written by a chatbot. Publishers have largely responded by adopting disclosure policies, by emphasizing human authorship, and in some cases by signing the Journalism Trust Initiative or affiliating with the Trust Project to signal editorial standards.
A fourth strand is the moderation question. Generative AI makes it cheap to produce harassment, scam content, non consensual sexual imagery, and propaganda at scale; the same systems are being used to moderate that content. Whether moderation can keep up with generation, and whether either should be controlled by a small number of large companies, is the dominant question in current platform policy.