Case Description
The Oversight Board has selected a case involving an apparently AI-created video of a British politician that misrepresents her views on immigration. The Board will evaluate Meta’s actions regarding this post in light of the company's policies and human rights responsibilities to ensure respect for freedom of expression while addressing the possibly harmful impact of deceptive content related to important public matters.
A Facebook user in November 2025 posted an album including a short video showing an impersonation of a UK Labour Party Councillor who represents an area in Scotland. The video, which appears to be AI-generated, shows the politician saying: “Refugees are welcome here, even if they rape our women, because white people do that too.” A second video in the album shows people protesting, including a man waving a Palestinian flag with seemingly AI-generated audio of him chanting praise for Antifa, a left-wing, anti-fascist movement. A final image that appears to be real rather than AI-generated shows several women, including the politician depicted in the first video, holding anti-far-right protest signs and names them. The album’s caption accuses the politician of tax evasion without providing evidence. The user did not disclose AI use, and no informative AI label was applied by Meta to the content. The content received fewer than 50 reactions and comments, and fewer than 50 shares.
Two users reported the content for violating the Bullying and Harassment policy, but Meta’s systems did not prioritize the post for human review so the content remained on Facebook. Both users appealed to Meta, but again the post was not prioritized for human review. One of the users then appealed to the Board.
In their statement to the Board, the reporting user alleges that the video is AI-generated and misrepresents the politician’s beliefs, in the context of a demonstration where she supported the housing of asylum seekers. The user said the content may pose a threat to the woman’s safety.
Anti-immigration protests have taken place across the United Kingdom in 2025 and 2026. Some protesters claimed the housing of asylum seekers in hotels around the UK created safety concerns for women. Counter-protesters, including the depicted politician, have publicly rejected these claims as racist disinformation weaponizing concerns about violence against women to fuel anti-migrant hate. They have also reported intimidation and threats against them for speaking out, including online harassment and defamation using AI-generated content. Immigration and housing for asylum seekers have been topics of intense political debate in the run-up to elections, including those in Scotland on May 7, 2026. The depicted councillor was not up for re-election in 2026, however.
When the Board selected the case, Meta’s subject matter experts concluded the post did not violate the company’s community standards and therefore did not require removal, or the use of an AI label.
According to Meta, the content did not violate its Bullying and Harassment policy because the politician is an adult public figure and therefore not protected from “unwanted manipulated imagery,” whereas private individuals can self-report such content for it to be removed. Meta also determined that the claim in the video that refugees commit rape did not violate its Hateful Conduct policy because it was an assertion against the actions of some refugees rather than a generalization equating all or most refugees with criminals. The company noted that the post was not removed as misinformation since no Trusted Partners flagged it as false and likely to “directly contribute to the risk of imminent harm,” and it was not election interference. Meta’s Trusted Partner program is a global network of independent organizations, agencies and researchers that flag emerging risks from content. The content was not reviewed by third-party fact-checkers.
Meta also determined that the content did not merit an AI label under its Misinformation policy to inform users that the content was digitally created or altered. According to Meta, it did not require a label under its manipulated media rule as the content was not posted during an election or crisis, is satirical, received little engagement and it is unlikely to “create a particularly high risk of materially deceiving the public on a matter of public importance.”
The Board selected this case to evaluate Meta’s human rights responsibilities when AI tools are used to impersonate politicians to misrepresent their views on matters important to public discourse. The case allows the Board to investigate how platforms can make design, policy and enforcement choices that respect freedom of expression, including to satirize or criticize politicians, while addressing harms that may result from deceiving the public on important matters.
The case falls within the Board’s Automation and AI, and Hate Speech Against Marginalized Groups strategic priorities.
The Board would appreciate public comments that address:
- The prevalence, impact and sources of deceptive AI-generated content in shaping public opinion on immigration in Europe and globally.
- How social media platforms should address AI-generated harassment, hate or deception, including against politicians, while respecting freedom of expression, especially for political criticism, satire and humour, including the necessity and proportionality of removal or alternative measures to reduce dissemination.
- If and how platform policy and design choices can create incentives, including through monetization and recommendations, for people to share deceptive AI-generated content and how to mitigate or prevent adverse human rights impacts from them.
- How deepfake techniques are used in harassment campaigns against people who speak out on controversial issues, including politicians, and the broader impacts of deceptive AI-generated attacks to discredit people on access to information and political participation.
In its decisions, the Board can issue policy recommendations to Meta. While recommendations are not binding, Meta must respond to them within 60 days. As such, the Board welcomes public comments proposing recommendations that are relevant to this case.
Comments
CENTRE FOR DEMOCRACY AND RULE OF LAW
Public Comment to the Meta Oversight Board
Introduction
Centre for Democracy and Rule of Law is a Ukrainian civil society organisation specialising, inter alia, in platform governance, digital rights, media law, and information integrity. We submit this comment as a practitioner organisation with documented experience of AI-generated political disinformation campaigns. Ukrainian officials, civil society figures, and journalists have been targeted by synthetic media campaigns structurally identical to the case before the Board — fabricated audio and video placing inflammatory statements in the mouths of real individuals to misrepresent their views and silence them.
Documented cases from Ukraine include: a 2022 deepfake of President Zelensky calling on Ukrainian forces to surrender, distributed via social media platforms; a 2022 deepfake of Kyiv Mayor Vitali Klitschko used to deceive European city mayors in fabricated video calls; a 2023 deepfake of then-Commander-in-Chief General Zaluzhnyi making fabricated statements about President Zelenskyy; a 2024 lip-sync deepfake of National Security Council Secretary Danilov falsely claiming responsibility for the Crocus City Hall attack; and a case involving Myroslava Keryk, Director of the Ukrainian House foundation in Warsaw, when a deepfake video replicating her appearance and voice circulated on social media, falsely depicting her making offensive remarks about Poles around Poland's Independence Day — a fabrication confirmed by the foundation and designed to inflame Ukrainian-Polish tensions. These cases demonstrate a consistent pattern: fabricated synthetic media is used to attribute inflammatory or operationally significant statements to real named individuals, in order to discredit them, manipulate their interlocutors, or influence public perception of contested events.
This comment addresses four issues related to those raised by the case before the Board: harassment protections for locally elected officials; AI labeling contingent on electoral timing; deepfake detection in the context of biometric data constraints; and deepfakes in targeted advertising as a vector of fraud and reputational harm. It follows the structure: challenge – rationale – suggested solution.
I. Inadequate harassment protections for locally elected officials
Challenge
According to Meta, its Bullying and Harassment policy exempts adult public figures from protection against “unwanted manipulated imagery,” reserving that protection for private individuals. Applied without distinction, this deprives locally elected representatives of any effective remedy against fabricated content designed to misrepresent their views and expose them to hostility — even when those individuals lack the institutional resources to counter such campaigns.
Rationale
The public/private figure distinction calibrates tolerance for criticism of those who enter public life. It was not designed to treat a local councillor and a cabinet minister identically for the purposes of harassment protection. A local councillor may have no press office, legal team, or communications infrastructure. The content at issue cannot be reduced to criticism or satire: it is a fabricated video placing words the politician never said into her mouth, combined with a real photograph identifying her by name, a second manipulated video, and an unsubstantiated tax evasion allegation. Taken together, its purpose seems to be reputational destruction, not political expression. Low engagement does not reduce this harm — the target’s reputation and safety are not protected by the post having fewer than 50 reactions. The harm is foreseeable and proportionate intervention is required.
Suggested solutions:
- Fine-tune a tiered public figure framework under the Bullying and Harassment policy, distinguishing between senior officials with institutional support and locally elected representatives who retain meaningful vulnerability to targeted harassment.
- Extend protection against manipulated and AI-generated imagery to all public officials where content fabricates statements they never made. Such content should be eligible for removal on self-report or report by third parties, adapting the existing mechanism accordingly.
- Introduce expedited human review for reports of synthetic content depicting named real individuals, regardless of the post’s engagement level.
II. AI labelling contingent on electoral timing is structurally inadequate
Challenge
Meta declined to apply an AI disclosure label because the post was arguable not made during an election or crisis period, received low engagement, and was assessed as unlikely to materially deceive the public. This framework produces systematic under-labelling of deceptive synthetic media by conditioning disclosure on prospective risk criteria that do not track actual harm.
Rationale
The harm caused by realistic synthetic media depicting a named person making statements they never made does not arise or dissolve according to electoral calendars. The depicted politician’s reputation and safety are not less at risk outside an election period; deceptive potential is not diminished by low initial engagement. Unlabelled content is treated by users as presumptively authentic.
The content here is realistically formatted, provides no disclosed signal of satire or AI generation, depicts a named person making an attributable statement on a contested political issue, and appeared in a documented context of real-world harassment against counter-protesters. Content that looks real and gives viewers no reason to doubt its authenticity cannot escape disclosure requirements by being labeled satirical after the fact.
Suggested solutions:
- Extend mandatory AI disclosure to all deepfakes depicting named real individuals making attributable statements, regardless of electoral timing, crisis context, or engagement level. The operative test: could a reasonable viewer be deceived into believing the person made the statement?
- Apply differentiated labeling — stronger visual warnings and pre-sharing friction — to synthetic content depicting a named individual on a contested political issue, formatted as authentic footage, or appearing in a context of documented harassment against that individual.
- Remove the satire carve-out where content provides no explicit signal of invention or exaggeration.
- Support and accelerate adoption of provenance standards (e.g. C2PA Content Credentials), reducing reliance on voluntary uploader declaration or post-hoc automated detection.
III. Deepfake detection and biometric data constraints
Challenge
Effective automated detection of deepfakes — particularly face-swap and voice-cloning techniques — commonly relies on processing biometric data: facial geometry, voice characteristics, and related biometric signals. In many jurisdictions, including EU member states, processing biometric data is subject to strict legal constraints under data protection law. This creates a structural tension: the most technically effective detection methods may not be legally deployable at scale, leaving platforms reliant on less accurate alternatives or on uploader disclosure. Meta’s current approach does not sufficiently address this tension. There is also no dedicated reporting option enabling users to flag the specific harm of a public figure being depicted in AI-manipulated video or audio content.
Rationale
The gap matters because biometric-based detection is currently arguably the most accurate. Where legal constraints prevent its use, platforms need viable alternative detection strategies. Approaches that analyse media provenance and integrity metadata (e.g. C2PA Content Credentials), generation artefacts (compression anomalies, lighting inconsistencies, audio-visual synchronisation errors), and contextual signals (account behaviour, posting patterns, upload metadata) can reduce but not eliminate the detection gap. The Board should encourage Meta to develop and publish a roadmap for detection methods that do not require biometric processing, acknowledging the current accuracy trade-offs transparently.
The absence of a specific reporting category for deepfake or AI manipulation of public figures compounds this problem. A dedicated option — for example, “Use of a real person’s likeness or voice in AI-manipulated content” — would both improve detection and generate better data on the scale and nature of the problem.
Suggested solutions:
- Pending resolution of applicable legal constraints on biometric data processing, develop and publish a roadmap for deepfake detection methods that do not require biometric data, including provenance metadata analysis (C2PA or equivalent), generation artefact detection, and contextual behavioural signals. Acknowledge the accuracy limitations of non-biometric approaches transparently in public reporting.
- Add a dedicated reporting category — “Use of a real person’s likeness or voice in AI-manipulated content” — in Meta’s reporting interface, distinct from general categories. Route these reports to a specialised triage queue with defined response time standards.
- Use data from this dedicated reporting category to publish aggregate transparency data on the volume, nature, and enforcement outcomes for reported synthetic media depicting real named individuals.
- Engage with EU institutions — including the European Data Protection Board, the AI Office, and the European Commission — on the need to develop guidance that balances biometric data protection requirements with the legitimate public interest in detecting synthetic media that threatens, inter alia, electoral integrity and freedom of expression. Current data protection rules were not designed with large-scale deepfake detection in mind; regulatory clarification is needed to enable proportionate detection measures that respect both privacy rights and the legitimate interest in detecting synthetic media that threatens democratic processes and undermines the integrity of public discourse.
IV. Deepfakes in targeted advertising as a vector of fraud and reputational harm
Challenge
The synthetic media enforcement gap documented in this case has broader implications — it also extends into Meta’s advertising ecosystem, where it takes a distinct and commercially significant form. Coordinated fraud campaigns, as documented by CEDEM, systematically use deepfakes of named public figures — politicians, news anchors, doctors, etc. — to promote fraudulent medical products and investment schemes. A fabricated video of a well-known doctor endorsing a “miracle cure,” or a deepfake of a politician recommending a fraudulent investment platform, exploits the reputational capital of real individuals to deceive consumers into financial or health harm. Unlike organic posts, these are paid placements: Meta receives advertising revenue from their distribution, the advertising post reach and impact more people, which creates a heightened duty of care.
Rationale
Advertising content passes through Meta’s ad review system before publication, yet deepfake-based fraud ads circulate at scale. Several structural features may explain this:
• Ad review systems are insufficiently designed to detect synthetic media manipulation of real individuals’ likenesses.
• Fraudulent advertisers exploit redirects to high-risk top-level domains (.xyz, .fun, .space and similar) and trigger-word combinations (“miracle,” “guaranteed returns,” “urgent,” “cure”) that are associated with scam content but are not systematically screened at the ad approval stage.
• Dietary supplements and similar products occupy a regulatory grey zone: they are not subject to the same advertising restrictions as prescription medicines, yet are frequently marketed using fabricated medical endorsements from politicians, real named doctors or public health figures.
The reputational harm to the individuals depicted is direct and serious: their name, face, and voice are used without consent to promote products/services they have not endorsed and may actively oppose. In Ukraine, CEDEM has observed cases of well-known public figures — including politicians, doctors, etc — whose likenesses have been used in fraudulent health and investment advertising on social media platforms. The individuals concerned typically have no effective remedy: by the time a complaint is processed, the ad campaign has run its course and the advertiser (which is usually a so-called “one-day page” created solely for the purpose of launching a fraudulent ad campaign) has disappeared.
This problem engages Meta’s human rights responsibilities on two axes: harm to the individuals whose likeness is misappropriated, and harm to consumers deceived by fabricated endorsements into making decisions damaging to their health or finances.
Suggested solutions:
- Enhance deepfake and synthetic media detection in Meta’s ad review pipeline. Ads containing video or audio content depicting real named individuals should be subject to an additional verification step before approval, with particular scrutiny applied to ads for medical products, dietary supplements, and financial services.
- Apply enhanced scrutiny to ads that combine any of the following risk signals: destination domains on potentially high-risk TLDs (.xyz, .fun, .space and similar); trigger-word combinations associated with health or investment fraud; claims of medical efficacy or guaranteed financial returns; and video or audio content that has not been verified as authentic.
- Extend the dedicated “Use of a real person’s likeness or voice in AI-manipulated content” reporting category (recommended above) explicitly to advertising content, and establish a fast-track review process for reports submitted by the depicted individual themselves or third parties (considering the impact and reach of advertising content).
- Publish disaggregated transparency data on enforcement actions against ads using synthetic media to impersonate real named individuals, including the categories of products advertised, volume of ads removed, and advertiser account actions taken.
Conclusion
CEDEM urges the Board to use this case to establish clearer standards for the governance of synthetic media targeting real individuals — grounded in international human rights law, operationally realistic about current limits of detection and enforcement, and extending across both organic content and paid advertising.
Please find attached our comment as a PDF file.
Please find the comment attached.
I am a senior reporter at the Bureau of Investigative Journalism in London, and for the past eight months I have been investigating the overwhelming amount of AI-generated misinformation targeting UK audiences on Meta platforms. I have researched this area extensively and published two major investigations and an opinion piece on the subject, as well as interviewing successful creators of commercial AI slop. I am therefore well placed to diagnose the structural factors which have led to UK Facebook being flooded with a particular kind of sensational, racist and anti-migrant content, much of which uses deepfakes of UK politicians in the same manner as the post being adjudicated by the Board.
The most fundamental point is that AI-generated content which makes sensationalised and racist claims performs well on Meta platforms. This has in turn produced an entire industry of creators based in the Global South who monetise these posts via Meta’s content monetisation tool. A Pakistani creator who ran a page called Britain Today told us he made $1,500 a month from this page alone; Geeth Sooriyapura, the Sri Lankan creator, claimed to have made $300,000 over the course of his Facebook career. We weren’t able to verify these figures, but both men were certainly making many times the average income in their countries. It is notable that the rise of this kind of content correlates with Meta paying out a record amount to creators, as was disclosed in a March 2026 report on the Creator Fast Track programme.
Even a couple of years ago, one would assume that misinformation about UK politicians and posts demonising migrants and other minorities were politically motivated, and expressed a sincere belief held by the person posting it. Now, a significant proportion of political content on Facebook is being produced by people thousands of miles away, who have absolutely no interest in UK politics beyond their ability to monetise posts.
The rise of generative AI tools have completely removed the need to have any knowledge of UK politics, or even of the English language. These tools are used at every stage of the content creation process: to brainstorm ideas, to write and translate captions and, most importantly, to create compelling images and videos. Since Meta removed access to Crowdtangle, it is difficult for journalists to determine the scale of commercial AI slop, but in my opinion a statistically significant proportion of UK politics content on Meta is commercially motivated and produced by creators in hotspots such as India, Pakistan and Vietnam.
I have manually reviewed a large amount of content produced by these pages. Some of it is certainly permissible under Meta rules, but much of it is hate speech, and when we shared examples Meta took many posts down, but not before they had been viewed millions of times. Often this hate speech is directed at Muslims and migrants in general, but some of it directly targets particular politicians: most commonly Sadiq Khan and Keir Starmer. Recent extreme examples I have reviewed include an AI-generated image of the prime minister text which reads: “Ostarmer Bin Laden Wanted For Terrorist Crimes Against White Christians Pensioners and the Disabled”; another deepfake video of Starmer shows his avatar making a speech which praises “Our crazy violent Muslim immigrants, including our precious criminals and rape gangs.” Later in the video, the avatar uses a racist term for Pakistani people. Before I reported this video to Meta it had 400,000 views.
The same creator had shared a video of Khan at a public iftar in Trafalgar square, described in a caption as a “colonisation event”. The conspiracy theory that Muslims are “colonising” the UK is a common trope in posts from these accounts.
Meta should certainly be more proactive in detecting and removing hate speech, and it is clear that its decision to scale back its trust and safety operations has exacerbated the problem. But by far the important thing Meta could do to fix this problem is adjust the incentives via the content monetisation tool. Even if Meta did more to take these pages down, if someone in Pakistan is able to make $1500 a month producing racist AI slop about the UK, new pages will always appear in their place. Meta has to approve each creator’s application to get access to the content monetisation tool: it is clear that there must be much more verification done at this stage, as well as ongoing monitoring and demonetisation if creators are found to violate Meta’s rules. There appears to be very little verification or monitoring of creators who have access to the content monetisation tool going on at present.
The problem is that as it stands, the system works well for Meta too: a creator in South Asia posts sensational, racist content, it gets a lot of views, and Meta records increased user engagement and ad revenue. However I would argue that, as well as the moral and legal argument for removing hate speech, the user experience on Facebook is being increasingly degraded by commercial AI slop, which will surely begin to affect Meta’s profitability.