AI-Generated Video in Israel-Iran Conflict

Name

Ila Rutten

Organization

CDAC Network

Country

Switzerland

Language

English

Sudan’s Information War: How AI-Generated and AI-Amplified Mis/Disinformation Endangers Civilians and Humanitarian Responders
Public Comment Submission to the Meta Oversight Board

1. Executive Summary
Sudan is experiencing one of the deadliest conflicts in the world, and one of the deadliest in history for humanitarian responders. More than 120 humanitarian workers have been killed since April 2023. Alongside mass displacement, the collapse of essential services, and widespread violence, the country faces a second and often less visible emergency: a weaponised information environment in which misinformation, disinformation, hate speech, and AI-accelerated content directly contribute to physical harm.
Since the war began, harmful information has been used deliberately by armed groups, political elites, and opportunistic actors to shape public perceptions, mobilise violence, obstruct humanitarian operations, and undermine social cohesion. CDAC Network’s 2025 report Sudan’s Information War documented fatal and life-altering consequences: from panic-induced displacement to attacks on aid workers triggered by false online claims; from doctored videos used to justify ethnic violence to rumours that obstruct lifesaving aid deliveries. It is in this context that Sudan demonstrates the exact category of high-risk environment the Oversight Board is examining in the Israel–Iran case: contexts where AI-generated and AI-amplified content undermine information integrity and result in real-world harm.
This submission draws on CDAC Network’s 2025 Sudan research, including the Local Lifelines project and the newly launched collaboration with Valent, which moves from diagnosis to response by co-designing safe, AI-supported tools with Sudanese Emergency Response Rooms (ERRs) and local civil society. Together, these research streams reveal critical insights about how Meta’s current policies, including the requirement that harmful content “directly contribute to imminent physical harm”—fail to capture the cumulative, cascading, and networked nature of harm in modern conflict information environments.

Sudan shows that AI-amplified posts boosted by bots, coordinated networks, and automated engagement systems can rapidly distort perceptions, provoke fear, and mobilise attacks. These harms disproportionately affect communities already facing severe access barriers, limited connectivity, and the near-total collapse of independent media.
This submission argues that Meta has human rights responsibilities to adapt its detection, escalation, and moderation systems to high-risk conflict settings—especially those marked by degraded media ecosystems, lethal misinformation, and systematic targeting of humanitarian responders. Sudan illustrates the urgent need for conflict-sensitive, locally informed moderation approaches; improved detection of coordinated inauthentic behaviour; and partnerships that elevate the voices of trusted local intermediaries.

2. Background: Sudan’s Information Battlefield
The war that erupted in Sudan in April 2023 produced one of the most dangerous operational environments for civilians and humanitarian responders in decades. Alongside mass killings, atrocities, and a dramatic expansion of displacement, Sudan’s information landscape has been transformed into an extension of the battlefield.
CDAC’s 2025 report documents how misinformation, disinformation, propaganda, malinformation, and hate speech are now “deeply intertwined with the violence on the ground”, often accompanying or following attacks on civilian infrastructure by both the Sudanese Armed Forces (SAF) and the Rapid Support Forces (RSF) . Social media is used to incite violence, manufacture suspicion, justify attacks, and delegitimise humanitarian actors. The result is a fragmented information ecosystem that directly shapes who receives aid, who is targeted, and who survives.

Collapse of Sudan’s Media and Connectivity Landscape
Sudan’s media infrastructure has been decimated. As documented in the CDAC report, hundreds of journalists have fled, dozens of outlets have been destroyed, and many online spaces have been overtaken by political or military propagandists . Connectivity blackouts—deliberately imposed or resulting from infrastructure collapse—further degrade verification capacity, heightening reliance on unverified social media content and WhatsApp chains.

In this void, Emergency Response Rooms (ERRs), informal mutual aid groups, and local volunteers became de facto frontline information intermediaries. They verify safe routes, aid availability, and displacement patterns during periods where communities have no access to radio, television, or formal humanitarian communication channels. Yet these same groups face coordinated online attacks that portray them as militia collaborators,placing them in extreme danger and inhibiting their ability to operate safely.

Weaponisation of Narratives
Sudan’s warring parties use strategic online campaigns to frame the conflict, incite ethnic polarisation, and justify violence. The CDAC report shows how RSF-aligned channels portray their struggle as a fight against historical marginalisation, while SAF-aligned influencers deploy hashtags such as “truce is treason” to suppress any calls for peace or humanitarian pauses .
These narratives have real consequences. In one documented incident, false claims on Facebook that a community-run kitchen in Khartoum had fed RSF fighters led to an attack in which local volunteers were killed. In another case, NRC faced widespread online allegations of smuggling weapons, leading to access disruptions and physical insecurity. These are not “online harms” in the abstract; they are directly connected to life-or-death outcomes for Sudanese responders and civilians.

3. The Role of AI-Generated and AI-Amplified Mis/Disinformation in Sudan
Sudan’s information war is not defined solely by generative AI–produced deepfakes. Rather, it is shaped by the combined effects of AI-generated content, algorithmic amplification, bot networks, and coordinated inauthentic behaviour,an ecosystem described in detail by the UNU-CPR report on peacebuilding in Sub-Saharan Africa and corroborated by CDAC’s Sudan analysis.
AI has fundamentally lowered the cost of producing and circulating harmful content. Generative AI tools now make it possible to create large volumes of false content text, video, and audio.at near-zero cost, allowing political and military actors to deploy influence operations with unprecedented ease . Additionally, botnets, troll farms, and automated posting systems are increasingly used to manipulate trending topics, distort public conversations, and overwhelm fact-checking ecosystems.

3.1.1 Documented Deepfake Use in Sudan’s Conflict
Multiple documented deepfakes have circulated in Sudan’s online conflict environment since 2023, including AI-generated images, audio recordings, and impersonations of political and military figures. Examples include:
• An April 2024 AI-generated image falsely claiming the Sudanese Armed Forces had bombed a building at Al-Jazeera University in Wad Madani, which went viral on Facebook and was reshared by political leaders.
• An AI-generated video in August 2023 falsely depicting the U.S. ambassador to Sudan announcing a U.S. plan to reduce the influence of Islam, identified by the Daily Mail.
• A BBC investigation in October 2023 revealing an AI-generated impersonation of former Sudanese president Omar al-Bashir, which received hundreds of thousands of views on TikTok.
• A March 2024 AI-generated audio recording attributed to the head of the Sudanese Armed Forces ordering the killing of civilians, which was viewed 230,000 times and shared by Sudanese politicians before debunking.
• Deepfake satire, such as a manipulated pro-SAF song showing RSF leader Mohamed Hamdan Dagalo (Hemedti) singing, which has sometimes been repurposed as disinformation.
• A March 2024 AI-generated audio recording purporting to show RSF leaders meeting with members of the Freedom and Change coalition to coordinate a coup, amplified by journalists and national television before being retracted.
• The “liar’s dividend” phenomenon, such as when Sudanese politician Mubarak Ardol claimed a recording of himself was fabricated with AI.
• Beam Reports, Sudan’s IFCN-verified fact-checking organisation, has analysed these deepfakes and noted that while deepfake usage has not surged recently, “misleading audio content has also been generated using AI and attributed to people active in Sudanese affairs.”
• At a UNESCO seminar in May 2024, Beam Reports highlighted that the collapse of on-the-ground reporting increases the spread of mis/disinformation, which is “amplified and complicated by the increasing use of generative AI in the production and dissemination of disinformation and hate speech.”
• Sudanese fact-checkers report difficulty detecting deepfakes due to lack of reliable detection tools, file compression issues, and model bias, especially for Sudanese dialects.
These examples establish that deepfakes, both visual and audio, are now part of Sudan’s conflict information terrain, even if many remain identifiable due to imperfect Sudanese dialect modelling or low training data availability.

3.2 AI-Amplified Narratives: A Dominant Threat
The most impactful harmful narratives in Sudan continue to derive not only from deepfakes but from the amplification of misleading content, whether synthetic or real. CDAC’s Sudan report highlights that miscaptioned real images, false claims about aid workers, and polarising narratives often spread faster and more dangerously than sophisticated fabrications
.
The Valent November dataset demonstrates:
• Bot driven amplification, driven by high-follower accounts, significantly increases the reach and perceived legitimacy of false narratives.
• The claim that SAF shot down an Emirati cargo plane over El Fasher received rapid, coordinated engagement from accounts with hundreds of thousands of followers, escalating suspicion towards humanitarian air transport
• Panic narratives about bombardments in Al-Daein similarly tracked with coordinated amplification and mass resharing driven by key nodes.
Deepfakes do exist in Sudan, but CDAC’s evidence suggests that the most operationally dangerous content consists of repurposed imagery, synthetic text overlays, and coordinated amplification .
3.3 Incentive Structures for AI-Driven Harm
Economic, political, and social incentives drive the use of AI for harmful information operations in Sudan :
• Warring parties use AI-boosted narratives to maintain morale, incite mobilisation, and justify attacks.
• Diaspora influencers monetise attention by posting sensational or unverified claims.
• Regional actors exploit Sudan’s conflict as a proxy theatre and deploy influence operations accordingly.
• Ordinary users share alarming or graphic posts faster than verified information, especially during offensives, sieges, or blackout periods.
In a setting where connectivity is intermittent and trustworthy sources are scarce, AI-amplified content effectively outcompetes verified information shaping community behaviour in dangerous ways.

4. Prevalence and Impact of AI-Driven Harmful Information in Sudan
CDAC’s research shows that harmful information in Sudan operates as a systemic, not episodic, threat. The Executive Summary of Sudan’s Information War notes that MDH is used to “distort narratives, fracture social cohesion, and obstruct humanitarian access” . When viewed alongside recent data sets a clearer picture emerges: harmful narratives are not only widespread, they follow identifiable patterns of artificial escalation and coordinated dissemination.
4.1 Impact on Civilian Behaviour
Harmful narratives influence life-and-death decisions:
• Communities flee neighbourhoods based on false claims of imminent attacks.
• Rumours about aid deliveries cause crowding or keep people away altogether.
• False accusations that particular ethnic groups support one side fuel ethnic reprisals and violence.
Examples from the November dataset show narratives with tens of millions of impressions, such as claims about atrocities in Al-Abyad or the need for “safe corridors” in El Fasher, spreading far faster than verified humanitarian information .
4.2 Impact on Humanitarian Access and Safety
CDAC’s Sudan report documents multiple cases where misinformation directly obstructed humanitarian operations:
• False claims that NGOs were smuggling weapons derailed convoys and fuelled suspicion toward aid workers.
• Misinformation about ERRs and grassroots responders led to killings, detentions, and intimidation.
• Accusations of affiliation with opposing armed groups undermined humanitarian neutrality.

A further example is the November 2025 incident involving the ICRC, in which a widely circulated social media image misidentified a vehicle as belonging to the organisation. The ICRC publicly clarified that the vehicle in the image was not affiliated with the ICRC and did not display the roundel used by the organisation in Sudan. Only a cropped red cross symbol was visible, and the ICRC stated it could not identify the vehicle. The organisation stressed that this kind of misrepresentation, circulated rapidly and without verification, places its staff and other humanitarian workers at risk by eroding the protections afforded under international humanitarian law. This incident demonstrates how even non-AI-generated, miscaptioned imagery can undermine humanitarian neutrality, shrink protected operational space, and create conditions in which aid personnel become targets.

In such a context, reporting tools that deprioritise Sudan or require imminent, direct harm before intervention fail to prevent violence. The “imminent physical harm” threshold is too narrow for a conflict where harm accumulates through spiralling misinformation cycles.

4.3 Impact on Trust and Institutional Legitimacy
The breakdown of Sudan’s information ecosystem has eroded trust between communities and humanitarian organisations. CDAC’s report shows that communities increasingly doubt the neutrality of aid actors and rely instead on informal networks.

Because local media is weak and telecommunications infrastructure is frequently destroyed or shut down, communities are left with few sources of verified information, making them extraordinarily vulnerable to manipulation.

5. Challenges in Detecting, Labelling, and Moderating AI-Driven Harmful Content in Sudan
5.1 Technical Barriers to Detection
There are multiple challenges in detecting AI-altered and AI-generated content: multilingual complexities, low-resolution media, rapid posting cycles, and the speed at which AI can create or modify content . These challenges are magnified in Sudan:
• Sudanese Arabic is under-resourced in global language models. Automated classifiers struggle with dialects, slang, and context-specific terms used by Sudanese users.
• Many posts combine real footage with false captions, a hybrid form of manipulation that AI systems typically misclassify.
• Content is often screenshot, re-uploaded, cropped, or recompressed—all of which reduce the efficacy of detection systems.
This means that harmful content, including AI-generated clips, often remains online long enough to shape community behaviour and provoke violence.
5.2 Collapse of Local Verification Ecosystems
CDAC’s Sudan report documents the near-total degradation of Sudan’s domestic media ecosystem and the exodus of journalists and fact-checkers . In the absence of independent verification:
• Communities rely on WhatsApp groups and Telegram channels where rumours spread unchecked.
• Mutual Aid Groups (MAGs) and ERR volunteers verify real-time safety information manually, often at enormous personal risk.
• Fact-checking organisations like Beam Reports operate from exile, with limited ability to scale verification or counter falsehoods.
This collapse means that platforms cannot rely on “crowdsourced” accuracy signals, community reporting, or local media debunking to correct false narratives.
5.3 Gaps in Meta’s Enforcement and Escalation
CDAC’s research shows that users widely perceive harmful content reporting on Meta platforms as futile, with content often left online long after it has been flagged . Responders report:
• No clear escalation pathways for high-risk conflict content.
• High false negatives: harmful content stays online long enough to provoke real-world consequences.
Meta’s current policy requires that a piece of content “directly contribute to imminent physical harm” before it is removed. Sudan provides overwhelming evidence that this standard is incompatible with conflict realities. In Sudan, the harm emerges from cumulative cycles of misinformation, not a single post. A rumour that mutates across channels can trigger panic, reprisals, attacks on responders, and large-scale displacement even if no individual post meets Meta’s threshold.

5.4 Coordinated Inauthentic Behaviour and AI-Assisted Amplification
The Sudan Pt. 2 project’s narrative monitoring confirms that AI-amplified content is a central feature of Sudan’s information landscape. Our dataset shows clear signatures of coordinated inauthentic behaviour:
• Inauthentic accounts with hundreds of thousands of followers simultaneously promote a narrative was present in major conflict-related claims, including:
• The false allegation that SAF shot down an Emirati plane over El Fasher.
• Panic narratives about bombardment in Al-Daein.
• Militia-clash narratives in Dongola.
These spikes reflect automated or semi-automated coordination that outpaces organic verification efforts .
• Accounts involved in amplification often show bot-like traits: high posting frequency, synchronized timing, and repetitive patterns of engagement.
These techniques are well within the capabilities of actors engaged in Sudan’s conflict .
Yet Meta’s moderation systems—designed for individual posts, not narrative-level patterns struggle to detect these coordinated campaigns. This leaves civilians and responders facing manipulation that the platform is not structurally equipped to address.

6. Human Rights Responsibilities of Meta in Armed Conflict
Meta’s global human rights responsibilities under the UN Guiding Principles on Business and Human Rights (UNGPs) require it to identify, prevent, and mitigate adverse human rights impacts linked to its platforms. Sudan provides a clear, urgent case where these responsibilities are both relevant and unfulfilled.
6.1 Impacts on Core Human Rights
Drawing on CDAC’s Sudan report :
• Right to life and security of person:
Misinformation has directly triggered attackse.g., the fatal attack on community kitchen volunteers, and threats or violence against ERR members and aid workers following online smear campaigns.
• Right to freedom of movement:
False narratives about safe corridors or imminent attacks cause dangerous displacement, expose civilians to bombardment, and impede safe evacuation.
• Right to non-discrimination:
AI-amplified hate speech reinforces ethnic polarisation between communities in Darfur, Kordofan, and the Nile region.
• Right to access to information:
When telecommunications networks collapse and local media is decimated, platforms like Facebook and WhatsApp become primary information channels. Over-removal of authentic crisis speech and under-removal of harmful misinformation both violate this right.
6.2 Harms to Humanitarian Operations
Meta has specific responsibilities when its platforms materially affect the ability of humanitarian actors to operate safely. Sudan offers several documented examples:
• Erroneous allegations that humanitarian organisations supported one side of the conflict have obstructed aid delivery.
• Coordinated attacks on ERRs and civil society actors—fuelled by online narratives—have resulted in detentions, threats, and killings.
• False rumours about cross-border relief convoys have undermined trust and cooperation with communities and local authorities.
These harms are foreseeable, repeated, and systematically linked to online content.
6.3 Why the “Imminent Physical Harm” Standard Is Insufficient
Meta’s focus on individual pieces of content, rather than narrative-level harm, misses the way information cascades in conflict settings.
Sudan demonstrates that:
• Harm emerges cumulatively across dozens or hundreds of posts.
• Panic, fear, and violence are triggered not by a single piece of content but by waves of coordinated or AI-amplified narratives.
• Humanitarian access is obstructed by atmospheres of suspicion produced over time, not by isolated violations.
Given this, Meta’s current policies do not adequately protect users—or responders—in high-risk conflict settings.

Name

Mahsa Alimardani

Organization

WITNESS

Country

United Kingdom

Language

English

Attachments

Shortened-WITNESS-Oversight-Board_-Iran-Israel-.docx

Name

Nathan Heath

Country

United States

Language

English

This case tragically showcases the weaponization of generative AI to fabricate "evidence" of successful strikes during active conflict. Meta’s decision to leave a video depicting significant fabricated damage in Haifa on the platform (justified by the claim it did not “directly contribute to the risk of imminent physical harm”) reveals a profound and dangerous miscalibration of risk for this highly volatile environment.

In the context of the June 2025 conflict, a high-virality, AI-generated video (700,000+ views) of alleged destruction in a major population center like Haifa is not merely misinformation, but also an immediate accelerant of escalation. In the delicate equation of deterrence between Israel and Iran, the scale of retaliation is highly sensitive to the perceived success of an adversary’s initial strike.

A video that fabricates successful missile impacts pressures political leadership to demonstrate an equally forceful response to both the domestic and regional audience. This synthetic information feeds the two rivals’ vicious escalation ladder, as public outrage, driven by visual “evidence” of destruction, compels swift kinetic action. Meta's criteria, which require a direct contribution to physical harm, entirely fails to account for the speed and impact of informational warfare on political decision-making and subsequent military responses. By the time fact-checkers might have rated the content, the political will for escalation may have already solidified.

The case details highlight the page self-identifying as a news source, a classic tactic in coordinated Information Operations (IO) designed to confer false legitimacy. Crucially, the content was flagged by an automated Meta classifier yet not prioritized for human review, despite multiple user reports citing "terrorism" and "violence." 

This pinpoints a critical flaw in the platform's content moderation architecture: the taxonomy of harm. In a live armed conflict, a realistic, high-volume AI-generated fake claiming a successful strike is not merely a policy violation related to "misinformation,” but intrinsically related to the immediate security environment and potential for violence. The further failure of third-party fact-checkers to rate the content, despite the internal classifier flag, strongly suggests that the sheer volume of synthetic media during the 12-day conflict overwhelmed existing verification pipelines. We should not rely on a system so easily surpassed by the speed of automated content generation.

Based on the failure modes demonstrated in this specific case, I propose the following policy recommendations for the Oversight Board to consider advising Meta:

1. Crisis Circuit Breaker for Synthetic Media: During officially designated high-risk conflicts, Meta must implement a "circuit breaker" on unverified, high-virality video content that is flagged by classifiers as likely AI-generated. This content should be automatically and severely down-ranked, or placed behind a mandatory "Unverified/Suspected AI" interstitial friction screen, until a human safety specialist or fact-checker can definitively clear or remove it.

2. Expanded Definition of Imminent Harm: The Board should recommend that Meta broaden its interpretation of "imminent physical harm" during active armed conflicts to include the active, high-volume, and fabricated dissemination of claims regarding:
• Successful military strikes on civilian centers.
• False declarations of imminent military action by state or non-state actors.

3. Mandatory Cross-Platform Signal Integration: Given the video's reported origination on TikTok, Meta must establish a policy to better integrate open-source intelligence and cross-platform threat signals. If a piece of content is identified as synthetic or deceptive on a peer platform (e.g., via open-source fact-checker networks), that signal must automatically elevate the content's priority score for human review on Meta’s platforms.

The Haifa video was a clear warning of how easily generative AI can be integrated into geopolitical IO. The platform's response to this synthetic content must be fundamentally restructured to prioritize the immediate de-escalation of conflict, which necessitates classifying fabricated evidence of attacks as a direct threat to safety

Name

Pavlo Burdiak

Organization

Centre for Democracy and Rule of Law (CEDEM)

Country

Ukraine

Language

English

CEDEM's submission to Meta Oversight Board case addressing AI-generated content in Israel-Iran conflict

This comment addresses Meta’s handling of a likely AI‑generated video depicting alleged war damage in Haifa during the June 2025 Israel-Iran conflict. It focuses on AI‑generated mis/disinformation in armed conflict, coordinated inauthentic behavior, engagement farming, and recommendation systems.

The comment draws on research in platform governance, disinformation, and information warfare, including experience from the Russian aggression against Ukraine, where similar tactics have been deployed and refined. It follows the structure: challenge – rationale – suggested solution.

I. AI‑generated conflict disinformation and the "imminent physical harm" threshold

Challenge

Meta’s current Misinformation Community Standards, focused on content that "directly contributes to the risk of imminent physical harm" or to interference with political processes, do not adequately cover large‑scale AI‑generated war disinformation that erodes information integrity and civilian protection without a single, easily provable causal link.

Rationale

During the June 2025 conflict, AI‑generated videos and images of fictitious missile strikes, destroyed Israeli cities, and false battlefield outcomes circulated widely, including via pages presenting themselves as "live news". Similar patterns have been observed in the Ukraine context: deepfakes of leaders and famous news anchors, AI‑generated videos of soldiers from the frontline, fabricated battlefield footage, etc.

Individually, a single AI-generated video may not meet the strict "imminent physical harm" or the "interference with political processes" test. However, cumulatively, systematic campaigns of such content can:
- Mislead civilians about where it is safe or unsafe to move, shelter, or evacuate.
- Disrupt humanitarian planning and assessment.
- Undermine trust in authentic evidence of violations of international humanitarian law ("liar’s dividend").
- Distort public and diplomatic decision‑making about escalation, ceasefires, or intervention.

Under Article 19 ICCPR, restrictions on expression must be lawful, necessary, and proportionate to a legitimate aim (for example, protecting the rights and safety of others). The UN Guiding Principles on Business and Human Rights (UNGPs) require companies to identify and address foreseeable human rights impacts, with heightened due diligence in conflict‑affected areas. Meta’s own Corporate Human Rights Policy commits it to these principles.

When AI‑generated content is falsely presented as documentary evidence of ongoing hostilities, and circulates at scale in an active conflict, the risk to the rights to life, security, and access to reliable information for self‑protection is foreseeable, even if the harm is systemic rather than tied to a single incident. Proper measures should be taken to address this challenge.

Suggested solution

1. Create a conflict‑specific sub‑rule under the Misinformation standards for AI‑generated content that:
- Falsely depicts real‑world damage, attacks, or military operations in an ongoing armed conflict.
- Is presented as authentic or “live,” rather than clearly labeled as AI-generated, satire, or artistic.
- Is likely, in context, to influence civilian protection decisions, humanitarian access, or understanding of the conduct of hostilities.

2. Allow measures short of removal (labelling, warning screens, distribution reduction) where full removal is not strictly necessary under Article 19 ICCPR, but the cumulative risk is high. This aligns with a proportionate, rights‑based approach.

3. Explicitly recognize in policy that in armed conflict, the threshold for intervention may legitimately be lower for AI-generated “documentary‑style” content than for opinion or commentary, because of its evidentiary and operational character.

II. Weak integration between misinformation enforcement and coordinated inauthentic behavior

Challenge

Enforcement against AI‑generated misinformation and against CIB is not well-integrated. This allows engagement‑farming pages and CIB networks to systematically use AI‑generated content to build audiences and only later face enforcement, often long after the damage is done.

Rationale

In this case, the page self‑identified as a "news source", posted an AI‑generated war video, accumulated over 700,000 views, and was reported multiple times. Only later were accounts linked to the page disabled for "engagement abuse and inauthenticity". This is consistent with a broader pattern, seen on Meta services and documented by researchers:
- Pages are created or hijacked, initially posting benign or viral "engagement bait" (sensational AI art, unrelated viral clips) to gain large followings.
- Over time, they shift to political or conflict‑related narratives, including AI‑generated war content, conspiracies, or propaganda.
- Networks of such pages cross‑promote and generate artificial engagement to trigger recommendation systems.

Meta has separate frameworks for Misinformation, Manipulated Media / Made with AI, and Coordinated Inauthentic Behavior. However, engagement‑farming operations intentionally sit at the intersection of these areas:
- They depend on AI‑generated visual content (manipulative visual content).
- They rely on inauthentic engagement patterns (CIB).
- They often distribute misleading or false content (misinformation).

Under the UNGPs, Meta is expected to take a systemic view of its adverse human rights impacts, not just a content‑by‑content view. Network‑level harms by organized actors fall squarely within this responsibility.

Suggested solution

1. Build a joint "AI‑Misinformation + CIB" enforcement track that:
- Automatically cross‑checks:
Pages repeatedly reported for "misinformation" or "AI manipulation";
Pages detected as heavy users of AI‑generated imagery/video on sensitive topics (e.g. conflicts);
Pages or accounts with CIB‑type signals (rapid name changes, sudden shifts in topic, synchronized engagement patterns, shared admins, or shared IP/hosting infrastructure).
- Triggers network‑level investigations, not just review of individual posts.

2. Introduce page‑level risk scoring for manipulative visual content:
- Factors: new or recently renamed pages, abrupt content strategy shifts, repeated posting of unlabelled AI‑generated war content, and engagement patterns suggesting farming.
- Consequences: stricter distribution limits, higher priority for human review, and, where appropriate, removal or disabling of whole clusters of pages/accounts.

3. Revise and clarify the "Manipulated Media" part of the Misinformation policy so that it explicitly covers systematic use of AI‑generated content as a growth and manipulation strategy by pages presenting themselves as news or "live updates".

III. Inadequate labeling, reporting, and user empowerment for AI‑generated content

Challenge

With the emergence of widely accessible generative AI services that produce highly realistic images and video, users can no longer reliably distinguish AI‑generated from authentic content, which becomes particularly relevant in times of conflict. This problem is compounded by limited, poorly targeted tools to report suspected AI manipulation, and by AI labels that are too generic and insufficiently connected to enforcement and user workflows.

Rationale

Meta has begun labeling some content as "Made with AI", following Oversight Board recommendations on manipulated media and Meta’s recent policy adjustments. However:
- Labels are not consistently applied or clearly differentiated (e.g. harmless artistic AI vs. deceptive AI-generated war video).
- Users lack a specific, intuitive reporting option for “AI manipulation” or “suspected AI‑generated war footage.”
- Reports of AI‑related issues may be merged into broader “misinformation” categories, slowing prioritization and obscuring patterns.
- Research suggests labels alone have limited impact unless paired with friction and context.

From a human rights perspective, the right to seek, receive, and impart information (Article 19 ICCPR) implies that users should be given tools to understand the nature of content they are shown, especially in high‑risk contexts like armed conflict.

Suggested solution

1. Introduce a dedicated "AI manipulation / suspected AI‑generated content" report category:
- Display clearly in the reporting user interface, with examples (e.g. "suspected AI‑generated war footage presented as real").
- Route these reports to a specialized triage queue with higher priority during conflicts or coming from conflict areas.

2. Strengthen and differentiate labels with context‑specific naming. Distinguish between and label clearly:
- "Creative AI" for benign/artistic uses (art, filters, obvious stylization);
- "AI‑Enhanced" for AI‑assisted edits of authentic content;
- "AI‑Generated" (with warning icon) for synthetic content presented as realistic, especially in sensitive topics (war, elections, disasters, health claims).
For the latter, use stronger visual cues and, where appropriate, interstitial warnings before resharing.

3. Pair labels with friction:
- When content is highly likely to be AI‑generated and depicts real‑world harm in an ongoing conflict, show a warning and require an extra click to share.
- Briefly explain that AI tools can fabricate realistic footage and encourage users to consult trusted sources.

4. Provide feedback loops to reporters:
- Inform users who report "AI manipulation" whether the content was confirmed as AI-generated and what action was taken (removed, labeled, demoted).
- This encourages high‑quality reporting and builds trust in enforcement.

4. Recommendation algorithms amplifying AI‑generated manipulative content

Challenge

Even passive interaction (simply viewing) with AI‑generated (including war-related) content can result in users being shown more of the same, because recommendation systems weight viewing, watch‑time, and engagement without adequately accounting for misinformation or synthetic nature.

Rationale

Meta’s shift toward "discovery‑first" feeds (Reels, algorithmic recommendations, suggested content) means:
- Users are increasingly shown content from pages and accounts they did not choose to follow.
- Engagement‑optimized ranking rewards content that provokes strong reactions – including sensational, fear‑inducing artificially generated war videos.
- Viewing, not only active engagement, can be treated as a positive signal, leading to more similar content in the feed.

Empirical research on recommender systems shows that engagement‑based ranking can systematically amplify misinformative and emotionally charged content, even when this conflicts with platform policies. This directly conflicts with Meta’s stated commitments to reduce the spread of misinformation and to respect users’ rights to a safe and trustworthy information environment.

Suggested solution

1. Audit the impact of recommendation systems on AI‑generated and misinformative content:
- Measure how much additional reach suspicious or policy‑violating content gains from recommendation algorithms compared to pure organic reach.
- Disaggregate for conflict‑related AI‑generated content.

2. Adjust ranking signals in sensitive domains:
- Down‑rank content from pages with a history of repeated misinformation, unlabelled AI‑generated war content, or CIB signals.
- Reduce the weight of passive signals (view time, autoplay) for high‑risk categories.

3. Exclude or severely limit high‑risk content from recommendations while under review:
- Content flagged by AI or user reports as likely AI‑generated conflict footage should be excluded from recommendations until verified.
- If verified as misleading, permanently exclude it from recommendation surfaces and apply stronger demotion to similar content sources.

4. Publish high‑level results of these audits to demonstrate compliance with Meta’s own policies and with systemic risk reduction duties under emerging regulatory frameworks (e.g. EU Digital Services Act’s systemic risk obligations, including on disinformation and AI‑generated content).

V. Fact-checker resource constraints and capacity gaps during conflict periods

Challenge

Fact-checkers — both those partnered with Meta and independent organizations — face severe capacity limitations during high-volume conflict disinformation periods. As the case shows, content was referred to fact-checkers but received no rating, suggesting either resource constraints, prioritization failures, or coordination breakdowns. These capacity gaps directly undermine Meta's ability to moderate AI-generated war content effectively, regardless of policy clarity.

Rationale

In this case, Meta's automated classifier flagged the video, referrals to third-party fact-checkers occurred, but no rating was issued — the content remained unverified and on the platform. This reveals a critical bottleneck: fact-checker capacity.

Research and direct testimony from fact-checkers globally document systemic resource constraints.

In January 2025, Meta announced the termination of its formal third-party fact-checking partnerships in the United States and an end to political content fact-checking. While Meta stated that fact-checking would continue outside the U.S. and that enforcement of its Misinformation community standards remains unchanged, the program restructuring creates additional uncertainty and likely coordination challenges.

At the same time, verifying AI-generated content requires technical skills, access to specialized detection tools, and familiarity with which generative models produce which artifacts. This is significantly more resource-intensive than fact-checking traditional false claims, yet most independent fact-checkers lack dedicated AI analysis capacity.

Under international standards — including the UN Special Rapporteur on freedom of expression, the ICRC's analysis of information warfare, and the UNGPs — platforms are expected to maintain moderation capacity proportionate to the scale of harms. During armed conflict, where misinformation can directly and/or systematically contribute to civilian harm, adequate capacity is not a discretionary enhancement but a rights-based necessity.

Suggested solution

1. Meta must maintain or expand dedicated fact-checking partnerships specifically for high-priority conflicts:
- Maintain and enhance (where applicable) relationships with regional fact-checkers covering Ukraine, Middle East, and other conflict-prone regions.
- Establish minimum response time commitments: for AI-flagged conflict content, fact-checkers should provide a preliminary assessment within 48 hours (e.g., "likely AI-generated," "insufficient evidence," "verified false").
- Full fact-checks can follow within 7 days, but initial rapid assessments unlock immediate enforcement decisions.

2. Increase Meta's internal capacity for AI-generated content verification:
- Hire or train specialists in AI detection tools, generative model artifacts, and multimedia forensics — with geographic focus on conflict-affected regions.

3. Create a coordinated "Conflict Verification Network":
- Convene fact-checkers, academic researchers, OSINT analysts, and humanitarian organizations into a pre-crisis coordination mechanism.
- Share detection models, training resources, and information sources to reduce duplication of effort.
- Establish shared infrastructure (shared database of verified claims, reusable video metadata analysis tools) to lower barriers for smaller fact-checking organizations.

4. Allocate sufficient resources proportionate to the risk:
- Conflict zones and crises demand more fact-checking resources, not fewer. Meta should consider providing direct grants to fact-checkers working on AI-generated conflict content, recognizing that this is a newly emerging and resource-intensive specialization.

VI. Transparency and human rights due diligence

Challenge

Without granular crisis‑specific transparency and formal human rights impact assessments, neither affected communities nor independent experts can evaluate whether Meta’s responses to AI‑generated war content are effective and rights‑respecting.

Rationale

Standard transparency reports are too high‑level to answer key questions:
- How fast are high‑risk AI‑generated conflict posts being flagged, reviewed, and acted upon?
- How many such posts are demoted, labeled, or removed?
- What is the false positive/false negative rate for AI detection on conflict content?
- How often are engagement‑farming networks involved?

Under the UNGPs, Meta is expected to carry out ongoing human rights due diligence and to communicate how it addresses salient risks. In armed conflict, this due diligence must be heightened. The UN Special Rapporteur on freedom of expression has also emphasized the need for transparency about disinformation‑related measures, especially when they affect access to information during crises.

Suggested solution

1. Publish conflict‑specific AI‑content transparency data, including:
- Number of posts flagged as likely AI‑generated in a given conflict;
- Time from flagging/report to human review and to enforcement action;
- Breakdown of enforcement outcomes (labels, demotion, removal, account or network action).

2. Conduct and publish human rights impact assessments focused on AI‑generated mis/disinformation in armed conflict:
- Assess impacts on the rights to life, security, access to information, and accountability for violations of international humanitarian law.
- Include perspectives from affected communities, journalists, fact‑checkers, and civil society in conflict‑affected countries, including Ukraine.

Name

Sam Stockwell

Organization

Centre for Emerging Technology and Security, The Alan Turing Institute

Country

United Kingdom

Language

English

Attachments

Meta-Oversight-Board-Submission-Nov-2025.pdf

Name

Wade Robertson

Country

Canada

Language

English

FACEBOOK NEEDS TO DO A LOT MORE TO ADDRESS A.I. CONTENT IN EVERY SITUATION, NOT JUST IN CONTENT GENERATED BECAUSE OF THE IRAN-ISRAEL CONFLICT. FACEBOOK NEEDS TO DO A 100% MORE EFFORT IN ADDRESSING SEX AND VIOLENT CONTENT ON FACEBOOK POSTS THAT PEOPLE ARE POSTING. I HAVE PERSONALLY SEEN HUNDREDS OF EXTREMELY GRAPHIC, VIOLENT VIDEO'S ON FACEBOOK; VIDEOS THAT SHOW MURDER AND PEOPLE BEING CRUSHED TO DEATH WHILE A VEHICLE HAS DRIVEN OVER THEM. EVERY VIDEO I HAVE SEEN HAS BEEN REPORTED. IN MANY OF THEM FACEBOOK HAS RESPONDED BAVK TO ME AND SAID WE'VE REMOVED THE PROFILE BECAUSE THEY VIOLATED YOUR STANDARDS.

IN FAR TOO MANY THEY HAVE BEEN ALLOWED TO STAY UP BECAUSE YOU SAID THEY DIDN'T VIOLATED YOUR STANDARDS OF USE. IN MANY OF THOSE VIDEOS, I HAVE APPEALED YOUR DICISION AND I GUESS AN ACTUAL SET OF HUMAN EYES VIEWED THE VIDEO AND THE VIDEO/ACCOUNT WAS REMOVED.

FACEBOOK IS RELYING WAY TOO MUCH ON AI TO SCREEN FOR HARMFUL CONTENT AND AI IS FAILING MISERABLY AT IT. YOU HAVE TO HAVE A MUCH BETTER SYSTEM SET UP TO SCREEN OUT THE GRAPHIC VIDEOS THAT SHOW SEXUALITY, VIOLENCE AND DEATH; KIDS DON'T NEED TO SEE THAT SHIT ON FACEBOOK.

FACEBOOK NEEDS TO MAKE THE ACTUAL PROCESS OF REPORTING SEX, VIOLENCE AND DEATH A LOT MORE EASIER FOR PEOPLE LIKE ME TO REPORT. WITH FAR TOO MANY VIDEOS AND PICTURES I DON'T EVEN PURSUE REPORTING OR APPEALING YOUR DECISION BECAUSE ITS CLEAR TO ME THAT HUMAN EYES ARE NOT VIEWING THE REPORT. I CAN'T BE THE ONLY PERSON ON FACEBOOK WHO IS FRUSTRATED WITH NOT BEING ABLE TO REPORT VIOLENT CONTENT IN A TIMELY MANNER.

THIS IS MY SECOND ATTEMPT AT BEING A PART OF THE FACEBOOK COMMUNITY. WITH MY PREVIOUS PROFILE, I MADE TOO MANY CONTROVERCIAL POSTS AND THE PROFILE WAS DEACTIVATED. I DON'T AGREE WITH YOUR DECISION TO DEACTIVATE MY PREVIOUS PROFILE BUT I UNDERSTAND IT. FACEBOOK SUGGESTED TO ME TO OPEN A NEW ACCOUNT EHICH I DID. YOU ALSO SAID I'D BE ABLE TO RECOVER MY WHOLE FRIENDS LIST SO I WOULDN'T HAVE TO RECREATE IT, BUT I CAN'T; I FOLLOWED YOUR INSTRUCTIONS AND NOTHING HAPPENED. I DON'T CARE ABOUT RECOVERING MY POSTS BUT, I WOULD LIKE TO RECOVER ALL MY PHOTOS FROM MY PREVIOUS ACCOUNT. FACEBOOK DOESN'T APPEAR TO HAVE THIS OPTION ON THIS PLATFORM AND THAT NEEDS TO CHANGE IMMEDIATELY.

YOU WANT FACEBOOK USERS TO BE A PART OF THE PROCESS OF KEEPING THE PLATFORM SAFE FOR EVERYONE BUT YOU ARE NOT IMPLEMENTING ANY DECENT PROCEDURES THAT WILL ALLOW THAT TO HAPPEN. YOU NEED TO USE AI PROPERLY TO SCREEN OUT DISTRUCTIVE, VIOLENT CONTENT; YOU ALSO NEED TO LIMIT ITS USE ENOUGH, IN ORDER FOR PEOPLE LIKE ME TO REPORT DANGEROUS VIOLENT CONTENT EFFECTIVLY AND PROMPTLY.