Meta Urged to Improve Oversight of Fake AI Videos After Haifa Case
Fake AI videos are becoming harder to separate from genuine footage just as social media users increasingly depend on smartphones for information about wars, natural disasters and political crises. A 23-second Facebook clip falsely claiming to show extensive destruction in Haifa has now become a case study in how Meta’s existing labelling system can fail when accurate information matters most.
The fabricated video was posted during the June 2025 conflict between Israel and Iran by a Facebook page presenting itself as a news source. It showed damaged buildings accompanied by the words “Live now – Haifa Towards Down” and attracted more than 700,000 views.
At least six users reported the post. A similar version had already been identified as false and AI-generated by an independent fact-checker. Meta did not review it, send it to a fact-checking partner or attach a prominent warning.
Meta’s Oversight Board later ruled that the video should have carried a “High Risk AI” label because it could materially mislead people about a major armed conflict. The Board did not say it should automatically have been removed. Instead, it argued that users needed clear information showing that the footage was synthetic.
The ruling was only the beginning. Meta responded in May 2026 but declined the Board’s central recommendation to create a separate Community Standard governing AI-generated content. The Board subsequently said the company needed to move faster and take bolder action.
The dispute now goes far beyond one false clip. It asks whether Meta can reliably identify fake AI videos at the speed and scale at which Facebook, Instagram and Threads distribute them.
The 23-Second Video That Exposed a Larger Failure
The video was posted on Facebook on 15 June 2025, two days into the 12-day Israel-Iran conflict. It purported to show severe damage to buildings in Haifa after an Iranian attack.
The account behind the page was located in the Philippines, despite presenting the page as a source of breaking conflict news. The caption contained numerous headline-style phrases, unrelated terms and hashtags designed to attract attention.
Several details should have raised concern. A closely related version had originated on TikTok and had already been debunked. Comments beneath the Facebook post also warned that the footage appeared to have been created with artificial intelligence.
Despite these signals, the fake AI videos review process failed at several points.
| Warning signal | Meta’s response |
|---|---|
| Page claimed to provide conflict news | Post remained available |
| Similar footage had been debunked | No fact-check was attached |
| Six users reported the Facebook post | Reports did not trigger a review |
| Comments identified the content as AI | No “AI Info” label appeared |
| Video concerned an active armed conflict | No “High Risk AI” label was added |
| Page showed suspicious engagement activity | Action came only after Board involvement |
Meta told the Oversight Board that the post did not violate its misinformation rules because it did not directly contribute to an imminent risk of physical harm or violence.
The Board accepted that the threshold for removal had not been met. It rejected the conclusion that no action was needed.
Why the Board Wanted a Label Rather Than Removal
The distinction between labelling and deleting fake AI videos is essential.
Not every AI-generated image or recording is harmful. Synthetic content can be used for art, satire, education, entertainment and political expression. Removing everything made with AI would restrict legitimate speech and produce countless moderation errors.
The question is whether a reasonable user could mistake the content for authentic evidence about an important matter.
The Haifa clip was designed to look like live news footage from a war zone. It did not announce itself as a simulation, artwork or parody. Its caption presented the destruction as real and current.
The Oversight Board concluded that it created a material risk of misleading the public at a critical moment. A prominent “High Risk AI” label would have preserved the post while warning viewers that the apparent evidence was not genuine.
This approach reflects Meta’s stated policy of providing context for manipulated media rather than automatically deleting it. Meta says it may attach a more prominent label when digitally created or altered media poses a particularly high risk of deceiving the public about a matter of importance.
In the Haifa case, that policy existed but was not used.
Six User Reports Led Nowhere
The most troubling detail may not be that Meta’s automated systems failed to recognise the footage. AI detection remains technically difficult, especially when videos have been downloaded, compressed, edited or stripped of metadata.
The deeper problem is that six people reported the post and none of those reports resulted in a review.
Meta told the Board that the content was not checked by its third-party fact-checking partners. It was also not escalated to the internal content-policy team that can authorise prominent AI labels.
The fake AI videos system therefore depended on pathways that the post never entered. The user who eventually appealed to the Oversight Board had to go outside Meta’s normal moderation process to obtain meaningful scrutiny.
Once the Board selected the case, Meta investigated the page more deeply. The company then disabled three accounts linked to it for engagement abuse and inauthentic behaviour, which removed the page and its content.
The post disappeared because Meta acted against the network behind it—not because the original reports caused the video to be recognised and labelled as artificial.
That sequence suggests that Meta possessed tools capable of detecting suspicious account behaviour but failed to connect those signals to the misleading video early enough.
Fake AI Videos Spread Faster During Information Vacuums
Conflict creates ideal conditions for synthetic misinformation.
People want immediate visual proof of missile strikes, damaged military facilities and civilian casualties. Journalists and fact-checkers may have limited access. Governments restrict information, internet services become unreliable and authentic videos arrive without complete context.
Fake AI videos can fill that gap within minutes.
They may exaggerate battlefield success, invent attacks that did not occur or create the impression that one side has suffered catastrophic damage. Accounts spreading the material can earn advertising revenue, attract followers or advance a government’s propaganda objectives.
The Oversight Board described the 2025 Israel-Iran conflict as an inflection point for deceptive generative AI. It cited reporting that three fabricated videos connected with the fighting accumulated more than 100 million views.
The problem intensified again during the 2026 Iran conflict. The Associated Press documented fabricated images and videos distributed by state-linked networks and opportunistic accounts, including false footage claiming to show successful missile attacks and burning buildings.
The News Ink’s reporting on AI deepfakes used in Russian disinformation shows that the same tactic is not limited to one region. Cheap generative tools allow governments, political groups and individuals to produce convincing visual propaganda at unprecedented speed.
Meta’s Current Labelling System Has Three Weak Points
Meta’s approach to fake AI videos rests largely on three signals.
First, the platform looks for technical metadata or industry-standard indicators showing that a file was created with an AI system.
Second, users uploading realistic synthetic video or audio are required to disclose that it was digitally created or altered.
Third, difficult cases can be escalated to specialist teams capable of attaching stronger warnings.
Each method has weaknesses.
Metadata can disappear
AI tools may attach Content Credentials or other metadata when a file is generated. That information can identify the software used and record how the file was created or edited.
However, metadata can disappear when a video is recorded from another screen, processed through an editing application or downloaded and reuploaded. Some AI generators do not attach useful provenance information in the first place.
Self-disclosure depends on honesty
A person creating fake AI videos to mislead viewers has little incentive to admit using artificial intelligence.
Self-disclosure may work for responsible creators sharing art or entertainment. It is far less reliable for propaganda, scams or engagement farming.
Escalations are too rare
The “High Risk AI” label is reserved for content capable of materially deceiving users about important subjects. Applying it generally requires special review.
The Oversight Board said a system that relies heavily on self-disclosure and infrequent escalations cannot handle the quantity and speed of modern AI-generated media.
What Meta’s “AI Info” Label Actually Means
Meta began expanding its AI labels in 2024. The company initially used “Made with AI” but changed the wording to “AI Info” after photographers and other users complained that minor editing tools could cause genuine images to receive misleading labels.
Meta now distinguishes between content generated by an AI system and material that was only edited or retouched using AI features. The most visible label is generally reserved for generated media, while information about minor editing may appear within a post’s menu.
That refinement addressed one problem: not all uses of AI are equivalent.
It did not solve detection.
A careful label on content Meta successfully identifies offers no protection when fake AI videos arrive without readable technical signals and the uploader refuses to disclose how they were made.
The Haifa case demonstrated that Meta’s policy can appear comprehensive on paper while leaving a deceptive wartime video unlabelled after hundreds of thousands of views.
Seven Changes the Oversight Board Recommended
The Board issued seven policy and enforcement recommendations intended to strengthen Meta’s response.
| Recommendation | Intended improvement |
|---|---|
| Proactive crisis review | Reduce dependence on alerts from outside organisations |
| Separate AI Community Standard | Put all AI rules in one clear public policy |
| Publish penalties | Explain consequences for failing to disclose synthetic media |
| Display Content Credentials | Let users inspect reliable provenance information |
| Improve detection | Identify AI-generated video, audio and images more accurately |
| Watermark Meta AI output | Ensure Meta’s own creations can be identified elsewhere |
| Expand high-risk labels | Apply prominent warnings at significantly greater scale |
These proposals treat fake AI videos as both a technical and governance problem.
Better classifiers alone will not solve unclear policies. Clearer rules alone will not identify stripped metadata. Labels alone will not stop abusive networks that repeatedly publish synthetic propaganda.
The recommendations were therefore designed to work together: detect the content, preserve its history, display useful information, investigate the account and respond quickly during crises.
Meta Rejected a Separate AI Community Standard
Meta responded publicly on 8 May 2026, within the required 60-day period.
The company declined the recommendation to create a separate Community Standard specifically for AI-generated content. It argued that the relevant requirements could remain distributed across its existing misinformation and manipulated-media policies.
The Oversight Board called that decision a missed opportunity for structural clarity.
A separate standard would allow users, creators and researchers to find the company’s full approach in one place. It could define the difference between ordinary AI content, deceptive synthetic media and content serious enough to warrant removal.
It could also explain:
- when users must disclose artificial creation or editing;
- which technical signals Meta recognises;
- how “AI Info” differs from “High Risk AI”;
- whether a labelled post is demoted in recommendations;
- what happens to repeat offenders;
- how rules change during wars or elections;
- how creators can appeal an incorrect label.
Without a unified policy, fake AI videos remain governed through several overlapping documents and internal procedures that ordinary users are unlikely to understand.
Meta Accepted Some Ideas but Left Others Under Review
Meta told the Board that one recommendation was already incorporated into its crisis processes. The company said it uses internal expertise and “Pre-Reviewed Harmful Claims” to prepare enforcement around foreseeable events.
The Board considered this only a partial answer because conflicts evolve quickly. A misleading claim that did not exist yesterday can become central to a propaganda campaign today.
Meta also said it was partially implementing work involving stronger Content Credentials, better detection, improved provenance information and more durable watermarks.
Two other recommendations remained under feasibility assessment: expanding the use of high-risk labels and publishing clearer penalties for users who fail to disclose digitally created or altered content.
The Board warned that feasibility reviews can remain open for years. It asked Meta to provide clearer implementation timelines rather than leaving major protections in an indefinite assessment stage.
The fake AI videos problem is already active. A response that may become effective after another war or election offers limited protection now.
Content Credentials Can Help, but They Are Not a Truth Machine
The Board placed significant emphasis on the Coalition for Content Provenance and Authenticity, generally known as C2PA.
Its Content Credentials standard can attach a verifiable history to a digital file. That history may identify the device or application that created it and show whether it was edited later.
For users, Content Credentials could function like an ingredient label. Rather than asking an unreliable detector to decide whether a video “looks AI-generated,” a platform could inspect signed information about the file’s origin.
The approach has important limitations.
Credentials prove what happened to a file only when the relevant tools attach and preserve the information. They do not automatically prove that the scene itself is true. A genuine camera can record a staged event, while a legitimate editor can remove important context without using generative AI.
Bad actors can also strip credentials by taking screenshots, recording screens or passing content through unsupported software.
Content provenance is therefore one layer of defence, not a complete answer. It works best alongside forensic detection, fact-checking, account investigation and clear labels.
Meta Was Criticised Over Its Own AI Output
The Oversight Board expressed concern that Meta was not consistently attaching and preserving Content Credentials even on material created using its own AI tools.
That criticism is especially serious because Meta controls both sides of the process. It develops generative products and operates the platforms where their output may be distributed.
The Board cited testing suggesting that only some Meta-generated images and videos carried credentials that could be detected and labelled reliably.
If a company cannot consistently identify material created by its own systems, it will have even greater difficulty recognising fake AI videos produced by thousands of outside applications.
The Board recommended that Meta add provenance information and invisible watermarks at the point of creation, then publish evidence showing how consistently those signals survive when content is shared.
One False Video Can Become Hundreds of Copies
The Haifa footage appears to have originated on TikTok before spreading to Facebook, Instagram and X.
That cross-platform movement exposes another weakness in content moderation. A fact-check or label applied on one service does not automatically follow the video elsewhere.
Users can crop the frame, reverse the image, change the soundtrack, add text or shorten the clip. Each variation may appear different enough to avoid an exact-match system while preserving the same false claim.
Meta and other platforms can respond through similarity matching. Once a reliable fact-check identifies one fake, automated systems can look for closely related copies and apply the same warning or reduce distribution.
The Oversight Board said broader matching could have limited the Haifa video’s reach.
Platform cooperation is also essential. The News Ink’s analysis of how online content can influence AI answers shows that false information can move between websites, social networks and AI systems rather than remaining isolated on one service.
The Business Model Rewards Sensational Fakes
The Facebook page behind the Haifa post had been eligible to earn money through Meta’s Stars programme.
That detail matters because fake AI videos are often commercial products as well as propaganda.
Synthetic footage of an ordinary street may attract little attention. A video claiming that missiles destroyed a major city can generate enormous engagement. Views become followers, advertising impressions or direct payments.
The account benefits from speed. Publishing first can be more valuable than publishing accurately, especially when a platform’s recommendation system rewards emotional reactions and rapid sharing.
Meta eventually removed the network for engagement abuse and inauthenticity. The intervention came only after the content had attracted more than 700,000 views and the Board had selected the appeal.
A stronger system would connect monetisation with integrity. Accounts repeatedly publishing undeclared synthetic news footage should not continue earning while their content waits for a slow review.
The News Ink’s report on Meta’s advertising and accountability dispute examines a related question: whether the company’s commercial decisions consistently match its public claims about safety.
Meta’s Wider Moderation Retreat Adds to the Concern
The warning about fake AI videos arrived after Meta had already reduced some of its professional misinformation infrastructure in the United States.
In 2025, the company ended its US third-party fact-checking programme and began replacing it with Community Notes. Meta argued that the previous system had produced excessive censorship and that crowdsourced context could operate at a greater scale.
Community Notes may add useful information to contested posts, but they require contributors to write, evaluate and agree on a note. Fast-moving conflict footage can gain millions of views before a suitable note reaches consensus.
The Oversight Board has separately warned that Community Notes should not simply be treated as a universal substitute for professional fact-checking, particularly in countries facing conflict, repression or limited access to reliable information.
Meta’s moderation strategy therefore contains a tension. It wants to reduce centralised judgments about truth while fake AI videos increasingly require rapid expert verification, technical analysis and coordinated enforcement.
Community participation can assist that work. It cannot replace every specialist function.
The “Liar’s Dividend” Damages Real Evidence Too
The greatest danger is not limited to people believing false videos.
As synthetic media becomes common, governments and other powerful actors can dismiss genuine footage as artificial. Researchers describe this as the “liar’s dividend”: the existence of deepfakes creates plausible deniability for authentic evidence.
A real video showing civilian casualties may be rejected as AI. A genuine recording of a politician can be dismissed as manipulated. Verified images from a war zone can lose their power because viewers no longer trust anything on screen.
The Oversight Board warned that fake AI videos contribute to general distrust of information. WITNESS, a human-rights organisation specialising in visual evidence, has similarly argued that platforms must address both sides of the problem: synthetic content accepted as genuine and authentic documentation falsely dismissed as fake.
The News Ink’s feature on why proving you are human is becoming harder explores how this erosion of trust affects people far beyond military conflicts.
Labels Must Be Visible Enough to Matter
A label that users never notice offers little protection.
Meta’s standard “AI Info” notice may appear near the content or within a menu depending on how the media was created or edited. The “High Risk AI” label is intended to be more prominent because the potential for serious deception is greater.
The Oversight Board wants Meta to explain these distinctions more clearly. Users should understand whether a label means:
- an uploader voluntarily disclosed AI use;
- technical metadata identified a generative tool;
- a fact-checker found the content false;
- Meta’s specialists judged it likely to deceive;
- the video contains minor AI editing rather than full generation.
Fake AI videos do not all carry the same risk. A transparent fantasy scene and fabricated evidence of a missile attack should not receive indistinguishable treatment.
The label must also appear before a person shares, comments on or reacts to the post—not after the false claim has already travelled across the platform.
Removal Should Remain Reserved for Greater Harm
Some critics may believe Meta should simply delete every deceptive AI war video. The Oversight Board did not adopt that position.
Removing misleading material can be justified when it contributes to imminent violence, fraud, voter suppression or another serious violation. But misinformation policies also affect journalism, satire, political advocacy and documentation of propaganda.
Keeping a fabricated video available with a clear warning can sometimes support research and public understanding. It allows journalists and educators to show how information operations work.
The appropriate response should depend on context:
| Type of content | Possible response |
|---|---|
| Clearly disclosed AI art | No intervention or ordinary AI label |
| Misleading but low-risk synthetic media | AI label and contextual information |
| Fabricated crisis footage | Prominent warning and reduced recommendations |
| Coordinated inauthentic network | Account investigation and monetisation limits |
| Content creating imminent physical danger | Removal under applicable safety rules |
The objective is not to erase artificial media. It is to prevent synthetic evidence from being mistaken for reality.
What Meta Needs to Change Before the Next Crisis
The Haifa case points towards a practical enforcement model.
Meta should activate dedicated synthetic-media teams whenever a major conflict, election or disaster begins. Those teams should include language experts, regional analysts and digital-forensics specialists.
Once a fake is verified, matching technology should locate altered copies across Facebook, Instagram and Threads. Prominent labels should appear automatically, and the accounts distributing those copies should be reviewed for coordinated or commercial abuse.
Creators using Meta AI should receive durable Content Credentials by default. Uploaders who intentionally remove provenance or repeatedly ignore disclosure requirements should face clear penalties.
Meta should also publish quarterly information showing:
- how many “High Risk AI” labels were applied;
- how quickly reported content was reviewed;
- which languages and regions received coverage;
- how many repeat offenders lost monetisation;
- how often labels were appealed or corrected;
- whether labelled content was removed from recommendations.
Transparency would allow the public to judge whether Meta’s fake AI videos response is improving or simply generating new policy language.
What Users Can Do When a Video Looks Suspicious
Platforms carry the greatest responsibility, but users can reduce the spread of fabricated footage.
Pause before sharing dramatic clips from unfamiliar accounts. Search for the same event through recognised news organisations. Check whether buildings, vehicles, shadows or human movement appear physically inconsistent.
Look for an “AI Info” label, but do not assume the absence of one proves authenticity. Meta’s own case shows why that assumption is unsafe.
Inspect the account. A page that recently changed its name, publishes unrelated breaking stories or uses excessive hashtags may be seeking engagement rather than reporting accurately.
Reverse-image searches and verification projects can also help. The News Ink’s cybersecurity guide offers broader advice for checking unfamiliar accounts, links and online claims.
Most importantly, avoid reposting suspicious footage merely to ask whether it is fake. Sharing can amplify it even when the caption expresses doubt.
The Oversight Board’s Power Has Clear Limits
Meta created the Oversight Board to make independent decisions on difficult content cases.
The Board can issue binding decisions about whether specific eligible content should remain available or receive different treatment. Its wider policy recommendations are not binding. Meta must respond publicly, but it can decline them.
That limit is visible in this case. The Board ordered a different outcome for the Haifa post and proposed seven broader reforms. Meta rejected the recommendation for a dedicated AI standard and left several other proposals only partly implemented or under review.
The Board can expose systemic weaknesses. It cannot redesign Meta’s systems by itself.
Real change still depends on whether Meta commits engineering resources, enforcement staff and product space to measures that may reduce engagement or slow the distribution of viral content.
The Bottom Line
Fake AI videos have moved from an emerging technology problem to a central challenge for public trust.
The Haifa case showed how one fabricated clip could collect more than 700,000 Facebook views despite user reports, prior fact-checking of similar footage and obvious indications that the page was not a reliable news source.
Meta’s Oversight Board concluded that the post did not meet the threshold for removal but should have received a prominent “High Risk AI” label. It also called for a separate AI Community Standard, stronger detection, visible Content Credentials, improved watermarks and far greater use of high-risk warnings.
Meta accepted or partially accepted parts of that programme but rejected the standalone policy. In May 2026, the Board responded that partial measures were not enough and urged the company to act faster.
The decision matters because the volume of fake AI videos is increasing while Meta reduces its dependence on professional fact-checkers and places greater weight on user disclosure, automated signals and Community Notes.
None of those tools is sufficient alone.
Reliable oversight requires technical provenance, forensic detection, human expertise, account-level investigations and labels visible enough to influence users before they share a post.
The alternative is not simply more people believing fabricated footage. It is a world in which authentic evidence also loses credibility because every inconvenient video can be dismissed as artificial.
Meta operates some of the world’s most influential information platforms and develops its own generative AI products. It is therefore not an outside observer of the problem. It controls important parts of both the creation and distribution system.
The Haifa video gave the company a clear warning. The next crisis will show whether Meta used it.
For more AI, online-trust and misinformation coverage, follow The News Ink on Medium or read our reports on AI disinformation campaigns and the growing deepfake trust crisis.
