AI Generated Video in Iran-Israel conflict

UPDATED MAY 8, 2026
2026-004-FB-UA
Today, November 18, 2025, the Oversight Board selected a case appealed by a Facebook user regarding a 23-second video depicting alleged damage to buildings in Haifa, Israel, during the 12-day conflict between Israel and Iran in June 2025. The video includes a text overlay that states, “Live now - Haifa.” and has a caption in English headline-style phrases linked to the conflict as well as disjointed terms and hashtags, without following a clear narrative. The video appears to be the same as one that was identified by independent fact-checkers as AI-generated.
Meta determined that this content did not violate our policies on Misinformation, as laid out in the Community Standards, and left the content up. Under our Misinformation policy, Meta removes “misinformation where it is likely to directly contribute to the risk of imminent physical harm” or interfere with the functioning of political processes. For other types of misinformation, we focus on reducing its prevalence or creating an environment that fosters a productive dialogue.
We will implement the Board's decision once it has finished deliberating, and will update this post accordingly. Please see the Board's website for the decision when they issue it.
Read the board's case selection summary
Case decision
We welcome the Oversight Board's decision today, March 10, 2026, on this case. The Board overturned Meta’s original decision to leave up the content without a High Risk AI label.
Meta will comply with the Board's decision within 7 days.
When it is technically and operationally possible to do so, we will also take action on content that is identical and in the same context. For more information, please see our Newsroom post about how we implement the Board’s decisions. After conducting a review of the recommendations provided by the Board, we will update this post with initial responses to those recommendations.
Recommendations
Recommendation 1 (Work Meta Already Does)
To ensure the swift review of misinformation that leads to risks of imminent physical harm or violence in crises, Meta should amend the Misinformation Community Standard to ensure that enforcement of this rule does not depend on signals from external partners. There should be a lever under the Crisis Policy Protocol to allocate resources for timely, proactive detection of such violating content, supported by in-house expertise, to identify, review and action content under the policy (including affixing labels under the Manipulated Media policy and investigating posting accounts and pages that show signals of engagement abuse).
The Board will consider this implemented when Meta updates its Misinformation policy to reflect these requirements for the Physical Harm or Violence category.
Our commitment: Meta will continue to enforce its Misinformation & Harm Policy only when an independent expert, possessing knowledge and expertise to assess the truth of content, tells us that a particular piece of content is false and is likely to directly contribute to the risk of imminent harm. However, as part of our current protocols and frameworks, we currently work with external independent experts ahead of critical events and crises to create Pre-Reviewed Harmful claims to help enforce our Misinformation & Harm Policy better at scale. While this is a separate process from the Crisis Policy Protocol, it is supported by in-house expertise. As such, we consider this Work Meta Already Does, and provide an explanation of this work in our considerations below.
Considerations: As the Board notes throughout their decision and as influenced by prior Board recommendations, we have a number of protocols and frameworks in place to respond to crisis situations as they unfold. These efforts include the Crisis Policy Protocol and a number of other enforcement and response processes that we explain in our Transparency Center. These response protocols work in tandem with our policies to balance our values of expression, safety, dignity, authenticity and privacy. For our Misinformation Community Standard specifically, our policies articulate different categories of misinformation and try to provide clear guidance about how we treat that content on the platform to align with these values.
As noted in our Community Standards, “[w]e remove misinformation where it is likely to directly contribute to the risk of imminent physical harm.” To determine whether content constitutes misinformation under this policy, “we partner with independent experts who possess knowledge and expertise to assess the truth of the content and whether it is likely to directly contribute to the risk of imminent harm.” However, our Misinformation policy and internal guidance includes “Pre-Reviewed Harmful Claims” (PHC), which are designed to serve as a proactive lever for at-risk markets and during certain critical events. The Crisis Policy Protocol is one criteria that informs which markets receive PHC designations along with input from Trusted Partners. When we make these designations, we work with Trusted Partners to determine whether any false claims made in the context of a crisis or ahead of a critical event may directly contribute to a risk of imminent physical harm. These designations are time-limited and still require a determination by an independent credible source that the content is making a PHC-designated false claim. This allows internal market teams to proactively identify and escalate potentially violating content for review by in-house experts who may consider removing content without additional outreach to a Trusted Partner, which may be impractical during a critical event or time of crisis. We consider these existing policies and processes as covering the Board’s recommendation and will not provide further updates.
To help advance trust in information on Meta’s platforms, Meta should create a Community Standard for AI-generated content, separate from the Misinformation Community Standard. The new Community Standard should provide comprehensive details on provenance preservation (i.e., capturing the detailed facts about the history of a piece of digital content), AI labeling protocols and self-disclosure rules.
The Board will consider this implemented when Meta publishes a new Community Standard specifically on AI-generated content.
Our commitment: We have a number of policies that cover how we address AI-generated content, and these policies build on the existing Community Standards. Updating our policies within the context of these existing structures and Community Standards frameworks assists in enforcement both at scale and on escalation. We also remain committed to sharing details about provenance preservation, AI-labeling protocols, and self-disclosure rules in our Transparency Center, including the details we’re sharing below. Ultimately, while creating a fully separate Community Standard on AI-generated content may not be the clearest way to explain our expectations to users, we will continue to consider additional ways to increase clarity on our approach to this type of content across our existing Community Standards, so that they can be easily accessed and understood.
Considerations: Each section of our Community Standards starts with a "Policy Rationale" that sets out the aims of the policy followed by specific policy lines that outline what content is not allowed, what content may require additional information or context to enforce, and what content may be allowed with a warning screen, age-gating, or a label.
AI-generated content is addressed under a number of our individual Community Standards. For example, under our Adult Sexual Exploitation policy, we may remove intimate imagery in certain contexts. We recently updated that policy to address AI-generated imagery rather than creating a standalone policy to separately address AI-generated imagery of this nature. This allows us to build on existing knowledge and enforcement structures to balance voice and safety considerations.
We recognize that AI is an evolving technology and we have made updates to our Community Standards to reflect this. We also understand that users are looking for greater transparency and more accessible information in this space. Accordingly, we have continued to update our Transparency Center and Community Standards previously with details about the ways that our Community Standards approach AI-generated content and other measures that we may take to address this content on our platforms. We do not expect to shift existing AI policies internally into a standalone Community Standard as this shift would require extensive changes to our internal enforcement approach.
While we will not implement this recommendation at this time, we will continue to consider ways to share more about our approach to AI-generated content throughout our Community Standards and Transparency Center as it evolves. We will have no further updates on this recommendation.
To improve the clarity of its rules, Meta should publish a clear explanation of penalties for failure to self-disclose digitally created or altered content. It should provide criteria for penalties and list which account features are consequently limited and for how long.
The Board will consider this implemented when Meta updates the Community Standard to include these penalty details and makes the revised guidance available in its public Transparency Center.
Our commitment: In our Community Standards, we note that we require people to use our AI disclosure tool whenever they post organic content with photorealistic video or realistic-sounding audio that was digitally created or altered. We also indicate that we may apply penalties if they fail to do so. We are exploring ways to improve the application of any potential penalties at scale and will assess how we can best share details about these penalties externally with our users.
Considerations: Our Transparency Center currently details our approach to strikes and penalties when users share content that violates our Community Standards. We separately may also apply penalties for users who repeatedly fail to disclose content with “AI Info” labels.
We are considering updates to our enforcement approach, and as part of those considerations will also assess how we can best communicate these updates with users. We will share updates in a future report to the Oversight Board.
To ensure users can reliably identify AI-generated content, Meta should implement Content Credentials (as laid out by the Coalition for Content Provenance and Authenticity) at scale and ensure that they are clearly and consistently visible and accessible to users whenever the provenance details are available. Provenance should not remain solely internally detectable or limited to back-end systems.
The Board will consider this implemented when Meta provides a report explaining the changes it made to its interfaces and products to ensure that Content Credentials are consistently and clearly shown to users when available.
Our commitment: We have built the technical foundation for detecting and surfacing Content Credentials at scale across our platforms and are committed to enhancing how we present content provenance details to users.
Considerations: We have invested significantly in building systems to detect, store, and show users information about how content is created. This includes reading Content Credentials (a digital record, based on the C2PA industry standard, that documents how a piece of content was made or edited) and AI origin tags (standardized labels embedded in image files by AI tools to indicate the content was generated by AI).
When we detect that content was created or edited using AI, we apply an "AI info" label, primarily across images, that users can see either directly on the post or by tapping the post's menu. This means Content Credentials already result in visible labels for users in our products across some content types today. We acknowledge, however, that there is meaningful work ahead in scaling our interventions across other content types so users can better distinguish between content that was fully created by AI and content that was only partially edited with AI tools, and how to ensure a consistent experience across Facebook, Instagram, and Threads.
While the core detection, ingestion, and labeling infrastructure is operational today, we are actively enhancing how this information is presented to users. We will provide the Board updates on the interface and product changes as we continue this ongoing work.
To improve detection and labeling accuracy, Meta should invest in stronger detection tools for AI-generated multi-format (audio, audio-visual and image) content. Tooling should support escalation teams to better identify generative AI content trends, including potential harms around deceptive content in crisis situations.
The Board will consider this recommendation implemented when the company confirms that stronger tools have been adopted and shares transparency data on the performance of these tools. These findings must be disaggregated by language and country, and whether the Crisis Policy Protocol was activated. This data must reflect comparable periods of time before and after the introduction of these changes.
Our commitment: We have made and continue to make significant investment in detection tools for AI-generated content across image, audio, and video formats. We are committed to strengthening these capabilities and will continuously maintain transparency around their performance. As generative AI technology consistently evolves, making it easier to create more convincing fakes, we are actively scaling our investments in detection to keep pace.
Considerations: We continue to work with industry partners to align on common technical standards that signal when a piece of content has been created using AI. Being able to detect these signals makes it possible for us to label AI-generated images that users post to Facebook, Instagram and Threads. We are building industry-leading tools that can identify invisible markers at scale – specifically, the "AI generated” information in the C2PA and IPTC technical standards – so we can label images from Google, OpenAI, Microsoft, Adobe, Midjourney, and Shutterstock as they implement their plans for adding metadata to images created by their tools. As companies start to include signals in their video generators, we are also working on detecting those signals so we can label videos created by tools from other companies.
This approach represents the cutting edge of what’s technically possible right now. But it is difficult to identify all AI-generated content at scale, and there are ways that people can strip out invisible markers. The rapidly evolving rate of deepfake technology makes it difficult for detection methods to keep up with the latest techniques. There are many different types of deepfakes, such as face swapping or voice manipulation. Each type may require a unique approach to detection. There is not a one-size-fits all solution. An important part of detecting AI-generated content relies on industry standard indicators that other companies include in content created using their tools, which help us assess whether something is created using AI. We are pursuing a range of options. We are working hard to develop classifiers that can help us to automatically detect AI-generated content, even if the content lacks invisible markers. At the same time, we are looking for ways to make it more difficult to remove or alter invisible watermarks. For example, Meta’s AI Research lab FAIR has shared research on an invisible watermarking technology we’re developing called Stable Signature. This integrates the watermarking mechanism directly into the image generation process for some types of image generators, which could be valuable for open source models so the watermarking can’t be disabled.
Given the work detailed above, we consider this recommendation implemented in part and complete.
To ensure more accurate labeling, Meta should attach provenance information and invisible watermarks to content created by Meta AI tools, so it can be consistently detected and labeled across platforms. This should include applying Content Credentials at the point of creation alongside using industry standard indicators for attribution to all content generated by Meta AI.
The Board will consider this recommendation implemented when the company shares with the Board a report on how consistently Meta AI attaches and preserves provenance data and invisible watermarks to content shared on the platform.
Our commitment: Our AI-generation tools already embed invisible watermarks on Meta-generated image content and we are actively evaluating how to improve provenance signals across our platform.
Considerations: When photorealistic images are created using our Meta AI feature, we may do several things to make sure people know AI is involved, including putting visible indicators such as labels as well as invisible watermarks and metadata embedded within image files. Using both invisible watermarking and metadata in this way improves both the robustness of these invisible markers and helps other platforms identify them. This is an important part of the responsible approach we’re taking to building generative AI features. For AI-generated audio, we have developed technology through our AudioSeal research, which embeds inaudible markers into audio that can later be detected to confirm it was created by AI.
We are also actively investing in making these watermarks harder to remove or tamper with. As mentioned in recommendation #5, our Fundamental AI Research (FAIR) team has published research on Stable Signature, a technique that builds the watermark directly into the image generation process itself, rather than adding it after the fact. This means the watermark is woven into how the image is created, making it significantly more difficult to strip out or disable. This approach is particularly valuable for open source AI models, where it helps ensure that watermarking remains intact regardless of how the model is deployed or by whom.
These invisible watermarks serve as a foundational layer of provenance. They travel with the content itself and do not depend on separately attached information that can be removed. This allows us, and potentially other platforms, to detect AI-generated content even when it has been modified or re-uploaded without its original context.
Going forward, we will continue to maintain our invisible watermarking technology to ensure it remains effective as AI tools evolve. In parallel, we are evaluating how to improve provenance signals across our products. We will share updates with the Board on our progress as this work advances.
To make the use of High Risk and High Risk AI labels on deceptive content more consistent, Meta should develop pathways for affixing those labels to content much more frequently, assisted by clearer escalation channels from automated systems and at-scale review, so that such labeling can occur at a significantly higher volume.
The Board will consider this recommendation implemented when there are new pathways for escalations to affix High Risk and High Risk AI labels to content, and Meta reports to the Board on the volume of these labels attached in 2026, by quarter. The absence of a denominator (i.e., the total volume of unlabeled content that does not meet this threshold) should not be a barrier to providing this information to the Board.
Our commitment: As announced in March, we are in the process of updating our automated systems, which may result in improvements to scaled AI content labeling. We are still assessing how this may impact escalation pathways, including for content under our Misinformation Community Standard. Moreover, we will assess how these updates and others can inform escalation pathways for applying the more prominent AI label to content.
Considerations: In our initial responses and related announcement to Oversight Board recommendations about applying labels to AI-generated content, we emphasized that we will continue to review our approach to this content as the technology and its uses progress. As indicated in more recent responses to Board cases such as Alleged Audio in Iraqi Kurdistan, we are continuing this commitment to review and iterate on this, be it by expanded access to languages for labels or updating internal guidance and training related to the application of labels.
As we have noted in prior responses and in our Community Standards, we apply a more prominent label to AI-generated content that may materially deceive the public on a matter of importance. This more prominent label may be applied in addition to other AI info labels that we apply to the content to provide context as noted in the recommendations above. In many cases, media may be edited in benign ways such as for artistic reasons, but other times we may consider restricting AI-generated content when it otherwise violates our policies. In rare instances, content may be AI-generated in a way that creates a particularly high risk of materially deceiving the public on a matter of public importance, and we rely on a clear set of internal criteria to consider applying the more prominent label in these cases. This criteria includes requiring reliable indicators of being digitally created or altered, which are assessed by specialized teams.
As the Board notes, applying this label to content requires clear escalation pathways, but it also requires careful consideration of the content by subject matter experts. For example, in our response to recommendation 1 in the Iraqi Kurdistan case, we noted that we had recently updated our guidance to also apply the more prominent label when we identify duplicate, identical content. However, we also highlighted that there are instances when identical content may be shared in a clearly condemning or debunking context and we would not apply the label in those cases.
Finally, we recently announced in March 2026 that we have begun introducing more advanced AI systems to help enforce our policies. These systems cover languages spoken by 98% of people online, helping us apply policies more accurately and consistently across billions of pieces of content and can also understand context and cultural nuance, including niche subcultures, rapidly changing and regionally specific code words, emoji meanings and slang. However, expert teams will continue to be crucial in enforcement including on potentially high-risk and high-impact decisions. As part of this ongoing work to introduce these systems and continue to iterate on our policies and their enforcement, we will assess how we can further improve escalation pathways for additional application of the more prominent AI label to content that meets our criteria. We will provide updates on the progress of this work in future reports to the Board.