Meta’s Shifting Response to AI‑Generated Misinformation By Lucie Čejková

This guest blog was written by Lucie Čejková, a DISC Virtual Visiting Fellow and a junior researcher at the Faculty of Social Studies at Masaryk University. Her post examines how Meta approaches the governance of generative AI–driven misinformation, drawing on a recent case of AI-fabricated health disinformation circulating on Facebook in Czechia. By close reading of the platform’s Community Standards and their evolution over time, she highlights a broader shift from content removal to labeling and user-side responsibility. The blog concludes with reflections on the implications of this approach for platform accountability and outlines key considerations for policymakers.

Earlier this year, an AI-generated news report featuring a prominent Czech cardiologist began spreading on Facebook, the second-most popular social media platform in Czechia after YouTube. The video contained an AI-fabricated news report by an actual news anchor of a major mainstream news television, announcing the death of the cardiologist by being pushed under a metro car, supposedly evidenced by a video (that was, however, from an unrelated incident in the New York metro). Ultimately, it ended with a seemingly exclusive, right-before-the-dead account by the cardiologist, AI-generated as well, accusing the medical profession and pharmacists of keeping their patients suffering while promoting an allegedly miraculous cure that somehow cleans your blood vessels.

The video and its content were false and fabricated. The Demagog fact-checking project, which cooperates with Meta, verified the video and flagged it for removal.

With half of the texts and up to a third of the videos on the Internet possibly AI-generated, the case described is just a drop in the ocean. And much like the amount of ocean water growing due to the climate crisis, the amount of AI-generated content will likely keep rising in volume.

AI-Generated Misinformation in the EU

While some of this AI-generated content might be benign by nature (although the environmental damage of generating even such content should be kept in mind), it is the false part of it that induces skepticism among audiences and is of particular concern. The EU has been addressing these issues with measures such as the Digital Services Act (DSA), the Code of Conduct on Disinformation, and the AI Act. It imposes obligations on social media platform providers to mitigate risks posed by the dissemination of AI-generated or manipulated content.

Coming back to the opening case, Meta, the parent company of Facebook, is one of the signatories to the 2022 Strengthened Code of Practice on Disinformation. It is a voluntary step that signals the acceptance of the measures and commitments outlined by the EU.

Translated from EU documents through platform practices, platform community guidelines constitute a readily available, front-facing summary of how policies on disinformation, including that employing generative AI, are adopted and communicated to platform users. In this regard, Meta is no different, but, unlike other social media platforms designated as Very Large Online Platforms (VLOPs) under the DSA, its Community Standards include both the current and previous versions of the text.

Ambiguity in Meta’s Misinformation Policies

Looking at the Meta’s Community Standards page dedicated to misinformation, three key issues stand out. First, the policy does not define mis- or disinformation, and it never did in its previous versions dating back to 2023. This omission is, unfortunately, a common issue when such terms are mentioned. Nonetheless, who else than a voluntary signatory of the 2022 Strengthened Code of Practice on Disinformation, which, among other goals, aims to reduce manipulative behavior and empower users, should set the standard?

Second, and relatedly, Meta’s approach to misinformation displays a surprising evasion of commitment. To quote the policy rationale as of June 2026: “The world is changing constantly, and what is true one minute may not be true the next minute. People also have different levels of information about the world around them, and may believe something is true when it is not. A policy that simply prohibits ‘misinformation’ would not provide useful notice to the people who use our services and would be unenforceable, as we don’t have perfect access to information.”

However, believing something is true when it is not is at the core of the definition of beliefs in misinformation – if only it was included. Revisiting the example from the beginning of this text, believing that there is a miraculous cure to clean one’s blood vessels that doctors and pharmacists are hiding from us does not make it any less false or any less of a concern when such potentially health-hazardous information is circulated on the platform.

Instead, Meta only clarifies specific areas of misinformation and its approach to them. Misinformation that poses a risk of physical harm and the promotion of miracle cures are among the types of misinformation removed from the platform, which is probably why the original post with the video described at the beginning of this blog post was ultimately removed.

Third, the opening case also paves the way for examining Meta’s approach to misinformation using generative AI. Interestingly, the Community Standards page of misinformation mentions AI only when explaining that the platforms require users to disclose their use of AI when posting photorealistic images, videos, or realistic audio, as this could mislead other users.

The Turn from Removal to Labelling

A look at earlier versions of the page reveals that, in the case of AI-generated misinformation, Meta gradually refrained from explicitly mentioning the removal of AI-generated misinformation to promote the labeling approach. Specifically, the Manipulated Media section addressing AI-generated content (although without explicitly calling it so) was notably shortened and moved from under “misinformation we remove” to “For the following content, we include an informative label” in July 2024.

These examples are further evidence of the need for increased attention from policymakers. Measures should move beyond simply addressing the presence of generative AI misinformation to also address how platforms individualize their management by shifting accountability onto users. Labeling AI-generated or potentially misleading content, while useful as a transparency tool, effectively delegates the burden of verification to individual users whose attention, media literacy, and may vary.

At the same time, ongoing cooperation outside of the United States with external fact-checking organizations and NGOs, although essential, often results in a cyclical pattern in which flagged content is removed or downranked only to reappear in slightly altered forms. This dynamic is exactly the case of the profile that posted the false AI-generated news report with the Czech cardiologist. The profile of the user who posted the content is back, up and running, spreading conspiratorial and manipulated content.

Policymakers should therefore move beyond endorsing labeling and third-party flagging as sufficient safeguards and instead require more systemic, enforceable interventions that address the recurrence and amplification of such content. This could include stronger accountability for repeat dissemination and obligations for platforms to demonstrate the effectiveness of their mitigation measures, rather than merely accepting their existence and outsourcing the labor of verification on their users.

Ambiguity in Meta’s Misinformation Policies

This blog was written as a part of the CDEP and DISC Critical AI Policy Virtual Fellowship.

The work is part of the project “On our own: Opportunities and Risks in the Individualization of Society (PRINS) CZ.02.01.01/00/23_025/0008710”, which is co-financed by the European Union.