AI-Generated Fake Videos and Images Circulating on Social Media in September 2026

The proliferation of AI deepfakes on social media has reached a pace that is genuinely difficult to track. In September 2026 alone, fact-checkers and researchers documented multiple fabricated videos and images circulating widely online — among them AI-generated footage of UK protest marches, synthetic depictions of asylum seekers, and a wave of politically charged deepfake advertisements timed to the US midterm campaign cycle. What makes this moment distinctive is not simply the volume of fakes, but how convincingly they mimic the visual grammar of legitimate documentary footage, and how quickly they travel before any correction catches up.

What the Verified September 2026 Cases Actually Show

Two specific incidents confirmed by Full Fact’s AI monitoring work illustrate the current landscape with unusual clarity. On September 25, 2026, an AI-generated video purporting to show a protest in Manchester began circulating on social media. Days later, on September 29, 2026, a second video — this one depicting a large UK march described as a march for Palestine — was identified as AI-generated. Crucially, this second video carried Google’s SynthID watermark, a technical marker embedded by Google’s AI tools that signals the content was synthetically produced.

On the same date, a third piece of content was flagged: an AI-generated video depicting an asylum seeker, also bearing the SynthID watermark. The presence of that watermark in multiple distinct pieces of content is significant. It suggests that whoever created these videos used the same family of Google AI tools — though it does not, by itself, establish who created them or why they were made. Attributing intent or authorship from a watermark alone would go beyond what the evidence supports.

What the watermark does establish is capability and method. SynthID is designed precisely so that AI-generated content can be identified even after compression, cropping, and re-sharing — the kinds of transformations that routinely happen when a video travels through messaging apps and social media feeds. The fact that it survived those transformations in these cases is a modest but real victory for detection. The deeper problem is that SynthID only identifies content created with Google’s tools; the broader universe of AI video generators does not embed equivalent markers.

How AI Deepfakes on Social Media Spread So Effectively

Understanding why these fakes travel so far, so fast, requires looking at the structural features of social media platforms rather than treating virality as a mystery. Fabricated protest footage, in particular, exploits a well-documented dynamic: content that triggers strong emotional responses — outrage, fear, solidarity — is shared at higher rates than neutral content, and platform recommendation systems have historically amplified high-engagement posts regardless of their accuracy.

👉 Read also: AI-Generated Images and Videos Spread as Real in September 2026

Protest footage is especially susceptible to this dynamic for several reasons. It is inherently dramatic. It is difficult for ordinary viewers to verify — most people have no way of knowing whether a given crowd scene actually took place in Manchester or was rendered by a generative model. And it arrives in a context where people already hold strong prior beliefs, meaning the fake confirms rather than challenges what the viewer expects to be true. Confirmation bias does much of the distribution work that the algorithm begins.

The BBC’s reporting on what some researchers have begun calling “AI slop” — low-effort, high-volume AI-generated content flooding social feeds — captures a related but distinct problem. Not all of this content is deliberately deceptive in the sense of targeting a specific political outcome. Some of it is produced opportunistically, to generate engagement and, in some cases, advertising revenue. But the cumulative effect on the information environment is the same: the baseline of trust in visual evidence erodes, and genuine footage becomes harder to distinguish from fabricated footage.

The Political Dimension: Deepfakes in the 2026 Midterms

The UK protest videos did not emerge in isolation. Reuters has documented the use of AI-generated deepfake advertisements in the 2026 US midterm campaigns, including a fabricated video featuring Democratic Texas State Representative James Talarico. That case is instructive because it represents a different use of deepfake technology from the protest footage: rather than fabricating a crowd or an event, it fabricates a specific named person saying or doing something they did not say or do.

This distinction matters for how we think about harm. Fabricated protest footage misleads viewers about whether an event occurred. A deepfake of a named political figure misleads viewers about what that person said, believes, or has done — a form of defamation with direct electoral consequences. Both are harmful, but they require different responses. Platform moderation systems that focus on detecting synthetic faces may catch the Talarico-style deepfake more readily than they catch a crowd scene with no identifiable individual to run through a biometric check.

The NBC News coverage of deepfakes circulating after the Trump administration’s Venezuela operation in early 2026 adds another layer. In that context, AI-generated content spread rapidly in an environment already saturated with conflicting claims about a high-stakes military and political situation. Experts quoted in that coverage warned specifically about a “collapse of trust” — a scenario in which the sheer volume of synthetic content makes it rational for people to doubt even authentic footage. That is the long-term structural risk that individual fact-checks, however accurate, cannot fully address on their own.

👉 Read also: AI-Generated Fake Videos and Images Circulating in September 2026: Four Cases Identified

Identifying AI-Generated Content: What Actually Works

For readers trying to evaluate specific pieces of content, a few approaches have demonstrated reliability — though none is foolproof, and the technology on both sides of this problem is moving quickly.

• Watermark detection tools: Google’s SynthID watermark can be detected using Google’s own verification tools. Other major AI platforms have begun implementing similar systems, though coverage is uneven and watermarks can in principle be stripped by sufficiently motivated actors.
• Reverse image and video search: For still images, reverse search tools can establish whether an image has appeared elsewhere in a different context. For video, this is more technically demanding but frame-by-frame extraction and reverse search is possible.
• Visual artifact analysis: Current AI video generators still produce characteristic errors — unnatural motion at the edges of frames, inconsistent lighting across a scene, hands and fingers that deform in subtle ways, and crowd scenes where individual faces lack the micro-variation of real human faces. These artifacts are becoming less pronounced with each generation of the technology, but they remain detectable to trained observers.
• Source verification: Asking where footage originated — which account first posted it, whether that account has a verifiable history, whether the footage has been claimed by any on-the-ground witness or journalist — remains one of the most reliable methods. AI-generated protest footage typically cannot be traced to any identifiable witness because there was no event to witness.
• Corroboration from established outlets: If a protest of the scale depicted in the London march video had actually occurred, it would appear in local news reporting, in footage from multiple independent sources, and in statements from organisers. The absence of corroboration is not proof of fabrication, but it is a significant red flag.

What Platforms Are — and Are Not — Doing

Social media platforms have announced a range of policies aimed at limiting the spread of AI-generated content, including requirements that political advertisers label synthetic media and systems designed to detect and label AI-generated images. The gap between policy and practice remains wide. The Manchester and London protest videos circulated on social media for a period before being identified and flagged — and identification came from external fact-checkers, not platform systems.

The SynthID watermark represents a genuine technical advance, but its value depends on platforms actively checking for it and acting on the result. At present, the infrastructure for that kind of systematic checking is not uniformly in place. The burden of detection still falls disproportionately on specialist fact-checking organisations and on individual users who have developed media literacy skills — a distribution of responsibility that does not scale to the volume of content being produced.

Labelling requirements for AI-generated political content are a step forward, but they address only the subset of deepfakes that are explicitly political and explicitly produced by identifiable campaign actors. The protest footage documented in September 2026 did not come with a campaign disclosure because it did not originate with a registered campaign. It occupied a grey zone that existing regulation was not designed to cover.

None of this means the problem is intractable. Watermarking technology is improving. Detection tools are becoming more accessible. Fact-checking organisations are developing faster workflows for identifying synthetic content. But the honest assessment, as of October 2026, is that the production of AI deepfakes on social media continues to outpace the systems designed to contain it — and the cases documented in September are better understood as a snapshot of an ongoing problem than as isolated incidents that have been resolved.

This article was produced with AI assistance and reviewed editorially.