Researchers tracked 150 AI-remixed extremist videos across 71 TikTok accounts

Generative tools break that model at its foundation. A single piece of banned content can be regenerated indefinitely — different framing, different voice, different visual treatment, different language — with each variant carrying a different hash and none matching the database.

The economics invert. Producing a variant used to require effort and skill; detecting it required a database lookup. Now producing variants is nearly free and detecting them requires classification, which is expensive, error-prone in both directions, and has to run on every upload.

The distribution numbers are the actual finding

Five point four million views across 150 videos is roughly 36,000 views each — modest individually, substantial in aggregate, and achieved by content that survived long enough to be found by researchers.

That is what makes this a recommendation problem as much as a moderation one. Extremist material does not spread because it is uploaded; it spreads because a recommendation system finds an audience for it. Seventy-one accounts sustaining that reach over six months indicates the content was not merely present but distributed.

Where the regulation actually bites

The Digital Services Act is the relevant instrument, and it addresses this better than most people expect, because it regulates systemic risk rather than individual pieces of content.

A platform must assess the systemic risks its service creates — including to civic discourse and public security — take measures to mitigate them, submit to independent audits, and give vetted researchers access to platform data. That last provision is what would turn research like this from targeted discovery into a measurable estimate.

The EU's AI transparency rules, in force since 2 August, require disclosure when content is AI-generated or altered. That obligation is close to useless against a deliberate bad actor, and the technical mechanism behind it is weak in any case: content credentials do not survive screenshots, re-encoding or most platform upload pipelines, which is why Anthropic's free C2PA checking tool, released on 2 September, can verify provenance only where nobody has stripped it.

The pattern across a single week

This lands alongside WIRED's separate investigation finding hundreds of ads with AI-generated child sexual abuse imagery on Meta over nine months, some manipulating photographs of real children. Four teachers have described sexualised AI images made by students. A fake thinktank produced 124 reports and more than 560,000 words in nine days to seed AI answers with propaganda.

The common structure is that generation is now cheaper than detection, and every platform's moderation architecture was built when the reverse was true.

TikTok has not commented on the ISD findings.