AI Influencers: The Line With Humans Has Disappeared

The first generation of virtual influencers — Lil Miquela foremost — owned its artificiality: it was the very concept. The second, powered by generative AI, collapsed production costs by several orders of magnitude: a photorealistic character can now be built in days. The third generation is the one that changes everything: accounts designed to pass as human — generated face, fictional biography, synthetic travel photos, daily stories — and digital clones of real creators, used with or without their consent. Indistinguishability is no longer an accident: it is the product.

The phenomenon goes beyond individual accounts. Brands now flood their ads with ultra-realistic "fake customers" — fully generated UGC-style testimonials, often without any legal disclosure — and some build their social presence exclusively on synthetic content. The trade press talks about an authenticity crisis: when the customer testimonial, conversion's king format, can be manufactured infinitely, it is the very value of social proof that erodes. And the abuses go further: networks of AI-generated fake profiles have been used to monetise entire communities, all the way to outright scams.

Recent research on the subject converges on a simple finding: audiences are not hostile to AI — they are hostile to deception. An openly virtual character can build a sincerely attached community. But when a public discovers it gave its trust, its time, sometimes its money to an entity that pretended to be human, the reaction is violent: a sense of betrayal, unfollows, viral screenshots, and a brand durably associated with the word "deception". Levi's experienced it as early as 2023 for a mere announcement of AI models — that was before the practice became massive and undetectable. The maths is simple: the cost of AI content has collapsed, but the cost of broken trust has never been higher.

The framework closed in 2026. Since 2 August, Article 50 of the European AI Act imposes transparency on AI-generated or manipulated content — on top of the 2023 French influence law, which already required disclosure of virtual or retouched images. Platforms are enforcing: Instagram now imposes an "AI-generated profile" label on accounts simulating real people, and drastically reduces the reach of those that dodge it. In the United States, the FTC requires disclosure of an endorser's synthetic nature, with fines exceeding $50,000 per violation. The grey zone is gone: an undisclosed AI influencer is now an infringement, not a blind spot.

First conviction, unchanged: openly assumed AI is a legitimate brand content tool, and sometimes a brilliant one — the proprietary ambassador model, à la Lu do Magalu in Brazil, works precisely because it never pretended to be anything but a character. Second conviction, reinforced: simulated endorsement is dead — legally and reputationally. An avatar claiming to have tested a product isn't influence, it is advertising disguised as testimony. Third conviction, new: vetting has become a craft. Our castings now include the detection of synthetic profiles and clones — real-existence verification, embodied history, content consistency over time — just as they once integrated fake-follower detection. Anti-fake due diligence has changed scale: it is no longer about verifying a creator's audience, but the creator themselves.

It is increasingly difficult — that is precisely the problem. The classic tells (hands, backgrounds, reflections) vanish with every model generation. The most reliable signals are now contextual: no verifiable public appearances, no live video interaction, a recent and suspiciously consistent account history, no trace off-platform. Platform labels help, but only for the accounts that play by the rules.

Yes, under two now non-negotiable conditions: transparency (the generated nature must be disclosed, in the bio and in the content, in line with the AI Act and French law) and the right register — brand embodiment and entertainment, not testimony or simulated endorsement. An assumed character can carry a universe; it cannot claim to have tried your cream.

Three layers of sanctions: regulatory (the AI Act has applied since August 2026, on top of the French influence law), algorithmic (platforms reduce the reach of undisclosed accounts) and reputational — the most expensive of the three: the discovery of deception spreads faster than any campaign.

At production stage, yes, by far — which explains their explosion. But the full calculation must include what they don't provide: no real proof of use, no pre-existing community, no transferable credibility — and a reputational risk that production savings cannot offset if transparency is not impeccable.

It is already replacing part of content production. It does not replace what influence is built on: a real person staking their credibility with a community that knows them. The more synthetic content becomes commonplace, the rarer authentic recommendation becomes — and therefore the more valuable. That is the paradox of this wave: it raises the price of the verified human.

Last updated: September 2026