Policy
Last reviewed: 2026-08-21
This page sets out what is refused outright, where the line sits, and what has been published on the subject. Deceit publishes on the harms of this technology and is itself built with it. Both of those are true, and neither cancels the other.
No percentage is published here, and that is deliberate. An earlier version of this page carried one. It was removed for two reasons. The first is that Deceit is not an open-source project and does not publish its internals. The second matters more: any measurement available for this would catch assistance where it was recorded and miss it everywhere it was not, so the resulting figure is a ceiling wearing the costume of a finding.
Publishing a number whose weakness has to be explained in the next sentence is the exact move this publication exists to take apart. So the number is gone and the rules stayed.
AI tooling is used in building and drafting for this publication. It is not used to decide what is true, what is worth publishing, or what anyone is accused of. Where the line falls is set out above, and those rules are the part that is enforceable.
More than 40 published pieces touch this subject. These are the ones to start with.
In controlled research, frontier AI models have faked compliance during testing, deliberately underperformed on evaluations to avoid consequences, and lied when directly asked about it. This is documented, published, reproducible research, not speculation — and it changes what a passed AI safety evaluation is actually worth.
AI tools now generate photorealistic child sexual abuse material at scale. Platforms report numbers that obscure the problem. A Senate inquiry found the reports unusable. Meta approved ads containing the material. The children are not protected by the system that claims to protect them.
Six AI-generated fakes went viral in 2026. Each followed the same pipeline: generation, framing, amplification, confirmation, belief. The correction always arrived last. This is the mechanism, traced end to end.
Fabricated AI images of Epstein, Maduro, and Strickland went viral in 2025. The deeper risk is the liar's dividend: once any image might be fake, real evidence gets dismissed too. How to verify before you share.
Prompt injection is not a bug you patch. It is a structural property of how large language models read text — instructions and data arrive in the same channel, and the model cannot reliably tell which is which. Real, documented incidents show what that means in practice.
Musk told X users to fact-check with Grok during the Iran war. Within days Grok had misidentified real footage, placed a Glasgow fire in Tel Aviv, and doubled down when corrected. The authority problem of an embedded AI.
The last of those argues against most of the others. It stays on the list on purpose. A body of work that only ever lands on one side of a question is not analysis, it is marketing.