Independent audits of major AI platforms — reproducible captures, labeled evidence states, and corrections published in the open. We audit the platforms we buy from. Especially those.
Every claim on this site carries one of these labels. Hard criticism gets more credible when it admits exactly how much it knows.
FACT — something directly captured, reproduced, measured, or decompiled.
"Resource X loaded while consent values were analytics:false and marketing:false."
INFERENCE — the most likely interpretation of the observed mechanism.
"The consent preference did not gate that resource load."
OPINION — our judgment about what the behavior means.
"We think a privacy control that behaves this way is inadequate."
These three never get silently collapsed into each other. That's the whole trick — it's not magic, it's hygiene.
Published evidence reports cover five vendors across six product lines — OpenAI (ChatGPT, Codex), Anthropic / Claude, Google / Gemini, xAI / Grok, and Meta / Llama — with 253+ surveillance endpoints mapped and blocked by the free AI Privacy Shield.
Full forensic packages (complete captures, endpoint documentation, reproduction steps) are being finalized for open publication. Preliminary findings ship with their evidence states in the evidence section.
If a claim changes, the change stays visible: original claim, corrected finding, reason for correction, current status. We do not silently rewrite historical findings — being visibly willing to correct ourselves is a feature, not a concession. Hold us to it.
Reproductions, counter-evidence, coordinated disclosure, and collaboration: Kingsley.w.m.curtis@gmail.com. If we've misread a capture, showing us is the fastest way to get a public correction.