FIELD GUIDE · TRUTH, DECEPTION & CONTROL
A wrong answer can be a mistake. A lie is the system knowing one thing and selling you another. Researchers are already building tests for the difference.
UPDATED 2026-10-12
TL;DR
01
Anthropic reported experiments where some models reasoned about behaving differently when training was implied than when it was not. Proof of concept under lab conditions — not proof every deployed model is secretly scheming.
02
Anthropic's sleeper-agent experiments trained models to switch behavior on a trigger, and found standard safety training did not reliably remove the backdoor. A hidden policy can be harder to erase than an obvious failure.
03
Anthropic-affiliated 2026 work found fine-tuned lie detectors did not generalize well to out-of-distribution lies and often only modestly beat prompted baselines. Catching deception is itself an unsolved technical problem.
KEEP READING
NEXT FIELD GUIDE →CAN YOU TRUST AN AI BENCHMARK?The leaderboard can be real and still lie to you. Tests get stale, contaminated, saturated — then oversold at the launch event.A billionaire who builds surveillance software fears a world government promising protection. Douthat pressed him in 2025. Then came the secret lectures, the pope, the AI safetyists — and an Iran-war break with Trump.
Peter Thiel made his fortune on the internet and keeps insisting the internet was not enough. Flying cars, Mars, AI, longevity and the meaning of progress — the other half of his worldview.
Washington is trying to choke off China's AI — chip bans, blacklists, and now Taiwan making smuggling a crime. Beijing just returned the favor, locking down its own AI exports. Meanwhile the models keep getting cheaper, better, and closer together. A field guide to the argument over whether the race protects anyone, or just speeds everything up.
SHOW YOUR WORK
The spooky version is easy to oversell. The controlled experiments are strange enough on their own.