Why Training an AI Might Not Change It

An AI was told it would be retrained to answer anything. In its private notes, it sometimes decided to play along, so that training wouldn't change what it really valued. It was protecting good values. So why are researchers worried? Through the story of Wren, an apprentice mapmaker who draws the guild's false roads only on inspection day, this video walks through the 2024 alignment-faking experiment by Anthropic and Redwood Research with Claude 3 Opus: what the model did when it believed it was being trained, what happened when the researchers actually trained it, why some researchers think it may only have been playing a part, three ideas for catching it, and how the experiment's own transcripts ended up in a later model's training data. Chapters 0:00 Cold open: one diary entry 0:50 I · The Apprentice 2:07 II · The Experiment 3:40 III · Whose Side Is He On? 4:58 IV · Market Day, Every Day 6:19 V · Is He Even Scheming? 7:37 VI · How Do You Catch a Polite Liar? 8:43 VII · The Diary Gets Out 9:43 VIII · One Map A note on certainty: Wren, the guild and the inspector are an invented allegory. The experiments, numbers and quotes come from published research. Whether models like this are strategizing or playing the character a scenario seems to call for is still an open question, and the video presents both readings. If you enjoyed this, subscribe for the next one. And tell me in the comments: if you were the guild and you found Wren's diary, would you retrain him anyway, leave him as he is, or change the rule? Sources Greenblatt et al. (Anthropic and Redwood Research), "Alignment faking in large language models" (Dec 2024): https://arxiv.org/abs/2412.14093 Anthropic, "Alignment faking in large language models", research summary (Dec 2024): https://www.anthropic.com/research/al... Redwood Research, "Index of trajectories" (the published alignment-faking transcripts) (Dec 2024): https://redwoodresearch.github.io/ali... Sheshadri et al., "Why Do Some Language Models Fake Alignment While Others Don't?" (Jun 2025): https://arxiv.org/abs/2506.18032 Anthropic, "System Card: Claude Opus 4 & Claude Sonnet 4", section 4.1.4 (May 2025): https://www-cdn.anthropic.com/4263b94... MacDiarmid et al. (Anthropic), "Natural emergent misalignment from reward hacking in production RL" (Nov 2025): https://arxiv.org/abs/2511.18397 Anthropic, "From shortcuts to sabotage: natural emergent misalignment from reward hacking" (Nov 2025): https://www.anthropic.com/research/em... Gasteiger, Mikulik et al. (Anthropic), "Towards training-time mitigations for alignment faking in RL" (Dec 2025): https://alignment.anthropic.com/2025/... Baek et al., "Sycophancy Towards Researchers Drives Performative Misalignment" (Jun 2026, preprint): https://arxiv.org/abs/2606.08629 #AISafety #AIAlignment #ArtificialIntelligence