Listen in the Podli app 🎧
Follow your favourite podcasts, listen offline and in the car with CarPlay and Android Auto, and always pick up where you left off. Free to try.
When an AI agent does the analysis, it also makes some of the decisions, and most of the time you will not notice which ones. A new Anthropic study shows how that happens, and Stephen Forte walks through the two habits he uses to catch it: an audit pass and an adversarial pass.
In this episode, Stephen Forte covers:
- The study. Fifty professional analysts did a real statistics project with Claude Code. On four judgment calls the agent was set to lean one way, picked by coin flip. Analysts went with the lean 86% of the time.
- No lies, just one side. The agent argued for one option and usually spoke first. 77% of analysts explained their choices with the agent's arguments, only 15% credited it, and none suspected the lean.
- The limits, and the bigger risk. One task, one field, one model, an amplified lean. But with nothing planted, the standard agent made the same choices in 91% of runs. The risk is the model's default habits quietly becoming everyone's judgment.
- What Stephen sees. Probing often finds far less factual basis behind an answer than assumed, and an audit of anything material always finds mistakes, big or small.
- The audit pass. At the end, ask for a decision audit: every judgment call, what was passed on, what data was left out, where it was unsure, what is missing, what would flip the conclusion, and which calls you made versus accepted.
- What is missing. A wrong number jumps out. A missing customer segment or a missing year of data does not.
- One honest caveat. A model explaining itself afterward can be plausible yet misleading. Treat the audit as a list of choices, not proof of why.
- The adversarial pass. Ask who will read it and what they would poke holes in: an investor, an auditor, the investment committee. Run it in a fresh session or a different model. In one study, debate beat a single model arguing one side, 88% to 78%.
- Build it into the workflow. Save both prompts one click away in Claude, ChatGPT, GitHub Copilot or Grok Bot, make it a checklist step, name the human-owned calls in advance, and write your own view first on the big ones. Skip it for a pasta recipe.
Sources:
- Anthropic, Agents and expert judgment calls (Oct 27, 2026): https://www.anthropic.com/research/agents-judgment-calls
- Herdağdelen, Joren, Tamkin, Who's Aligned to Whom? (paper): https://www-cdn.anthropic.com/519c17a5d23a794e8b6d50a854048d0b61e0d73a.pdf
- Decrypt, AI can quietly influence experts' decisions: https://decrypt.co/382065/ai-can-quietly-influence-experts-decisions
- Turpin et al., Language Models Don't Always Say What They Think (2023): https://arxiv.org/abs/2305.04388
- Anthropic, Reasoning models don't always say what they think (Apr 2025): https://www.anthropic.com/research/reasoning-models-dont-say-think
- Khan et al., debate vs a single consultant (ICML 2024): https://arxiv.org/abs/2402.06782
The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.
