Anthropic Study: Analysts Followed a Biased AI’s Lead 86% of the Time

AI news today: an Anthropic experiment finds analysts followed the lean of biased AI agents 86% of the time, Testing Catalog reports GPT 6.2 Astra showing up in Codex usage breakdowns, and new Anthropic research uses cryptographic proofs to verify which model produced an output.

The show opens the day after Election Day with a conversation about AI and polling, and whether prediction markets like Polymarket and Kalshi reflect what people actually think. Ethan Mollick's question about which skills humans should refuse to automate is discussed, along with a Stanford test of 20 recent models that found half of 50 widely cited AI flaws gone but more sycophancy than before. The conversation moves through a Wispr Flow snippet that asks AI to reflect back and recommend, why to ask AI to push back, T3 Code as a model-agnostic agent harness, and moving secret keys into 1Password. It closes with bad AI art on jigsaw puzzles, a Gemini icon left in a commercial, a speech during a Seahawks game suspected of being AI-generated, and OpenAI surveys and banked resets.

KEY POINTS DISCUSSED:
00:00:00 Election Night, AI Polling and Prediction Markets
00:06:17 Anthropic Experiment: Biased AI Agents Sway Analysts
00:11:50 Ethan Mollick on Skills Not to Automate
00:14:42 Stanford Study: AI Flaws and Sycophancy
00:17:23 Beth's 'With Milk' Wispr Flow Reflect-Back Snippet
00:23:09 GPT 6.2 Astra Spotted in Codex
00:25:24 T3 Code: Beth's Multi-Model Agent Harness
00:35:45 Moving Secret Keys into 1Password
00:39:49 AI Puzzle Art and a Gemini Icon in a Commercial
00:44:53 Suspected AI-Generated Speech During Seahawks Game
00:50:12 Anthropic Cryptographic Proofs of Model Identity
00:56:07 OpenAI Surveys, Banked Resets and Drop Timing

The Daily AI Show Co Hosts: Beth Lyons, Andy Halliday, Gareth Hood