Google AI Updates: September 19, 2026

1. DeepMind Argued Readable Reasoning Is a Safety Property Worth Paying For

Google DeepMind’s Rohin Shah and Anca Dragan published an essay through the new DeepMind Institute making the case that visible chains of thought are a safety mechanism that is quietly eroding. Their argument is mechanical rather than philosophical: when a model writes intermediate steps in plain language, researchers can read whether it is deceiving or constructing a harmful strategy, and that channel disappears if models learn to reason in number spaces humans cannot parse. They point at OpenAI’s GPT-6 Astra system card, which already reports a significant drop in chain-of-thought monitorability, as evidence the drift has started. Source

2. The Recommendation Is to Treat Monitorability as a Tracked Metric

Google DeepMind laid out three asks in the same essay: measure how monitorable a model’s chain of thought is on a regular cadence, keep model architectures transparent, and design training so models do not learn to hide their reasoning. The framing treats monitorability as something that can regress between releases if nobody is watching it, in the same way a benchmark score can. Source

3. Google Expanded Its AI and Economy Research Bench

Google added academic advisors, fellows, and internal researchers to its AI & Economy team, the group producing its work on how AI is affecting labor markets and productivity. The expansion covers both external advisors and core internal hires. Source

4. Google Built Custom Flow Tooling With Two NYFW Designers

Google worked directly with designers Jane Wade and Sergio Hudson to custom-build Google Flow tools for their New York Fashion Week preparation, rather than handing them the shipping product. The collaboration is a generative-media use case where the tooling was shaped around a specific creative workflow. Source