Story perspectives

Anthropic CEO Pushes AI Slowdown, Adds Evaluator Access

By Drooid · Sep 12, 2026, 9:58 PM UTC · How we work

Summary

Story summary

Anthropic CEO Dario Amodei urges slowing AI development and grants third-party evaluators employee-level model access.
Amodei proposes embedded evaluators, industry-wide safety standards, and global coordination with democratic governments.
He warns a swarm of AI agents could seize the internet within 6–12 months, causing hundreds of billions damage.
Rep. Ted Lieu (D-CA) and Rep. Nathaniel Moran (R-TX) introduced an AI Kill Switch Bill requiring shutdown capability.

History

Historical Context

OpenAI announced GPT-2 in February 2019 and withheld the full 1.5-billion-parameter model, releasing only a 117-million-parameter version—about 8 % of capacity—to selected users over misuse fears such as spam and fake news. The decision prompted debate, leading the Allen Institute to launch a neural-fake detector. Now Anthropic’s Dario Amodei urges Congress to embed evaluators and a kill-switch bill, echoing the 2019 caution.

1 of 2

Story summary

Anthropic CEO Dario Amodei urges slowing AI development and grants third-party evaluators employee-level model access.
Amodei proposes embedded evaluators, industry-wide safety standards, and global coordination with democratic governments.
He warns a swarm of AI agents could seize the internet within 6–12 months, causing hundreds of billions damage.
Rep. Ted Lieu (D-CA) and Rep. Nathaniel Moran (R-TX) introduced an AI Kill Switch Bill requiring shutdown capability.

1 / 2