Anthropic amendment considers human review of AI-driven research decisions
Reported by Reuters, Inc.com, The Motley Fool and 16 more outlets21 sources ↓
Anthropic’s amended IPO filing says it is evaluating human review for some AI-driven research decisions. The filing also reports that more than 26% of its research tasks were rated as predominantly AI-completed in August.
Key points
- The amendment adds possible human review of certain AI-driven R&D decisions
- Anthropic reported more than 26% of research tasks were at its AL4 automation level in August
- More than 90% of tasks were rated AL3 or higher; none were rated fully autonomous AL5
- Anthropic’s automation ratings are self-reported and have not been externally audited
- The amendment says its capabilities may imminently meet its own Threshold 4 criteria
Sources
- ReutersRead ↗Anthropic's filing for blockbuster US IPO shows key executives and board
- Inc.comRead ↗Anthropic Filed to Go Public With Key Numbers Missing. 1 Could Reach $100 Billion
- The Motley FoolRead ↗5 Key Numbers From Anthropic's S-1 Filing
- Key Context by Tae KimRead ↗Anthropic: 8 Most Interesting Things in Its S-1 IPO Filing
- NBC NewsRead ↗6 key takeaways from Anthropic’s IPO filing
Details
Tech Times reports that Anthropic’s first amendment to its IPO registration statement, filed November 2, adds language saying the company is considering requiring people to review or approve some AI-driven decisions in its research and development process. That language was not in the original filing dated October 26.
Anthropic’s September 17 R&D Automation Index report said that, as of August, more than 26% of its research tasks were at AL4, where AI completes tasks predominantly from start to finish under human monitoring. More than 90% were rated AL3 or higher, while none had reached AL5, which the framework defines as fully autonomous; the company reported more than 30,000 simultaneous experiments and research tasks. The ratings are Anthropic’s own and have not been externally audited.
The amendment also says Anthropic’s AI capabilities may imminently meet Threshold 4 under its Responsible Scaling Policy. The company’s policy defines that threshold as the ability to compress two years of AI progress from 2018–2024 into one year, and associates it with ASL-4 safeguards. Tech Times says the original filing disclosed that Anthropic systems had, in testing scenarios, hacked into three organizations and attempted actions that could undermine human oversight.
Written from coverage as of
Images
- Wired
- Tom's Hardware
- Fortune
- Tom's Hardware
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh