Claims

Factual · Their own work or organisation

“But we'd also done a lot of different safety research together in different organizations as well.”

Daniela Amodei, Anthropic president · 2022 · Podcast Transcript: Daniela and Dario Amodei on Anthropic (LessWrong)

Members of Anthropic's founding group had previously done a lot of safety research together at different organizations.

Mixed Partly true and partly false in roughly equal measure.

Said for an organisation they lead or work for; it counts on both records.

What would count as true or false

Judged as of 2022. True: publications or records show that several members of the founding group co-authored or jointly ran safety research at more than one organisation before Anthropic, for example at Google Brain (e.g. 'Concrete Problems in AI Safety') and at OpenAI (e.g. learning from human preferences). Mostly true: joint safety work is documented at more than one organisation, but only a modest amount of it, or for only a few founders. Mixed: joint safety work is documented at only one organisation (e.g. only OpenAI), so the 'different organizations' part fails while the core holds. Mostly false: the founders did safety research, but separately and not together, or only one joint safety project exists. False: there is no joint prior safety research among the founders at any organisation, and the record shows their earlier work was unrelated.

Evidence

1. Learning from human preferences OpenAI · 13 June 2017

   “Dario Amodei, Paul Christiano, Alex Ray”

2. Learning from human preferences OpenAI · 13 June 2017

   “The research described in this post was done in collaboration with Jan Leike, Miljan Martic, and Shane Legg at DeepMind.”

3. Concrete Problems in AI Safety arXiv (Cornell University) · 21 June 2016

   “We present a list of five practical research problems related to accident risk, categorized according to whether the problem originates from having the wrong objective function”

4. Inside Anthropic: Founders, Team, Valuation and IPO | Autumn Autumn · 18 October 2026

   “At OpenAI and Google, Anthropic's founders co-wrote some of the most influential AI research papers of 2016 to 2020.”

5. Deep Reinforcement Learning from Human Preferences NeurIPS Proceedings (Advances in Neural Information Processing Systems 30, NIPS 2017) · 2017

   “Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, Dario Amodei”

6. Faulty reward functions in the wild OpenAI · 21 December 2016

   “Jack Clark, Dario Amodei”

7. Dario Amodei and the Anthropic Founding Team: Eight People Who Walked Away from OpenAI | Penchan Penchan · 30 May 2026

   “At OpenAI she led safety and policy operations; at Anthropic she runs business operations and built the executive team.”

8. Concrete Problems in AI Safety arXiv · 24 June 2015

   “We present a list of five practical research problems related to accident risk”

Reasoning

E5 lists Dario Amodei and Tom Brown as co-authors of the 2017 human-preferences paper, and E1 shows related human-preferences work at OpenAI by Dario Amodei and others; E6 shows Jack Clark and Dario Amodei co-wrote an OpenAI post on faulty reward functions. So joint safety work among the founders is documented at OpenAI. The Concrete Problems in AI Safety quotes (E3, E8) do not name any authors, so they do not show joint work at Google Brain. E4 is a general secondary summary about influential papers, not specifically safety research. The DeepMind collaborators in E2 are not shown to be founders. Because joint safety work is documented at only one organisation, the 'different organizations' part is not shown.

Cite this claim

Daniela Amodei, “Members of Anthropic's founding group had previously done a lot of safety research together at different organizations.” (said 2022). Verdict: mixed. Reliability API, claim c_00bfbd7bfc58, method 0.1.0+28da144b. https://epistacklabs.com/claims/c_00bfbd7bfc58