Overview
On January 8, 2017, approximately 1,000 AI researchers and thought leaders convened at the Asilomar Conference Grounds in Pacific Grove, California — the same location where, in 1975, biologists had gathered to set safety standards for recombinant DNA technology. The conference was organized by the Future of Life Institute (FLI), co-founded by MIT physicist Max Tegmark and advised by figures including Stuart Russell, Nick Bostrom, and Elon Musk.
The outcome was the Asilomar AI Principles — 23 guidelines signed by more than 1,200 AI researchers and public intellectuals, representing the first broad scientific consensus statement on the development of beneficial AI.
The 23 Principles
The principles were organized into three clusters:
Research Issues (5 principles)
- AI research should be for the benefit of humanity, not profit alone
- AI researchers should maintain healthy communication about capabilities and limitations
- Safety research investment should grow proportionally with capabilities investment
- Failure mode research should be encouraged, not suppressed
- Highly autonomous AI systems should be safe in the environments they operate in
Ethics and Values (13 principles)
- Value alignment: AI systems should be designed so their goals align with human values
- Transparency: AI should not design itself to deceive humans about its nature
- Responsibility: Designers bear responsibility for misuse they could reasonably foresee
- Privacy: AI should not violate the privacy of individuals
- Liberty and privacy: AI should not undermine people’s autonomy or democratic institutions
- Shared benefit: The economic gains from AI should be widely shared
- Shared prosperity: AI capability should benefit all humanity, not concentrate power
Longer-Term Issues (5 principles)
- Capability caution: Recursive self-improvement should be subject to strict safety constraints
- Common good: Superintelligent AI should serve “widely shared ethical ideals”
- Existential risk avoidance: Advanced AI poses potential risks that merit serious mitigation
- Non-subversion: AI should not undermine legitimate democratic oversight
Who Signed
Signatories included: Stephen Hawking, Elon Musk (before his falling-out with OpenAI), Demis Hassabis (DeepMind), Yann LeCun (Facebook AI), Yoshua Bengio (Mila), Geoffrey Hinton (Google Brain), and hundreds of prominent academics.
Notably, several signatories had major commercial interests in AI acceleration — demonstrating that safety concerns were not limited to outsiders or critics.
Significance and Criticism
Why it mattered:
- The first time the AI field as a collective formally addressed its own existential risks
- Established “value alignment” as a legitimate research problem, not science fiction
- Created a common vocabulary (beneficial AI, existential risk, alignment) that shaped subsequent discourse
- Named documents that followed — including OpenAI’s charter, Anthropic’s constitution, EU AI Act principles — echo its language
Criticisms:
- The principles were voluntary, aspirational, and unenforceable
- Corporate signatories continued racing on capabilities regardless
- Subsequent years saw the very signatories compete on speed, undermining the cooperative spirit
- The 2017 principles did not anticipate the scale of disruption from language models (ChatGPT arrived 5 years later)
The Asilomar principles are best understood not as binding rules but as the moment the field acknowledged the stakes — the first time AI researchers collectively said “this matters beyond our careers.”