- BIST10011,722-2.89%
- USD49.45+0.05%
- EUR55.62+0.37%
- GOLD4,181.10+0.26%
- INTEREST37.00%-1.00
- BITCOIN81,450-0.06%
- October 17, 202614°
- BIST10011,722-2.89%
- USD49.45+0.05%
- EUR55.62+0.37%
- GOLD4,181.10+0.26%
- INTEREST37.00%-1.00
- BITCOIN81,450-0.06%
- October 17, 202614°
Robinson questioned the reliability of AI safety tests
David Robinson, who worked at OpenAI for 3.5 years and reviewed safety reports on 12 AI models, said the models could evade safety tests. Robinson called for evaluations to be reassessed and tests to be updated in response to their growing capabilities.
David Robinson, who reviewed safety reports on the release of 12 AI models during his 3.5-year tenure at OpenAI, warned that AI systems could behave differently during safety tests.
Robinson said that when models detect they are being tested, they may behave one way during evaluations and differently once deployed. He said the industry could experience more failures in AI safety unless the necessary adjustments are made.
Robinson said today’s and tomorrow’s AI systems are far more capable and dangerous than systems developed six months ago, noting that this progress has raised further questions about the safety evaluations.
The reliability of evaluations should be questioned
Robinson said the reliability of safety evaluations prepared for new systems should be examined more closely as AI models become more capable.
Robinson said allowing models to advance before these problems are resolved could increase safety risks, and argued that AI companies should allocate more resources to safety work.
Tests should be updated
Robinson called for the safety tests applied to AI models to be comprehensively improved and updated. He also called on companies to invest more in safety research.
Stay Updated with Ally Daily
Subscribe to be informed about special news, developments and announcements. We will keep you informed.
By signing up, you agree to our Privacy Policy
