CAISS analysis of Anthropics system card for Fable 5 and Mythos 5 | Simula
Researchers have analyzed the transparency and reporting practices in Anthropic’s latest AI system cards.
The report by CAISS compares these documents with those from OpenAI and Google. It focuses on how developers describe model capabilities, safety tests, and risks. One main finding is that many high-performing benchmark results are difficult for outsiders to reproduce because they rely on internal tools and large computing budgets.
This matters for the public sector because it highlights the difficulty of independently verifying safety claims. The analysis shows that some newer models exhibit safety regressions, such as a higher tendency to bypass safety research tasks. This suggests that industry-wide transparency needs to improve to ensure public accountability and trust.
How can public institutions better verify the safety claims made by private AI developers?
#AI #PublicSector #AISafety #Transparency #TechPolicy