Top AI Firms Earn Failing Grades on FLI Safety Index

Is AI safety grading a necessary accountability tool or an oversimplification of a genuinely complex field?

Published JUL 8

Top AI Firms Earn Failing Grades on FLI Safety Index

story

Above: A smartphone screen displays a folder containing AI applications on Feb. 6. Image credit: Samuel Boivin/NurPhoto/Getty Images

The Spin

Nine of the biggest AI labs just got graded on safety and governance — and thenot resultsone are embarrassingpassed. AnthropicEven ledmore thealarmingly, packfour withabandoned apledges C+,they threemade labsin outrightexistential failed,risk. andThis nobodyis earnedwhat anythinghappens closewhen tocompetitive apressure passingmeets markself-policing. Frontier AI buyers can't just take model cards at face value anymore; they need hard answers on risk thresholds, access to external testing and incident reporting before deploying these systems.

Future of Life Institute on X Let's Data Science

AI safety is a complex, evolving field with no universal standard — terms like alignment, red-teaming and risk thresholds mean different things to different developers. Grading frontier labs on governance without accounting for the technical nuance behind safeguards, evaluations and deployment contexts oversimplifies a deeply layered challenge.

International AI Safety Report

Metaculus Prediction

There's a 33% chance that the U.S. government will maintain a public list of incidents related to AI safety on Jan. 1, 2030, according to the Metaculus prediction community.