Red Herrings, where Red Flags should be!

AI-washing is serious, but AI Safety-washing is worth an extra furrow over the eyebrows. When hats are packaged as helmets, the furrow leads to a deeper and darker rabbit-burrow.

- Pratima H

How many times do we ask- is this AI safe? No, that’s not the problem. The problem is something else - How do we define ‘safety’ to begin with? Do we measure it with the right inch-tape at all? Are we putting on helmets inside cars, and parachutes on bikes?

Some time back, in a paper ‘Safetywashing: Do AI Safety Benchmarks Actually Measure Safety Progress?’ Richard Ren and other researchers like Steven Basart, Alice Gatti and others (from Center for AI Safety, University of Pennsylvania, UC Berkeley, Stanford University, Yale University and Keio University) did a comprehensive meta-analysis of AI safety benchmarks, empirically analysing their correlation with general capabilities across dozens of models - also providing a survey of existing directions in AI safety. Here, the findings revealed that many safety benchmarks highly correlate with both upstream model capabilities and training compute, potentially enabling ‘safetywashing’. Basically, something where capability improvements are misrepresented as safety advancements. It also discussed how corporate entities often engage in safetywashing for the sake of appearances. This paper also talked about ‘portraying capabilities advancements in terms of safety progress by reporting correlated safety metrics to project an image of responsible AI development.’ Interestingly, and understandably, this behaviour is particularly pronounced when there is significant public pressure or regulatory scrutiny.