Study Shows We Don't Need Super AI to Make AI Safer

There's been a big argument inside artificial intelligence: do we need to build smarter AI before we can make AI safe? Some lab leaders say yes, because safety tools — like testing and oversight — need highly capable AI to work. Others say no, and that pouring money into bigger models just creates more risk. This report is the first to actually count, using data.

The researchers looked at 521 safety-focused papers and blog posts published from 2019 to 2025. They scored each one to see if its central finding depended on access to top-tier AI, like what OpenAI or Google build. The result? Only 2.7% of papers truly needed frontier models. The other 97% got their results using ordinary, non-frontier AI, or no models at all.

But there's an important twist. When the team looked at the most highly cited, influential papers, the share that depended on frontier AI more than doubled — to over 7%. So advanced AI isn't broadly necessary for safety progress, but in the trend-setting work that shapes the field, it clearly helps. It makes techniques stronger and shows dangerous behaviors more visibly.

What it doesn't do is create a bright line. Different safety specialties have very different needs. For example, papers about evaluating new AI systems lean on capable models pretty heavily. But papers in theory and interpretability — understanding how AI thinks — mostly don't. The takeaway: the "smarter AI or safer AI" debate misses the middle ground. What matters is the job you're trying to do.