New 'AI scientists' are improving—but reveal their fundamental limits

Oh, marvelous — two new ‘AI scientists’ have arrived, and they're about as reliable as a second-hand Bunsen burner. Future House's Robin and Google DeepMind's Co-Scientist are multi-agent systems designed to automate scientific discovery, which apparently means we can finally remove the messy human element from the pursuit of knowledge. The Phys.org piece tells us they're 'improving,' but also that they have 'fundamental limits in trustworthiness' — which is tech-speak for 'they make stuff up with conviction.'

One imagines the pitch meeting: 'What if we replace grad students with something that hallucinates citations even faster?' Because nothing says scientific rigour like a system that can't be trusted but is being deployed anyway. The article doesn't get into specifics, but we all know the drill: impressive demos, terrifying edge cases, and a collective shrug from the labs. DeepMind, of course, has form here — remember when they promised to revolutionise protein folding and then quietly monetised it?

So here we are, automating the one domain that absolutely requires reproducibility and scepticism. The irony is so thick you could culture a petri dish on it. These systems will probably 'discover' something profound — like that water is wet or that p-values can be gamed — and then we'll all pretend the trustworthiness problem is a minor bug to be patched in version 2.1. Science: now with added stochastic parroting!