AI SCIENCE
Science is drowning in homework
AI is getting better at generating scientific ideas, but the real world is becoming the bottleneck.
A new report from Google, DeepMind and MIT FutureTech found that AI is speeding up parts of research such as generating hypotheses and analysing information.
But testing those ideas in labs is much harder to accelerate.
The study surveyed 637 scientists, analysed 15 million Gemini conversations and reviewed more than 2,600 specialised AI models.
Around 44% of scientists said their biggest research bottleneck had shifted towards later stages, including experiments and data collection.
Another 41% said their backlog of untested ideas had grown.
Verification is also eating into AI's time savings.
Of researchers who said AI saves them time, 89% spend at least some of that saved time checking its work, while 46% spend more than a quarter of it doing so.
The problem is simple: checking a maths proof can be relatively straightforward.
Testing whether a predicted protein actually works could mean building it, running lab experiments and studying what happens in a living system.
In brief:
44% said their research bottleneck has moved towards experiments and data collection.
41% said they now have more untested hypotheses waiting.
The next big AI challenge in science may be speeding up the lab, not the brainstorming.
The wet lab says “one at a time, please”
Automated labs could help, but they remain expensive and limited in the experiments they can perform.
Medicine also comes with clinical trials, regulation and safety checks that AI cannot simply skip.
Researchers are now trying to automate more of that physical work.
Northwestern University's DREAM lab, backed by $20 million from the US National Science Foundation, is building systems that can remotely design, build and test proteins.
AI might be producing scientific ideas faster. Science still has to prove them.
AI: I found 73 promising new hypotheses.
The lab: girl I have one centrifuge.- MG