Op-Ed: Who cleans up after AI? AI discoveries and the weird new world of ‘whatever’ as science

You may or may not have noticed that the hype about AI comes from everyone but the people working with it.

Rather than just another useless AI hate session, let’s have a brief Socratic discussion about the virtues of AI:

“What are the virtues of AI?”

“It does exactly what it’s supposed to do. To a point.”

“What is it supposed to do?”

“AI things.”

“What can it do?”

“We’re not sure yet.”

“What can’t it do?”

“We don’t know.”

“Let’s be clear. What won’t it do?”

“We have no idea.”

“What doesn’t it do?”

“We’re trying to find out.”

There, in 60 words of very mildly extrapolated 2400-year-old human logic, is the problem. Plato and Xenophon, please excuse the extreme lack of sophistication. This just isn’t a very sophisticated time in human thought.

AI can now be described as a third party to human logic, science, and basic comprehension on any subject, the ultimate “whatever”.

In science, a third wheel might be a more apt description. Part of the problem is defining who manages the logic of AI behaviour in science.

AI in real science is a very different thing

AI in science, particularly in the field of discoveries, is subject to the same peer review processes, critiques, and afterthoughts as human science. This form of AI has nothing in common with gratuitous chatbots and folksy tech sales spiels.

AI discoveries are now continuous and in large quantities. these vast volumes of new data are a real burden. That’s putting a lot of pressure on the scientific review process.

As if fraudulent human scientific papers weren’t enough of a millstone, AI work has to be reviewed meticulously. That’s expensive and time-consuming for researchers and the mainstream science business as a whole.

This is where the question of “who cleans up after AI” comes into play. There are some interesting developments in this area, intentionally and/or otherwise.

According to Stanford Medicine, whole texts can be turned into AI agents which can discuss and collaborate. This very good idea could be easily turned into a self-audit process, identifying areas of agreement and disagreement.

This approach is getting desperately needed extra teeth and capacity with Google Deep Review, a dedicated verification platform. Deep Review has credentials across 16,000 actual deployments and is now launching publicly. Deep Review is specifically tasked with managing the gigantic volumes of materials for review, and this really is a “just-in-time” moment.

“But…?” Of course, there’s a “but”.

Healthy and necessary as the cross-referencing agents and Deep Review are, there’s a built-in caveat.

AI can, by definition, be scripted. Any “scientific finding” created purely to prove a point can be ready-made by a few prompts. In theory, you can “prove” that the Moon is made of Gruyère with a script. Scripts are relatively easy to spot, but they’re still wasteful when used for such motives.

Devoting resources to a scientific game of automated Whack-a-Mole leaves a lot to be desired. It’s hard to believe that the fraud factories will just shut up shop and go back to health insurance.

At what point does the review process tell “spurious science” where to go? There will need to be clear cutoff points for scientific drivel. It would be nice if the cutoff process were sufficiently ferocious to deter the time-wasters and fakes.

Pure research is a special case. Pure research is quite capable of turning up new data and new information that is hard to cross-check simply because it’s so new. Automated reviews would need to refer upwards to human scrutiny. It’s still a filter, but a careful filter.

Commercial research is vulnerable to fraud and self-seeking findings. This issue could be “painful but valuable”. The quality of research is critical, and the “paper mills” need to be exterminated. It’s reasonable to assume that commercial researchers could put strict guidelines on inputs, particularly AI inputs, to ensure proper referencing for citations, supporting evidence, etc. “Zero trust” has to be the unavoidable bottom line for researchers.

There’s also the issue of “just plain wrong” to consider. One of the problems with AI in general and AI agents in particular is their sourcing. Let’s leave out the issues of lousy AI management and vague sloppy guidelines and focus on whether information is correct. AI is perfectly capable of getting things completely wrong both in process and in output. The cleanup process needs to start there.