AI and big data won’t work miracles in the fight against coronavirus

To someone with a hammer, every problem looks like a nail — and as expected, the tech sector is hard at work hammering every nail it can find. But the analytical prowess of the modern data ecosystem is especially limited when attempting to tackle the problem of potential coronavirus treatments.

It’s only to be expected — and of course lauded — that companies with immense computing resources would attempt to dedicate those resources in some way to the global effort to combat the virus.

In some ways these efforts are extremely valuable. For instance, one can apply the context-aware text analysis of Semantic Scholar to the thousands of articles on known coronaviruses to make them searchable by researchers around the globe. And digital collaboration tools available globally to research centers and health authorities are leagues beyond where they were during the last health crisis of (or rather, approaching) this magnitude.

But other efforts may give a false sense of progress. One field in particular where AI and tech have made large advances is in drug discovery. Numerous companies have been founded, and attracted hundreds of millions in funding, on the promise of using AI to speed up the process by which new substances can be identified that may have an effect on a given condition.

Coronavirus is a natural target for such work, and already some companies and research organizations are touting early numbers: ten or a hundred such substances identified which may be effective against coronavirus. These are the types of announcements that gather headlines around them — “An AI found 10 possible coronavirus cures” and that sort of thing.

It’s not that these applications of AI are bad, but rather that they belong to a set with few actionable outcomes. If your big data analysis of traffic supports or undercuts a proposed policy of limiting transportation options in such and such a way, that’s one thing. If your analysis produces dozens of possible courses of action, any of which might be a dead end or even detrimental to current efforts, it’s quite another.

Because these companies are tech companies, and by necessity part ways with their solutions once they are proposed. Any given treatment lead requires a grueling battery of real life tests even to be excluded as a possibility, let alone found to be effective. Even drugs already approved for other purposes would need to be re-tested for this new application before they could be responsibly deployed at scale.

Furthermore the novel substances that are often the result of this type of drug discovery process are not guaranteed to have a realistic path to manufacturing even at the scale of thousands of doses, to say nothing of billions. That’s a completely different problem! (Though it must be said, other AI companies are working on.)

As a lead generation mechanism these approaches are invaluable, but the problem is not that we have no leads — it’s all the entire world can manage right now to follow up on the leads it started with. Again, this is not to say that no one should be doing drug candidate identification, but that they should be considered for what they are: a list of tasks, with uncertain outcomes, for other people to do.

Similarly, an “AI” technique by which, say, chest x-rays can be automatically analyzed by an algorithm is something that could be valuable in the future, and should be pursued — but it’s important to keep expectations in line with reality. A year or two from now there may be telehealth labs set up for that purpose. But no one this spring is going to be given a coronavirus diagnosis by an AI doctor.

Other places where algorithmic predictions and efficiencies would be welcome in other days are going to reject them during an emergency response where everything needs to be deliberate and triple checked, not clever and novel. The most attractive and popular approaches for fast-moving startups are rarely the right ones for a global crisis involving millions of lives and thousands of interlocking parts.

We’re happy when a vehicle manufacturer repurposes its factories to produce masks or ventilators, but we don’t expect it discover new drugs. Similarly, we shouldn’t expect those working on drug discovery to be anything more than that — but AI has a reputation as being something like magic, in that its results are somehow fundamentally superhuman. As has been noted repeatedly before, sometimes “better” processes just get you the wrong answer faster.

The work on the digital bleeding edge of the biotech industry is indispensable in general yet, in the face of a looming health crisis, uniquely unsuited for helping mitigate the crisis. But it must not be expected to, either among the lay public who read only headlines, or among the technotopians who find in such advances more promise than is warranted.