AI Weekly: Watson Well being and semi-autonomous driving failures present the risks of overpromising AI

This week in AI, actuality got here knocking — each for AI well being tech and semi-autonomous driving methods. IBM agreed to promote the belongings from its Watson Well being enterprise to funding agency Francisco Companions, following the division’s steeply declined efficiency. In the meantime, the Insurance coverage Institute for Freeway Security (IIHS), a nonprofit financed by the insurance coverage trade, introduced a brand new rankings program designed to guage how properly “partial” automation methods like Tesla’s Autopilot present protections to stop misuse.

The dual developments are emblematic of the AI trade’s perennial downside: acknowledging AI’s limitations. Slate’s Jeffrey Funk and Gary Smith do a thorough job of recapping overly optimistic AI predictions lately, together with Ray Kurzweil’s pronouncement that computer systems could have “human-level” intelligence and “the power to inform a joke, to be humorous, to be romantic, to be loving, to be attractive” by 2029.

As any professional will verify, AI is nowhere near approaching human-level intelligence — emotional or in any other case. (Kurzweil’s new estimate is 2045.) Equally, autonomous vehicles and AI-driven well being care haven’t achieved the lofty heights futurists as soon as envisioned they’d attain. It’s an vital lesson in setting expectations — the longer term isn’t straightforward to foretell, in any case — but in addition an instance of how the pursuit of revenue supercharges the hype cycle. Below strain to point out ROI, some well being tech and autonomous automobile corporations have collapsed below the burden of their overpromises, as this week’s information exhibits.

Driving excessive on the Watson win in opposition to Jeopardy! champion Ken Jennings, IBM launched Watson Well being in 2015, positioning the suite of AI-powersed companies as the way forward for augmented care. The corporate’s gross sales pitch was that Watson Well being might analyze reams of medical knowledge — far quicker than any human physician might, ostensibly — to generate insights that enhance well being outcomes.

IBM reportedly spent $4 billion beefing up its Watson Well being division with acquisitions, however the expertise proved to be inefficient at finest — and dangerous at worst. A STAT report discovered that the platform typically gave poor and unsafe most cancers therapy recommendation as a result of Watson Well being’s fashions have been educated utilizing faulty, artificial medical data fairly than actual affected person knowledge.

Watson Well being’s demise could be partially attributed to IBM CEO Arvind Krishna’s shifting priorities, however rising cynicism about AI’s well being capabilities little doubt additionally performed a task. Research have proven that almost all eye illness datasets come from sufferers in North America, Europe, and China, that means eye disease-diagnosing algorithms are much less sure to work properly for racial teams from underrepresented international locations. An audit of a UnitedHealth Group algorithm decided that it might underestimate by half the variety of Black sufferers in want of higher care. And a rising physique of labor means that pores and skin cancer-detecting algorithms are usually much less exact when used on Black sufferers, partially as a result of AI fashions are educated totally on photos of light-skinned sufferers.

Semi-autonomous, AI-powered driving methods are coming below related scrutiny, significantly as automakers ramp up the rollout of merchandise they declare can almost drive a automobile themselves. In October 2021, Tesla was ordered to show over knowledge to the Nationwide Freeway Site visitors Security Administration as a part of an investigation into the corporate’s vehicles crashing into parked automobiles. The suspicion was that Tesla’s Autopilot was accountable — both partially or entire — for the damaging behaviors.

It’s not an unreasonable assumption. Late final yr, Tesla rolled out an replace to Autopilot with a bug that precipitated the automated braking system in Teslas to interact for no obvious cause. This precipitated vehicles to quickly decelerate as they traveled down the freeway, placing them susceptible to being rear-ended.

Tesla isn’t the one vendor that’s struggled to good semi-autonomous automobile expertise. A sobering 2020 examine from the American Vehicle Affiliation discovered that the majority semi-autonomous methods in the marketplace — together with these from Kia and BMW — bumped into issues a median of each eight miles. For instance, when encountering a disabled automobile, the methods precipitated a collision 66% of the time.

In 2016, GM was compelled to push again the rollout of its Tremendous Cruise characteristic as a consequence of unspecified points. Ford just lately delayed its BlueCruise system as a way to “simplify” the tech.

That brings us to this week’s information: the Insurance coverage Institute’s rankings program to guage semi-autonomous methods’ security protections. The group hopes it’ll encourage automakers to create higher designs as soon as the primary set of rankings, presently in growth, are issued this yr.

“The best way many of those methods function offers folks the impression that they’re able to doing greater than they are surely,” Insurance coverage Institute analysis scientist Alexandra Mueller mentioned in a press release. “However even when drivers perceive the restrictions of partial automation, their minds can nonetheless wander. As people, it’s more durable for us to stay vigilant once we’re watching and ready for an issue to happen than it’s once we’re doing all of the driving ourselves.”

That’s all to say that AI — whether or not self-driving or disease-diagnosing — is fallible, identical to the people who design it. Visions of a Jetsons future is likely to be tantalizing, however when lives are at stake, historical past has proven that it’s finest to be overly cautions.

For AI protection, ship information tricks to Kyle Wiggers — and you’ll want to subscribe to the AI Weekly publication and bookmark our AI channel, The Machine.

Thanks for studying,

Kyle Wiggers

Senior Employees Author