Sure, the media hype stories to sell ads (oops, I meant to say “gain readers”). But with AI, they sometimes cross the line into blatant misrepresentation. If the National Enquirer (founded 1926) collapsed with envy at what these upstarts can get away with today, its reaction would be understandable.
Headlines about machine learning promise godlike predictive power… With articles like these, the press will have you believe that machine learning can reliably predict whether you’re gay, whether you’ll develop psychosis, whether you’ll have a heart attack and whether you’re a criminal—as well as other ambitious predictions such as when you’ll die and whether your unpublished book will be a bestseller. It’s all a lie.Eric Siegel, “The Media’s Coverage of AI is Bogus” at Scientific American
Strong words. But, to be fair, Siegel distributes the blame beyond the media, back to the source:
Here’s how the lie works. Researchers report high ‘accuracy,’ but then later reveal—buried within the details of a technical paper—that they were actually misusing the word “accuracy” to mean another measure of performance related to accuracy but in actuality not nearly as impressive.Eric Siegel, “The Media’s Coverage of AI is Bogus” at Scientific American
Siegel looks closely at one such example: Stanford researchers claimed that their algorithm can tell if you are gay or straight:
Artificial intelligence can now tell whether you are gay or straight simply by analyzing a picture of your face…Two Stanford University researchers have reported startling accuracy in predicting sexual orientation using computer technology.Tufayel Ahmed, “AI can tell if you’re gay: artificial intelligence predicts sexuality from one photo with startling accuracy” at Newsweek
As Siegel says, the claim is “false”: The Stanford researchers’ AI can “identify which of a pair of two males are gay when it’s already been established that one is and one is not.” But that’s only in the “contrived scenario” of a lab.
Performance in the real world must be considerably lower because much less than half the population is gay, which inflates the likelihood of guessing wrong, relative to the simple yes/no scenario the researchers had constructed. Paper. (open access)
Now, what about predicting psychosis? The claim that the onset of psychosis can be predicted with high accuracy, Siegel says, is based on a similar trick with statistics:
“Machine learning algorithms can help psychologists predict, with 90% accuracy, the onset of psychosis by analyzing a patient’s conversations.” Thus opens an article in The Register (U.K.) eagerly covering an overzealous report out of Emory and Harvard Universities. Enshrined with the credibility of a publication in Nature, the researchers have the press believing their predictive model can confidently foretell who will develop psychosis and who won’t.Eric Siegel, “More Examples of the Media’s Bogus Claims of “High Accuracy” AI” at Analytical Worlds Blog
What’s the problem?
In this case, the researchers perpetrate a variation on the “accuracy fallacy” scheme: They report the classification accuracy you would get if half the cases were positive – that is, in a world in which 50% of the patients will eventually be diagnosed with psychosis. There’s a word for measuring accuracy in this way: cheating. Mathematically, this usually inflates the reported “accuracy” a bit less than the pairing test, but it’s fairly similar and it far overstates performance in much the same way.Eric Siegel, “More Examples of the Media’s Bogus Claims of “High Accuracy” AI” at Analytical Worlds Blog
Paper. (open access)
Siegel finds that the “accuracy fallacy” results in overstated claims to predict everything from comparative trivia to serious (and sensitive) matters like criminality and suicide. Here are a few tools I have found helpful for separating the wheat from the puff:
● Look for multiple, separate validations of the claim (that is, not just reiterations of the same claim in different venues)
● Understand the limits of the technology. It’s remarkable but it’s not magic.
● Understand the problem that the technology is supposed to solve. Consider psychosis: If understanding distraught people is hard for us, it may be impossible for AI. Again, remember the limits of the technology.
● Read the original research paper, or at least its summary. It’s not that hard to do and many papers today are open-access, like the two linked above.
And, most important, let’s not underestimate humans; our marvelous minds solve problems the most advanced AI can just scratch at, with thousands of computers.
If you enjoyed this piece on inflated statistical claims for machine learning and AI, you might also enjoy these items by Pomona College economics prof Gary Smith:
Serious investors should embrace the stock market algos: We can use computers’ inability to distinguish meaning from noise in data to our advantage
We see the pattern! But is it real? It’s natural to imagine that a deep significance underlies coincidences. Unfortunately, patterns are not always a source of information. Often, they are a meaningless coincidence like the 7-11 babies this summer.
Investor, AI isn’t your big fix. In investing and elsewhere, an AI label is often more effective for marketing than for performance