In early 2011, Ken Jennings looked like humanity’s last hope. Watson, an artificial intelligence created by the tech giant IBM, had picked off lesser Jeopardy players before the show’s all-time champ entered a three-day exhibition match. At the end of the first game, Watson—a machine the size of 10 refrigerators—had Jennings on the ropes, leading $35,734 to $4,800. On day three, Watson finished the job. “I for one welcome our new computer overlords,” Jennings wrote on his video screen during Final Jeopardy.
Watson was better than any previous AI at addressing a problem that had long stumped researchers: How do you get a computer to precisely understand a clue posed in idiomatic English and then spit out the correct answer (or, as in Jeopardy, the right question)? “Not a hit list of documents where the answer may be,” which is what search engines returned, “but the very specific answer,” David Ferrucci, Watson’s lead developer, told me. His team fed Watson more than 200 million pages of documents—from dictionaries, encyclopedias, novels, plays, the Bible—creating something that sure seemed like a synthetic brain. And America lost its mind over it: “Could Watson be coming next for our jobs in radiology or the law?” NPR asked in a story called “The Dark Side of Watson.” Four months after its Jeopardy win, the computer was named Person of the Year at the Webby Awards. (Watson’s acceptance speech: “Person of the Year: ironic.”)
But now that people are once again facing questions about seemingly omnipotent AI, Watson is conspicuously absent. When I asked the longtime tech analyst Benedict Evans about Watson, he quoted Obi-Wan Kenobi: “That’s a name I’ve not heard in a long time.” ChatGPT and other new generative-AI tools can furnish pastiche poetry and popes wearing Balenciaga, capabilities that far exceed what Watson could do a decade ago, though ones still based in the ideas of natural-language processing that helped dethrone Jennings. Watson should be bragging in its stilted voice, not fading into irrelevance. But its trajectory is happening all over again; part of what doomed the technology is now poised to chip away at the potential of popular AI products today.
The first thing to know about Watson is that it isn’t dead. The machine’s models and algorithms have been nipped and tucked into a body of B2B software. Today IBM sells Watson by subscription, folding the code into applications like Watson Assistant, Watson Orchestrate, and Watson Discovery, which help automate back-end processes within customer service, human resources, and document entry and analysis. Companies like Honda, Siemens, and CVS Health hit up “Big Blue” for AI assistance on a number of automation projects, and an IBM spokesperson told me that the company’s Watson tools are used by more than 100 million people. If you ask IBM to build you an app that uses machine learning to optimize something in your business, “they’ll be very happy to build that, and it will probably be perfectly good,” Evans said.
From the very beginning, IBM wanted to turn Watson into a business tool. After all, this is IBM—the International Business Machines Corporation—a company that long ago carved out a niche catering to big firms that need IT help. But what Watson has become is much more modest than IBM’s initial sales pitch, which included unleashing the machine’s fact-finding prowess on topics as varied as stock tips and personalized cancer treatments. And to remind everyone just how revolutionary Watson was, IBM put out TV commercials in which Watson cheerfully bantered with celebrities like Ridley Scott and Serena Williams. The company soon struck AI-centric deals with hospitals such as Memorial Sloan Kettering and the MD Anderson Cancer Center; they slowly foundered. Watson the machine could play Jeopardy at a very high level; Watson the digital assistant, essentially a swole Clippy fed on enterprise data and techno-optimism, could barely read doctors’ handwriting, let alone disrupt oncology.
The tech just didn’t measure up. “There was no intelligence there,” Evans said. Watson’s machine-learning models were very advanced for 2011, but not compared with bots like ChatGPT, which have ingested much of what has been published online. Watson was trained on far less information and excelled only at answering fact-based questions like the kind you find on Jeopardy. That talent contained obvious commercial potential—at least in certain areas, like search. “I think that what Watson was good at at the time kind of morphed into what you see Google doing,” Ferrucci said: surfacing precise answers to colloquial questions.
But the suits in charge went after the bigger and more technically challenging game of feeding the machine entirely different types of material. They viewed Watson as a generational meal ticket. “There was a lot of hyperbole around it, and a lot of lack of appreciation for what it really can do and what it can’t do, and ultimately what is needed to effectively solve business problems,” Ferrucci said. He left IBM in 2012 and later founded an AI start-up called Elemental Cognition.
When asked about what went wrong, an IBM spokesperson pointed me to a recent statement from CEO Arving Krishna: “I think the mistake we made in 2011 is that we concluded something correctly, but drew the wrong conclusions from the conclusions.” Watson was “a concept car,” Kareem Yusuf, the head of product management for IBM’s software portfolio, told me—a proof of technology meant to prod further innovation.
And yet to others, IBM may have seemed more concerned with building a showroom for its flashy convertible than figuring out how to design next year’s model. Part of IBM’s problem was structural. Richer, nimbler companies like Google, Facebook, and even Uber were driving the most relevant AI research, developing their own algorithms and threading them through everyday software. “If you were a cutting-edge machine-learning academic,” Evans said, “and Google comes to you and Meta comes to you and IBM comes to you, why would you go to IBM? It’s a company from the ’70s.” By the mid-2010s, he told me, Google and Facebook were leading the pack on machine-learning research and development, making big bets on AI start-ups such as DeepMind. Meanwhile, IBM was producing a 90-second Academy Awards spot starring Watson, Carrie Fisher, and the voice of Steve Buscemi.
In a sense, IBM’s vision for a suite of business tools built around machine learning and natural-language processing has come true—just not thanks to IBM. Today, AI powers your search results, assembles your news feed, and alerts your bank to possible fraud activity. It hums in the background of “everything you deal with every day,” Rosanne Liu, a senior research scientist at Google and the co-founder of ML Collective, a research nonprofit, told me. This AI moment is creating even more of a corporate clamor for automation as every company wants a bot of its own.
Although Watson has been reduced to a historical footnote, IBM is still getting in on the action. The most advanced AI work is not happening in IBM’s Westchester, New York, headquarters, but much of it is open-source and has a short shelf life. Tailoring Silicon Valley’s hand-me-downs can be a profitable business. Yusuf invoked platoons of knowledge workers armed with the tools of the 20th century. “You’ve got people with PDFs, highlighters,” he said. IBM can offer them programs that help them do better—that bump their productivity a few points, or decrease their error rates, or spot problems faster, such as faults on a manufacturing line or cracks in a bridge.
Whatever IBM makes next won’t fulfill the promise implied by Watson’s early run, but that promise was misunderstood—in many ways by IBM most of all. Watson was a demo model capable of drumming up enormous popular interest, but its potential sputtered as soon as the C-suite attempted to turn on the money spigot. The same thing seems to be true of the new crop of AI tools. High schoolers can generate A Separate Peace essays in the voice of Mitch Hedberg, sure, but that’s not where the money is. Instead, ChatGPT is quickly being sanded down into a million product-market fits. The banal consumer and enterprise software that results—features to help you find photos of your dog or sell you a slightly better kibble—could become as invisible to us as all the other data we passively consume. In March, Salesforce introduced Einstein GPT, a product that uses OpenAI’s technology to draft sales emails, part of a trend that Evans recently described as the “boring automation of boring processes in the boring back-offices of boring companies.” Watson’s legacy—a big name attached to a humble purpose—is playing out yet again.
The future of AI may still prove to be truly world-changing in the way that Watson once suggested. But the only business that IBM has managed to disrupt is its own. On Monday, International Workers’ Day, it announced that it would pause hiring for roughly 7,800 jobs that it believes AI could perform in the coming years. Vacating thousands of roles in the name of cost-saving measures has rarely sounded so upbeat, but after years of positive spin, why back down now? Yusuf swore that IBM’s future is just around the corner, and this time would be different. “Watch this space,” he said.