An Optimist’s Account of Artificial Intelligence
We find ourselves at an extraordinary pass where the biggest stories in business, finance, science, medicine, education, and warfare are all the same story: artificial intelligence (AI). Thus, the timeliness of Josh Tyrangiel’s “AI for Good,” a vigorously reported glass-half-full account of reasons to be optimistic about AI improving lives.
At a thematic level, Tyrangiel complicates my glib opening sentence by demonstrating that the AI story is, in fact, many stories, that neither AI doomers nor cheerleaders deserve our deference, and that particular applications of AI demand individualized scrutiny before we conclude that the technology will resolve all human woes (it will not) or erase all meaning from human life (it could, in theory, but it doesn’t necessarily have to).
“AI for Good” offers a series of journalistic reporting excursions to places where well- intentioned people are struggling to use AI to teach algebra and history, diagnose and monitor hospital patients, and untangle government bureaucracy. Tyrangiel explores his topic by means of narrative, not polemic. His approach reminded me of books by Michael Lewis, in which the genial, nonexpert author invites you to come along for the ride as he learns about a complex, partly hidden world filled with quirky, endearing people.
Unlike some Lewis works, however, there isn’t a Hollywood-ready plot line here, which I mean as a compliment. This is an episodic book that raises more questions than it answers. It is selective in its reach, not comprehensive. And it doesn’t purport to explain in any detail how the technology works. But it’s an engaging antidote to the usual hair-on-fire coverage of the AI debate.
In the spirit of full disclosure: Tyrangiel was my boss for five years when he edited Bloomberg Businessweek in the early 2010s. Subsequently, he worked at VICE and produced documentaries for HBO, Netflix, and Apple TV. Today, he is a staff writer for The Atlantic.
Getting the Context Right
The first dozen pages of “AI for Good” set the context for what’s to follow. They remind us that AI already was playing a big role in our lives when a start-up called OpenAI rolled out ChatGPT for the general public in November 2022. The software providing your driving instructions, filtering your social media feeds, determining whether you got a mortgage loan or had your taxes audited, analyzing your medical scans, powering facial-recognition systems, and steering military drones—it all relied on AI, just not on the generative AI that so captured consumer and corporate attention beginning in late 2022.
“Artificial intelligence” is an umbrella term of vast and imprecise reach. For good and ill, AI is already deeply integrated into modern life. That doesn’t mean that either the well-established or latest iterations of AI should be accepted just because they exist. But the choices we can—and ought to—make are not all brand new ones. Just one example: For years, lenders have used AI to make split-second decisions on borrower applications; the speed and seeming efficiency are terrific, but the optimization sometimes relies on historical biases linked to applicants residing in predominantly minority neighborhoods or having sparse credit records. The upshot is that AI software, like almost any technology, can be shoddily designed or recklessly deployed; its hazards can outweigh its benefits.
The AI ship may have sailed, but we collectively face crucial decisions on how we will navigate. The consumer launch three-and-a-half years ago of generative AI caused many people to realize just how far we’d already traveled. This variety of the technology allows ordinary users to type natural-language prompts on practically any topic and receive nearly instantaneous responses phrased in confident, human-sounding terms. It’s uncanny, mesmerizing, and sometimes wrong (but improving). It’s allowing office workers to save oodles of time organizing emails, meeting notes, and company policies. It’s also eliminating certain customer service jobs, facilitating classroom tutoring (and cheating), and offering endless opportunities for pernicious manipulation, including disinformation, fraud, and revenge porn.
Should AI Teach Schoolchildren?
Among his case studies, Tyrangiel examines the application of generative AI to childhood education. The decades-long history of educational technology (“ed-tech”) is littered with failures and frauds, as Tyrangiel concisely recounts. Ed-tech’s designers and marketers have promised “personalized” education that allows students to learn at their own speed, but the reality has tended toward software that keeps students glued to screens and facilitates daydreaming, distraction, and corporate data-harvesting. The classroom attention-deficit mayhem only worsened in the 2010s with the advent of smartphones and social media, leading in recent years to a nascent move by some school districts to rip out ed-tech curricula and restrict phone use.
Amid this maelstrom, a rare and saintly online educator named Sal Khan has tried to harness generative AI to advance his mission. Khan’s celebrated nonprofit, Khan Academy, serves some 190 million students around the world, providing a library of thousands of free video lessons, most notably in math. Historically, Khan’s content has been carefully crafted, one video at a time, highlighting the founder’s wonky-but-empathetic persona. When deployed with diligent in-person teacher supervision, Khan Academy stands out as an exception in the junky ed-tech landscape.
In 2021 and 2022, OpenAI approached Khan in hopes of forming a joint venture that would lend the company some of Khan’s credibility while providing Khan Academy with ChatGPT’s capacity to respond to virtually any question and even explain its answers. Problems ensued. For all its computing muscle, ChatGPT proved “inconsistent” at elementary math, Tyrangiel writes. It “hallucinated,” the industry term for making up falsehoods. “When prompted, it could create nonexistent sources to support its nonexistent facts.” Some of these flaws were ironed out by crafting a customized ChatGPT for a system ultimately called Khanmigo, but the labor-intensive fine-tuning generated “a fresh cascade of hallucinations and unpredictability.”
Khan Academy’s strong reputation led 53 school districts with 65,000 students to try Khanmigo. The results have been mixed, Tyrangiel reports. One adventurous superintendent in Indiana saw promise but admitted that many of her students, conditioned to passively absorb huge volumes of digital content via social media, “don’t know how to ask questions,” undermining the putative benefit of a generative AI system that can answer any query. Students Tyrangiel interviewed in Newark were underwhelmed. One told him that AI generally will be “a big thing” but that Khanmigo “is not close to the same thing as an actual teacher.”
Khan himself told the author that “the self-motivated kids—the high flyers—they’re getting it. But there are some kids who just aren’t engaging. And even among the ones who are, some are only fishing for answers.”
This left me wondering whether the high-flyers would be doing fine in math classes without a hallucination-prone generative AI overlay—or without any ed-tech at all. Just because we have cool technology doesn’t mean it necessarily solves difficult problems. Reducing class sizes and putting better-trained and more generously compensated teachers in those classrooms seem like more promising, if more expensive, responses to eroding student achievement.
Another question unaddressed by “AI for Good” is whether a generative AI system purpose-built for teaching algebra and basic physics would do better than a jury-rigged ChatGPT. OpenAI’s large language models are gargantuan programs trained on huge swaths of the information stored on the internet. The enormous scale contributes to their substantive reach but also may help explain their unpredictability. Generative AI models can be assembled in a far more bespoke manner for narrower purposes than answering any question under the sun. Sal Khan and other reformers might be better advised to go that route.
Artificial Intelligence, M.D.
The potential of purpose-built AI becomes clear when Tyrangiel parachutes into the prestigious Cleveland Clinic, where physicians are improving on a variety of highly specialized AI systems to help them better assess patient scans and monitor for potentially dangerous symptoms hidden in forbidding data flows. AI can facilitate the generation and analysis of far more thorough cardiac MRIs, for instance, than an unaided-human clinician, assuming that the hospital in question can afford the expensive computer capacity that’s required.
AI ambient scribe software can operate inside an app on a doctor’s phone, eavesdropping on conversations with patients. If it works properly, this technology transcribes the audio in real time, filters out the dross and initiates action based on medically significant details. It can suggest diagnoses, craft treatment plans, and even prepare orders—all requiring the physician’s ultimate sign-off, of course.
The potential pitfalls of ambient scribe software are obvious. Without rigorous human review, a hospital is looking at a slew of misunderstandings and mistakes, not to mention a robot-driven medical malpractice bonanza. (The second season of the HBO Max emergency room drama “The Pitt” includes an episode in which a young doctor’s failure to proofread AI ambient scribe output leads to a near catastrophe.) For its ambient scribe contract, the Cleveland Clinic has turned not to OpenAI but to a much smaller, more specialized company called Ambience, whose app helped reduce physician burnout and attrition, according to one study.
Detecting sepsis, a severe form of infection that can spiral out of control and lead to death, is a major challenge for hospitals. The Cleveland Clinic has introduced an AI-driven monitoring system that flags sepsis symptoms. From 2021 to 2024, sepsis mortality declined by 40 percent, but the institution’s leaders say that AI doesn’t deserve all of the credit. Clinicians may have been more on their toes because they knew that the technology was looking over their shoulder. And the AI sepsis detector too often sounds alarms that turn out to be false positives, requiring nurses and physicians to expend unnecessary time and attention. Still, Tyrangiel concludes, reasonably, that “AI does not have to be perfect to be useful.”
The Palantir Dilemma
The variety of illustrations in “AI for Good” conveys the pervasiveness of the technology. The most moving section of the book describes a scientist’s determined quest to deploy AI to communicate with her extremely autistic son, based on his obscure vocalizations.
Yet another section recounts how military officers worked with Palantir to create a data dashboard that tracked the materials needed to deliver hundreds of millions of Covid vaccines as part of Operation Warp Speed in 2020. Co-founded by the far-right financier and Donald Trump backer Peter Thiel, Palantir has drawn criticism for enabling excessive government surveillance by means of its facial-recognition technology, among other services. It currently provides software to enable the Trump administration’s brutal anti-immigrant enforcement policies.
Many people would not categorize Palantir or Peter Thiel under “AI for Good.” But that discussion does not interest Tyrangiel, who, in a misstep, dismisses it as “kind of trivial.”
Overall, though, the book injects a welcome dose of shoe-leather reporting into the debate about AI, without succumbing to the exaggeration that too often taints the topic. It’s certainly not the last word on artificial intelligence. But it provides an important reminder: that AI may help address some difficult problems, but only with arduous human labor and copious amounts of common sense. Otherwise, buyer beware.
