We find ourselves at an extraordinary mountain pass: artificial intelligence (AI), where the biggest stories in business, finance, science, medicine, education, and war are all the same story. Josh Tyrangiel’s AI for Good is thus timely, as it energetically reports on reasons to be optimistic about how AI will improve our lives.
On a thematic level, Tyrangiel teaches that the story of AI is actually many stories, that neither its destroyers nor its cheerleaders deserve our respect; The specific application of technology complicates my glib opening sentence by showing that individual scrutiny is needed before concluding that technology solves all human suffering (it doesn’t) or erases all meaning from human life (theoretically possible, but not necessarily).
AI for Good offers a series of journalistic trips to places where people of goodwill are striving to use AI to teach algebra and history, diagnose and monitor hospitalized patients, and untangle government bureaucracy. Tyrangiel explores his themes through stories, not polemics. His approach reminded me of Michael Lewis’s books. In the book, the mild-mannered non-specialist author invites you along for the ride as he learns about a complex and partially hidden world filled with quirky and endearing people.
However, unlike other Lewis works, this film does not have a Hollywood-style plotline. This is meant as a compliment. This is an episodic book that raises more questions than it answers. It is selective in scope and not comprehensive. It also does not provide a detailed explanation of how this technology works. But it’s a fascinating antidote to the usual intense coverage of AI debates.
In the spirit of full disclosure: Tyrangiel was my boss for five years when I edited Bloomberg Businessweek in the early 2010s. He then worked at VICE and produced documentaries for HBO, Netflix, and Apple TV. He currently serves as a staff writer for The Atlantic.
Get the context right
The first few dozen pages of “AI for Good” set the background for what follows. These are reminders that AI was already playing a big role in our lives when a startup called OpenAI released ChatGPT to the public in November 2022. Software that provides driving instructions, filters social media feeds, determines whether you’ve taken out a mortgage or had your taxes investigated, analyzes medical scans, powers facial recognition systems, and pilots military drones, all relied on AI. Generation AI It has attracted the attention of consumers and businesses since the second half of 2022.
“Artificial intelligence” is a broad and imprecise umbrella term. For better or worse, AI is already deeply integrated into modern life. That doesn’t mean we should accept established or modern AI just because it exists. But what we can and should choose is not only something completely new. This is just one example. Lenders have been using AI for years to make instantaneous decisions about borrower applications. Although speed and efficiency seem impressive, optimizations can rely on historical biases related to applicants residing in minority-dominated areas or applicants with sparse credentials. As a result, AI software, like almost all technology, can be sloppily designed or recklessly deployed. The risks may outweigh the benefits.
The AI ship may have sailed, but we collectively have to make important decisions about how to navigate it. The launch of generative AI for consumers three and a half years ago made many realize how far we had already come. This range of technologies allows ordinary users to enter natural language prompts on virtually any topic and receive near-instantaneous responses expressed in confident, human-like language. It’s creepy, fascinating, and sometimes wrong (but improving). This allows office workers to save a lot of time organizing emails, meeting notes, and company policies. It also eliminates certain customer service jobs, facilitates tutoring (and cheating) in the classroom, and provides endless opportunities for harmful operations such as disinformation, fraud, and revenge porn.
Should AI be taught to schoolchildren?
In his case study, Tyrangel explores the application of generative AI to early childhood education. As Tyrangiel succinctly describes, the decades-long history of educational technology (“edtech”) is littered with failures and injustices. While edtech designers and marketers have promised a “personalized” education that allows students to learn at their own speed, the reality is that the trend is toward software that keeps students glued to screens and facilitates daydreaming, distraction, and corporate data collection. The disruption caused by inattention in the classroom was further exacerbated in the 2010s with the advent of smartphones and social media, leading to early moves in recent years to eliminate educational technology curricula and limit phone use in some school districts.
In the midst of this chaos, a rare and saintly online educator by the name of Sal Khan set out to leverage generative AI to advance his mission. Mr. Khan’s prominent nonprofit organization, Khan Academy, serves approximately 190 million students worldwide and offers a library of thousands of free video lessons, with a particular focus on mathematics. Historically, Mr. Khan’s content has been carefully crafted, one video at a time, to highlight the founder’s eccentric but empathetic persona. Khan Academy stands out as an exception in the janky ed-tech industry, with dedicated face-to-face instruction from teachers.
In 2021 and 2022, OpenAI approached Khan in hopes of forming a joint venture that would provide Khan Academy with ChatGPT’s ability to respond to virtually any question and even explain its answers, along with providing the company with some of Khan’s credibility. A problem has occurred. Despite its computational power, ChatGPT proved to be “inconsistent” in elementary mathematics, Tyrangiel wrote. It was a “hallucination”, an industry term meaning to fabricate a falsehood. “When prompted, they may create non-existent sources that corroborate non-existent facts.” Some of these flaws were eventually addressed by creating a customized ChatGPT for the system called Khanmigo, but the labor-intensive tweaks introduced “a cascade of new hallucinations and unpredictability.”
Khan Academy’s strong reputation has attracted 65,000 students from 53 school districts to try out Canmigo. Tyrangiel reported mixed results. One enterprising school superintendent in Indiana was hopeful, but acknowledged that the putative benefits of generative AI systems that can answer any question are undermined because many students, conditioned to passively absorb vast amounts of digital content through social media, “don’t know how to ask questions.” Tyrangiel, a student interviewed in Newark, was overwhelmed. One said that while AI in general will be a “big thing,” Cummigo is “nowhere near the equivalent of a real teacher.”
Kahn himself told the author, “Spontaneous children, the high flyers, understand. But some of them are not interested at all. And even among those who are interested, there are some who are just looking for answers.”
This left me wondering whether high achievers would be able to do well in math classes without a hallucinogenic generative AI overlay, or even without any pedagogical technology at all. Just because we have great technology doesn’t necessarily mean it will solve difficult problems. Reducing class sizes and filling those classrooms with better-trained, more generously compensated teachers seems like a more likely, if more expensive, response to declining student achievement.
Another question not addressed in “AI for Good” is whether a generative AI system built to teach algebra and basic physics would perform better than the jury-rigged ChatGPT. OpenAI’s large-scale language models are huge programs trained on vast amounts of information stored on the internet. Its huge size contributes to its substantial scope, but may also help explain its unpredictability. Generative AI models can be assembled in a much more bespoke way for narrower purposes than answering general questions. Sal Khan and other reformers may be encouraged to go down that path.
Artificial Intelligence Doctor of Medicine
The potential of dedicated AI becomes apparent when Tyrangiel parachutes into the famed Cleveland Clinic. There, a variety of highly specialized AI systems are being refined to help doctors better assess patient scans and monitor potentially dangerous symptoms hidden in forbidden data flows. For example, assuming the hospital in question has the necessary expensive computing power, AI can facilitate the generation and analysis of cardiac MRIs with far more detail than a clinician without human assistance.
The AI ambient scribe software works within an app on a doctor’s phone and can eavesdrop on conversations with patients. When working properly, this technology transcribes audio in real time, filters out dregs, and initiates action based on medically important details. They can also suggest diagnoses, create treatment plans, and write instructions. Of course, all of this requires final approval from your doctor.
The potential pitfalls of ambient scribing software are obvious. Without rigorous human review, hospitals are open to numerous misunderstandings and mistakes, not to mention medical errors caused by robots. (The second season of HBO Max’s emergency room drama “The Pit” includes an episode in which a young doctor fails to calibrate the output of an AI ambient scribe, leading to near-disaster.) For the ambient scribe contract, the Cleveland Clinic turned to a much smaller, more specialized company called Ambience instead of OpenAI. One study found that the app helped reduce physician burnout and turnover.
Sepsis is a serious infection that can get out of control and lead to death, and is a major challenge for hospitals. Cleveland Clinic has introduced an AI-driven monitoring system that flags symptoms of sepsis. From 2021 to 2024, sepsis mortality rates decreased by 40 percent, but agency leaders say AI doesn’t deserve all the credit. Clinicians may have been more nervous because they knew technology was looking over their shoulders. Additionally, AI sepsis detectors too often sound alarms that turn out to be false positives, requiring nurses and doctors to spend unnecessary time and attention. Still, Tyrangiel reasonably concludes that “AI doesn’t have to be perfect to be useful.”
Palantir’s Dilemma
The colorful illustrations in “AI for Good” convey the widespread use of technology. The book’s most moving section explores a scientist’s determination to deploy AI to help his severely autistic son communicate based on his slurred vocalizations.
Yet another section details how, as part of Operation Warp Speed in 2020, military officers worked with Palantir to create a data dashboard to track materials needed to distribute hundreds of millions of coronavirus vaccines. Palantir, co-founded by far-right capitalist and Donald Trump supporter Peter Thiel, has drawn criticism for using facial recognition technology and other services to enable excessive government surveillance. It currently provides the software that enables the Trump administration’s brutal anti-immigrant enforcement policies.
Many people wouldn’t classify Palantir or Peter Thiel as “AI for good.” However, this discussion does not interest Tyrangiel, who dismisses it as a “trivial matter”.
But overall, the book injects a welcome amount of shoe-leather reporting into the AI debate without succumbing to the hyperbole that too often taints the topic. Of course, this is not the last word on artificial intelligence. But it does provide an important reminder. AI may help address some difficult problems, but it requires hard human labor and a great deal of common sense. Otherwise, buyer beware.
