A happily retired dentistheard some of AI's own builders warn the world,and at 81ended his retirement that day.

Why would anyone listen to a retired dentist about artificial intelligence? His answer is his life.

At Cornell, his published biography says, he took a double major in economics and chemistry: one subject for the tooth, one for the ledger. In 1969 he graduated from the Harvard School of Dental Medicine with his DMD, and later taught there. For about ten years, by his account, he taught dental economics to dental students, at Harvard and at Columbia: not how to repair a tooth but what a tooth costs, who pays, and why.

Then he turned from the tooth to the ledger. Dental insurance reached only a few hundred thousand people when he started, he says. His biography lists what followed: he founded and led American Dental Examiners, Better Benefits and US Dental Plan, created the Dental Health Alliance, and designed the dental plan for the United States government. He tells it himself: “I spent forty years building the dental insurance industry. I built Protective Life’s dental insurance business from the ground up, and I founded Dental Health Alliance, a national dental network later acquired by Sun Life Financial. I owned a reinsurance company and worked alongside Swiss Re and Hannover Re.”

Harvard kept him close. He advised the dental school’s dean for more than a decade, his biography says, and on June 7, 2019, the school gave him its Distinguished Alumni Award, “recognized,” in its words, “as an innovator in the field of dental insurance.” Accepting it, he said: “This institution has the ability to inspire its graduates.”

The craft under all of it was pricing risk: putting a number on what might go wrong, deciding who should carry it, and collecting that number before the harm arrives rather than after. Hold that thought. It returns.

He describes himself plainly.

Good student, silver spoon, but always hard work and I love challenges. Change is what you expect and prepare for, Clear vision of the future has always come naturally and Confidence I can do anything and I do. I am 81 years old, a dentist, and 5’4”

He tells, against himself, that when he sat his school’s entrance exam in seventh grade he was told he had the lowest score of anyone actually admitted.

And he sailed: by his count, his own boat the equivalent of three times around the world. His family has been coming to Nantucket for more than fifty years, and his letters are signed from the island.

He has never worked alone. The people he has built things with, in one line: “We are all from the for Profit world. We are highly action oriented. We call ourselves the doers.”

Of all the dangers he heard described, Bob chose the one closest to the ground. Children.

“That subject is the very reason I came out of retirement,” he wrote to a friend. Two efforts carry it, deliberately apart.

The first is a nonprofit that tests: Keep AI Safe, led by Yarin Kirchen, its co-founder and chief executive. It tests the AI toys, chatbots and companion apps children actually use, publishes what it finds, is building a safety standard for children’s AI, and takes no money from the companies that make those products. Bob co-founded it with Yarin and is one of its principal funders. He does not run it; its work is independent of his companies. He insisted early that certification stay “completely arms length” from the children’s company he would co-found.

The second is a company that builds; he has a financial interest in it. Boundlia, co-founded in 2026 with Evren Yiğit, who leads it, began with talking toys and grew into a platform where companies building AI for children can define and test what the AI may draw on, remember (with parents in control), do and refuse. Limits on what a system can reach, store and do can be built in; what a language model says can only be narrowed and tested, never guaranteed, so testing before launch is part of the design. It is not finished.

Bob’s own conviction, his and not a Keep AI Safe standard, is simple and unfashionable. A general-purpose chatbot, however brilliant, is not in his view for a child under thirteen. Thirteen is his line, not a settled finding, though American privacy law draws it there. On privacy and safety he is absolute. “Once a child speaks or writes through AI they are not protected.” And:

A 99 percent safety rating is a complete failure when the remaining one percent can mean serious harm to a child, or worse.

No system, bounded or not, reaches zero. His argument is about where the bar is set for a child, and who must prove it has been met before a product ships. The rare failure, he holds, is the one that counts.

For children already using these tools, his instinct is to make the child work: “We are better off insisting on education for children, with games reinforcing that education.”

He is candid that children also bring very different people to one table, and that they are the point, not the means. “No part of humanity will oppose keeping children safe,” he wrote in July 2026. “And all agree that children are the most vulnerable. So we lead with them to move the entire train.”

In the summer of 2026 the thing Bob had been circling for a year came into focus. It was very old.

“The root problem,” he wrote in July, “is that we have never clearly established that the owners of these systems are responsible for the harm they cause — a principle as old as the Bible: man must be accountable for his actions.” To a colleague he traced it to Deuteronomy and put it in his own words: “He who builds it is responsible for its consequences.” Every faith, he says, teaches some version of this. Only with the most powerful tools ever built had the rule gone missing.

He is blunt about the excuse he will not accept.

You can not say I did so much good it is ok if I kill a few children.

What separates Bob’s argument from a sermon is where he learned it. His years in insurance taught him, he says, that safety holds when ignoring it becomes too costly. In his telling, regulators arrive after the harm and write the rule, while insurers, who will pay, have every reason to arrive first, if the law makes the builder liable and the harm can be measured. Lloyd’s Register, he points out, was classing ships before governments regulated their safety, and underwriters wrote building codes that cities later adopted. He ran the same architecture, he says, for twenty years in dental coverage.

His proposal borrows from the oldest market he knows. Define the harm so it can be measured. Make the builder liable for it. Let insurers put a price on that liability, and watch what happens to safety budgets. ”…Go after their pocketbook and they will pay attention.” He calls it insurance that pays for safety instead of asking permission for it.

He is putting his own hand to it, too. In August 2026 he began setting up a small consulting practice to help insurers learn to cover AI agents, with the kind of underwriting and claims manuals he once wrote for dental coverage. It is at an early stage.

He is careful about one limit. “We must only insure what is measurable.” That rule is part of why he keeps children at the front: harm to a child has the broadest moral agreement of any harm being debated, and some of it can be tested directly: what a product says to a child, what it keeps, whom it tells. When a scientists’ working group left children off its list of urgent domains, partly, as he read it, for political feasibility, he answered in a draft letter that insurance “does not require political feasibility, only measurability, and harm to children is measurable.” Then: “I will defend that line to anyone.”

“It is not too much to ask,” he wrote in the same draft. “It is, as it happens, the one demand civilization already knows how to enforce.”

Half of Bob’s work is a warning. The other half is a demonstration that the same tools, kept inside the right fences, can do a great deal of good. He has said that both grew from the same concern: the risks of AI developed without constraint.

So in 2026 he started building companies again.

Medical Discoveries is a development-stage company that Bob co-founded and chairs. It is the one effort here that sits outside the children’s work: AI for good in medicine. The idea came from his son’s published research on target trial emulation, a method that designs the analysis of existing records, such as hospital notes, to emulate a randomized trial. Bob says the company was inspired by his son’s use of it. That research is independent of the company and is not evidence for it.

Hospitals, in Bob’s view, are full of knowledge that almost no one can use. Much of what a doctor or surgeon learns about a patient is typed into free-text notes, and there it stays. The hospital already holds the raw material, as he sees it; what it lacks is the layer that makes the raw material usable. The company’s aim, in the careful words its team chose for launch, is to turn that documentation into structured information that can help researchers investigate clinically meaningful questions, and to help hospitals use it in their own work: in medication safety, in keeping accreditations current, in giving a physician a one-page view of where a patient stands. Data used for research would have identifiers removed under the hospital’s privacy rules, with any key kept inside the hospital. All of it is in development. Bob is not promising cures.

The company is led by a physician chief executive, and Bob spent the summer of 2026 assembling its board. He is not a physician himself, and he does not pretend to be. He is a dentist who spent decades building large systems in health care and insurance, and that is the experience he brings to the table.

Boundlia, the bounded-AI platform for children described earlier, is the other company on the building side of his ledger. Of that one he wrote, in September: “I am in the business of building the impossible.”

What he is building, taken together, is an argument made in companies rather than speeches: that intelligence can be put to work for people without being let off the leash.

Read about this work →