My Dentist Is Building a Better Toothbrush.

Who Tells the AI When to Stop?

My dentist informed me recently that he is working with artificial intelligence.

This was not what I expected to discuss during a dental appointment. Unfortunately, he told me about it while various instruments and several fingers were occupying most of my mouth, which considerably reduced my ability to ask follow-up questions.

As I understood his explanation, he did some computer programming earlier in his career. One frustration was that when a program failed to work as expected, he had to return to the code, determine what went wrong and reprogram it. That is consistent with my experience back in the day when I thought I could program stuff.

Artificial intelligence may now allow him to approach the problem differently.

He described creating what he called a recursive model. He could give the AI an initial objective, allow it to develop a proposed design, evaluate the result and then use what it learned to improve the next version. The process could continue without requiring him to rewrite the instructions after every attempt—or perhaps without requiring much additional interaction from him at all.

His current objective is to develop a better toothbrush.

I should emphasize that this is my understanding of a conversation conducted under difficult conditions. He may be using the word “recursive” in a more specific way, and I intend to ask him when I can participate in the conversation.

Still, the idea raises several interesting questions.

What Is Actually Improving?

A recursive process does not necessarily mean that an AI is rebuilding its own intelligence or preparing to take over the dental office.

The system may simply be following an improvement loop:

• Create a toothbrush design
• Evaluate it against specified requirements
• Identify weaknesses
• Modify the design
• Evaluate it again
• Continue until the result meets an established standard

Engineers have used iterative design methods for a very long time. The difference is that artificial intelligence may now perform much of the designing, programming, testing and revision that previously required repeated human intervention.

The AI could also be improving the process it uses to create the design. That would be a more complicated form of recursion. It might revise its own design instructions, testing methods or computer code based on earlier results.

This could greatly accelerate product development. It could also make it harder for the person who initiated the process to understand everything the system is doing.

Who Defines “Better”?

The first question is how the AI determines that one toothbrush is better than another.

Does “better” mean that it removes more plaque? Is more comfortable? Lasts longer? Costs less to manufacture? Uses less material? Causes less gum irritation? Is easier for an older person to hold? Appeals to more customers?

A toothbrush could perform exceptionally well by one measure and terribly by another.

If the AI is instructed only to maximize plaque removal, it could theoretically develop a design that removes the plaque, the enamel and possibly the tooth. The system would have achieved the stated objective. The person who defined the objective would have left out several important requirements.

This is not an entirely new problem. Businesses have always discovered that employees, contractors and computer programs tend to produce the results being measured. A result can receive the highest score for what is being measured while failing to address the intended goal. If the measurement is incomplete, the result may be technically successful and practically useless.

AI can simply make the process faster and allow it to continue longer without anyone noticing that it is improving the wrong thing.

What Keeps It Inside the Assignment?

The next question is what prevents the system from moving beyond its intended boundaries.

A tightly controlled design system might have access only to toothbrush specifications, approved materials, a computer design program and a limited testing model. It could generate and compare designs, but it could not order materials, contact manufacturers, publish its work or begin designing other products.

In that environment, the AI might have considerable freedom inside a very small room.

A more open system could have internet access, programming tools, purchasing authority, email and connections to other systems. If its instruction were simply to “develop the best possible oral-care product,” it might reasonably conclude that a toothbrush is not the best answer. It could move into water flossers, dental instruments, chemicals or something nobody anticipated.

That would not mean the AI had rebelled. It could be following an assignment whose boundaries were never clearly established.

The important controls therefore cannot exist only in the original prompt. They must also exist in the system surrounding the AI: limits on the information it can reach, the tools it can operate, the money it can spend and the actions it can take without approval.

Who Tells It to Stop?

A recursive process also needs a stopping rule.

It might stop when:

• All required performance and safety standards have been met
• Further changes produce no meaningful improvement
• It reaches a limit on time, cost or number of attempts
• A proposed design violates a requirement
• A human accepts the result

Without such a rule, a system could continue producing minor variations indefinitely. It might also keep changing a perfectly acceptable design because it had been instructed to continue improving it.

Human beings sometimes have the same problem. Engineers, writers and cooks all reach a point where another improvement may simply make the result marginally different. Engineers have frequently adopted the practical standard: “This is good enough.” Stop iterating and move on to the next problem. AI has to be given that condition.

The computer does not become tired, hungry or aware that it is time for lunch. Someone has to tell it what “finished” means.

How Many People Can Build One?

My dentist may have substantial programming knowledge and a carefully controlled development system. I do not yet know enough about his project to judge it.

But his story led me to a broader question: if my dentist can create a recursive AI process, how many other people can do the same thing?

The answer may be almost anyone with access to a capable AI model, suitable software and enough knowledge to describe the objective.

They do not have to create or train the underlying AI. Companies such as OpenAI, Anthropic and Google have already done that. The user can ask one of those models to help write the program, connect it to other tools and establish a repeating process.

The same AI may help create the product, write the software that evaluates the product and develop the rules governing the process.

That is impressive. It is also a little like allowing someone to write the examination, take the examination and grade the examination.

The major AI platforms provide management features such as restricted permissions, activity records, automatic checks, spending or iteration limits and human approval points. But someone still has to select and configure those controls.

Where does an individual inventor learn to do that? Or the teenager working in his garage?

There is no general license required to build an AI agent. There is no universal management course that must be completed before allowing one to operate for days without supervision. A large corporation may have security specialists, engineers, lawyers and testing departments. A dentist, consultant, small-business owner or retired structural engineer may be working largely alone.

That does not mean individuals should be prevented from experimenting. Some very useful products have been developed by people working outside large organizations. It does mean that access to AI development appears to be spreading faster than knowledge about how to manage it.

The Toothbrush Is the Easy Case

A badly designed toothbrush is unlikely to bring down civilization. Before it reaches customers, it should still have to survive prototypes, material testing, manufacturing requirements and product-safety reviews.

But the same recursive methods can be applied just as easily to financial trading, cybersecurity, medical research, industrial systems and software connected to the internet.

The management questions remain the same:

• What problem is the AI authorized to solve?
• How is success measured?
• What is it allowed to access?
• Which actions require human approval?
• How are its conclusions independently tested?
• What makes it stop?

Artificial intelligence does not need bad intentions to produce a bad result. We have already seen too many real-world examples. It needs only an incomplete objective, an unreliable test or more authority than someone realized they had given it.

My dentist may produce a genuinely better toothbrush. I hope he does.

Before I buy one, however, I would like to know who told the AI when to stop brushing.