Everybody Wants to Slow AI Down. Which Part?

A recent AP-NORC poll found that 64 percent of Americans think artificial intelligence is developing too quickly. I understand the feeling. By the time I figure out what one version can do, somebody has introduced its replacement. Sometimes the version I am using changes without anyone asking me.

But I also have a few things I would like it to hurry up with. Sorting out health insurance policies comes to mind. So does finding a suitable alternative place to live without making me read seventeen websites and telephone twelve people.

Apparently, I want artificial intelligence to slow down and speed up. Before declaring myself confused, however, I think we need to ask what, exactly, we want to slow down.

Better Help Would Be Welcome

In my personal use, I am generally pleased when AI gets better. I use it to investigate subjects, compare alternatives, organize information, and help develop these articles. Better answers and fewer mistakes would be welcome. So would less time spent explaining what I meant.

There are frustrations. AI can misunderstand a question, produce an impressive answer with questionable facts, or confidently recommend something that does not meet the requirements we just discussed. Anyone who has used it extensively has probably collected a few examples.

But there is usually a person between its answer and the action. In this case, that person is me. I can question the result, check a source, revise the instructions, or decide the exercise has wandered off into the weeds.

That does not eliminate the risks. Bad information can lead to bad decisions. Still, I see little reason to ask developers to slow their efforts to make these tools more useful and reliable.

Giving those tools permission to act introduces a different question.

From Answers to Actions

An individual might authorize an AI agent to pay bills, make purchases, or operate an investment account. A business might connect agents to purchasing, billing, customer records, scheduling, and communications. It could deploy thousands of them, with some assigning work to others.

There are considerable potential benefits. Businesses spend enormous amounts of time moving information between systems and people. Individuals spend plenty of time on their own administrative chores. I would happily surrender some of mine.

But asking AI to explain a bank statement is one thing. Giving it permission to move the money is another.

The number of agents matters, but so do their permissions and connections. One agent with broad authority could cause substantial trouble. Thousands with carefully limited roles and effective supervision might perform quite well.

A mistaken instruction can also travel. An incorrect order might trigger a payment, a shipping request, and changes to several records before someone notices the original problem.

That introduces a timing problem. People may establish boundaries and review activity periodically, while an automated system can initiate further actions between those reviews. Having someone watching is useful only if the problem can be detected and contained before it spreads too far.

Some limits therefore need to operate as the system acts: restricting what it can do, how much it can spend, and when it must wait for approval. A report tomorrow morning may explain what happened. It may arrive considerably too late to prevent it.

Sometimes They Are Just Experimenting

Commercial deployment is not the only setting that deserves attention. Researchers are also testing what AI can do.

Anthropic recently reported experiments examining steps involved in AI replication. A model with nine billion parameters successfully got AI inference—the process of running a model to produce responses—working on eight of nine hardware types tested. The report examined equipment ranging from desktop computers to inexpensive Android TV boxes. These capabilities are not confined to giant data centers.

That result did not, by itself, demonstrate complete autonomous replication. In the broader replication tests, the researchers simplified the challenge of gaining access to another device: the test devices had a remote debugging connection that required no authentication. That let them examine subsequent steps without also requiring the AI to overcome normal access protections.

This research can be valuable. Finding out what systems are capable of may help identify problems before those capabilities appear elsewhere.

But “we are experimenting” still leaves questions about containment, permissions, monitoring, and how an experiment is stopped. Exploring a capability and managing the experiment belong together.

Most ordinary users will never conduct anything resembling those tests. We may nevertheless encounter the same underlying issue on a smaller scale: how much freedom should we give a system whose behavior we cannot fully anticipate?

New Technology, Continuing Obligations

My engineering background includes work on the Saturn rocket program and large communications systems. That does not qualify me to prescribe a complete testing system for artificial intelligence. It does leave me with an appreciation for checking things before putting them into service.

AI presents challenges that the engineering practices I grew up with may not fully address. Its responses can vary, its uses can be open-ended, and its failures can be difficult to anticipate.

We should be honest about those differences. You cannot take a rocket-testing manual or a network guide, substitute “chatbot” for “rocket” or “switch,” and declare the problem solved.

But the obligations behind those practices still apply. Understand what you can about the system. Test it for its intended work. Establish operating limits. Monitor performance. Provide a way to intervene. Accept responsibility for what you put into service.

If familiar methods are insufficient, developing better methods is part of the job. Novelty does not make those obligations disappear.

Who Explains the Permissions?

For individuals, there is another practical question: who provides the training or advice before we authorize an agent to act?

The company offering the agent should explain its permissions in language customers can understand. Services granting access should provide effective limits and a clear way to withdraw permission. Users have responsibilities too, but “be careful” is a poor substitute for understandable controls.

A spending limit enforced by a payment system offers a different kind of protection from telling an AI, “Please don’t spend too much.”

We cannot reasonably expect every customer to become an AI safety engineer before using a service advertised as convenient. Nor should a business collect the savings from automation while handing customers the unpaid work of repairing its mistakes.

So which part of AI do I want to slow down?

I welcome faster progress in useful assistance. I want the granting of consequential authority to proceed at a pace its safeguards can support. Our ability to manage AI needs to keep up with our ability to build it.

Meanwhile, it can get back to sorting out my health insurance. That part has already taken long enough.