AI's last mile is law enforcement

I was thinking about how much of a gym's front desk could already be run by software.

A new member could walk in, talk to an AI agent, choose a plan, sign the waiver, pay, and receive an access pass. The same agent could handle freezes, renewals, and most support requests. A remote employee could pick up the handful of cases that go wrong. None of this requires a person to sit at a desk all day.

Still, I would not expect many gyms to remove that person.

The obvious reason is vandalism. Someone can walk in behind a member, damage equipment, threaten a customer, or simply refuse to leave. The employee at the desk may spend most of the day answering questions, but their presence also says that somebody is watching and can intervene.

That second job is easy to miss because it rarely shows up in a job description. It only becomes visible when the chair is empty.

Suppose the gym replaces the employee with cameras and an AI system. The system watches the entrance, notices unusual behavior, gives a warning over a speaker, and alerts a remote operator. If there is immediate danger, it calls 911 and sends the footage. For a less urgent incident, it contacts security or the police non-emergency line.

At that point, the gym can operate with no employee on site. Enforcement has simply moved outside the building.

I think this reveals a larger constraint on AI in the physical world. Most discussion focuses on whether the model can understand video, make a good decision, and use the gym's software without getting stuck. Those capabilities are improving quickly. AI does not need police to complete ordinary tasks. It needs a credible escalation path for the moment a person ignores its decision.

On the internet, software can often enforce its own rules. If a payment fails, the subscription can be disabled. If an account lacks permission, the server rejects the request. There is no physical argument with an access-control check.

A door is less certain. One member can hold it open for someone else. A warning on a screen can be ignored. Even a perfect camera does not stop the person it identifies.

That gap explains why AI may remove much of the work performed at a location before it removes the need for any human presence there. A receptionist is partly doing customer service and partly standing at the point where a rule might have to be enforced. Store clerks and transit employees often have the same unspoken responsibility.

A capable robot would move the boundary, but it would also create a more serious problem. If a robot blocks a doorway or restrains someone, it is using force. Someone has to decide when that is legal, how much force is allowed, and who is responsible when the robot gets it wrong. Better motors and better models do not answer those questions.

The authority has to come from institutions. Police response is one part of it. Courts, emergency services, private security, insurance, and clear liability rules also matter. An unattended business works only if escalation leads somewhere.

Law enforcement is not the only constraint on physical AI, but it is the final backstop when persuasion, access controls, and private security fail. That does not mean aggressive policing or sending officers for every broken gym rule. A system that calls 911 whenever its classifier gets nervous would waste public resources and put innocent people at risk. Constant surveillance has costs of its own.

What matters is that serious calls receive a predictable response, mistakes can be challenged, and the people exercising authority are accountable. Without those constraints, giving an automated system easier access to enforcement would be dangerous.

The same issue will show up in unattended stores, hotels, warehouses, and robotaxi fleets. AI can handle ordinary transactions and watch for trouble. Its weakness appears when somebody refuses to cooperate and the software runs out of buttons to press.

Physical automation will rely on a trusted institution at the other end of the call. Another leap in model intelligence will not replace it.