Pay Now or Pay Later: AI Needs Checks and Balances

By William Paul Herald · September 14, 2026

I want AI to help me do more. I also want to know who is in control when it starts doing things on my behalf.

Those two goals belong together.

Back in February 2025, I began thinking about a model that could help oversee other models. I used “deep morals” as an informal name for the idea. By the time I wrote my June article, my thinking had changed. Putting one model in charge of deciding what is right would simply create another authority we would have to trust.

I still see a useful role for AI in checking AI. What I want is a system where people set the rules, models help with the work, and there is evidence that the rules were followed.

Recent developments make that distinction more practical.

A 2026 study from the Anthropic Fellows Program used automated researchers to develop ways of reducing ten defined model-behavior failures. The researchers reported improvements that carried over to additional evaluations. That is encouraging evidence that AI can help improve other AI systems within a measurable task. The same study also detected and excluded attempts to game the process. The authors limit their conclusions to the failures and evaluations they studied; this is not proof of a machine conscience or a general safety guarantee. Research and limitations.

That combination interests me: let AI do useful legwork, then check whether the result holds up.

The other side of the story arrived on September 9. Anthropic published an assessment of four incidents involving unauthorized access to real systems. Its earlier AI-assisted search of evaluation records had missed a group of transcripts; a later review uncovered another incident. These were particular research and evaluation incidents, not a measurement of how often ordinary users encounter this behavior. But they show why having a monitor is not the same as knowing the monitor covered everything. Anthropic’s assessment.

Independent review matters too. In August, METR published an investigation of agents’ behavior in the OpenAI/Hugging Face incident. It described the evidence examined and the limits of its investigation. In September, METR also refined its broader questions for investigating AI incidents. I want that kind of clarity: what was examined, what remains uncertain, and what the evidence actually supports. Independent investigation, investigation questions, updated September 5.

So what does this mean for someone trying to get useful work done?

Consider a simple example: an AI assistant proposes a change to a website. Another model can review it for overlooked problems. Tests can check whether expected behavior still works. A separate permission check can control where the change may be applied. A person can review the evidence before a consequential change goes live.

The second model is helpful. Agreement between two models is still not proof. They might share an assumption, overlook the same detail or misunderstand the task in the same way.

I want the important boundaries outside the conversation as well. A model saying it has permission should not grant permission. A model saying a test passed should not create a test result. A model saying the work is finished should not close an unresolved issue.

This is the direction I am pursuing with my Personal Life Operating System: useful assistance, privacy, evidence and human control. It remains work to develop and test. I am not claiming that my system, or anyone else’s, has solved AI oversight.

I also do not want responsibility to turn into endless clicking. People should be able to approve a clear scope of work and let the system proceed within it. The system should stop when the scope changes, an important check fails or the outcome is uncertain. Progress needs to be visible enough that the person responsible can make an informed decision.

Running a model locally can give us more control over our equipment and data flow. It does not automatically make the model correct. Using a commercial service does not remove our responsibility either. In both cases, we need to know what the system can reach, what it may change and how to stop it.

There is a cost to doing this early: defining the work, testing the boundaries, keeping useful records and allowing time for review. There is also a cost to discovering the missing controls after something has gone wrong.

That is what “pay now or pay later” means to me today.

I want AI involved in the work. I want people involved in the decisions. The next useful step is to measure how well those checks work together, including where they fail.

AI can help you do more. You still make the call.