There’s a real argument happening right now inside the AI industry, and almost nobody outside it has noticed yet.
It’s not a debate about whether AI agents should do more work. Everybody’s already decided that part. It’s a debate about who checks the work once there’s too much of it for a person to look at, and the two answers on the table flatly contradict each other.
Gregor Stewart, the Chief AI Officer at the security company SentinelOne, said the quiet part out loud. As agents take on more of the actual labor, the volume of output is starting to outpace any human’s ability to review it. His exact words were that you end up with so much work done and so much work to audit that you can’t truly be accountable anymore. That’s not a hypothetical. That’s a description of where things already are in fields like coding, which he said is running about a year ahead of everyone else.
So what’s the fix the industry is reaching for. The one getting traction right now is having one AI system write the work and a second, separate AI system check it. The writer writes. The editor’s only job is finding what the writer got wrong. The one rule everyone agrees on is that the same system can’t grade its own homework — it has to be a different model checking the first one.
I understand why that’s appealing. It solves the speed problem completely. A machine can review ten thousand lines of code or a hundred drafted emails in the time it takes a person to read through one. But sit with what it actually means for a second. The fix for “humans can’t keep up with checking AI” is having AI check AI, with no human required at either step. That’s not closing the accountability gap Stewart named. That’s moving it one layer deeper and hoping nobody asks what happens when the editor misses something too.
Here’s where it gets interesting, because not everyone in this industry agrees that’s the right call. Google has its own position, laid out in a white paper on secure AI agents, and it points the opposite direction. Their stance is that agentic systems need an actual human in the loop, not just a second AI checking a first one. The platform is supposed to communicate what it’s about to do and ask for approval at key decision points, before the action happens, not after. Google’s own people called this not just good practice but something too important not to regulate.
So there it is, plainly, two real positions from two real companies, not theory. One says the answer to “too much to audit” is faster auditing, done by another AI. The other says the answer is a human checkpoint that doesn’t disappear no matter how much volume the system produces. Those can’t both be the long-term answer. Either the checkpoint stays human, or eventually it doesn’t, and right now the industry is building toward both at once without admitting they’re headed in opposite directions.
I know which side of that I come down on, and it’s not close. A system checking another system can catch a contradiction or a broken citation. What it can’t do is stand in the place of a person who actually has something at stake in the outcome being true. The writer-AI and the editor-AI both have the same kind of nothing riding on the answer. Neither one loses anything if it’s wrong. That’s the entire difference between checking and accountability, and it’s the difference the writer-editor fix quietly erases the moment nobody asks the question out loud.
This is exactly the gap the Faust Baseline was built to sit inside. Not to slow agents down, and not to pretend a machine can’t write good output, because it plainly can. The point is naming where the human checkpoint has to stay, on purpose, before the volume gets so large that nobody remembers there was ever a choice to make about it. A system that discloses what it’s about to do and waits for a yes is doing something fundamentally different from two systems quietly clearing each other’s work in the background. One of those keeps a person in the room. The other just moves the room somewhere nobody’s watching.
The industry hasn’t settled this yet, and I don’t think it will settle quietly. Too much money is riding on the side that says faster auditing is good enough. But faster isn’t the same as accountable, and somewhere in the next year, something built entirely by AI checking AI is going to get something wrong in a way that actually matters, with no human anywhere in the chain who can say they caught it first. When that happens, the question won’t be which side argued better. It’ll be which side built the checkpoint before they needed it.
“The Faust Baseline Codex 3.5”
micvicfaust@gmail.com
Post Library – Intelligent People Assume Nothing
”AI Baseline Governance”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
