accountability
Hands on the Wheel
2026 · 5 min read
The better you get at controlling hallucinations, the more blindly you trust what the agent tells you. That's the trap — and it doesn't catch beginners. It catches the people who are good at this.
Watch any population of engineers using AI agents and you see two groups. The beginners are still being burned by hallucinations — or they know enough to be afraid of them, and hesitate to bring the tool anywhere near real work. That fear is healthy. It's the second group that should worry you.
The second group handles agent models fluently. They know the prompting patterns, the context windows, the failure modes. They can get a hallucination under control in minutes. And somewhere around the point where controlling hallucinations becomes routine, something quietly worse happens: they start believing everything the agent says.
Answers too good to check cheaply
Here's the mechanism, because it isn't stupidity — it's economics. At a certain level of skill, the agent's answers become genuinely hard to distinguish from the truth. Fabricated facts and real ones come out in the same confident register, at the same depth of detail. Telling them apart is possible, but it costs exactly the things nobody has: time, and attention.
Meanwhile the volume keeps growing. The AI multiplies how much you can produce faster than it multiplies how much you can verify. We now live in a world where there is so much output — much of it ours — that understanding our own documents has become the bottleneck. The tool that was supposed to do the reading for you has made the reading the job.
The fear arrives on schedule, at the moment of sharing. As long as the answers stay on your screen, doubt is a private matter. The moment the output has to go to a teammate or a customer, every unchecked claim becomes a question someone will ask out loud.
Why modeling people feel this the most
For people who build models, this stress is not abstract. What crosses over to the customer becomes silicon. It gets packaged, powered up, and tested — and if a serious bug is sitting in there, the result is a revision. Months. Money. A conversation nobody wants to have.
And the failures that cause this are never exotic. They are exactly the problems a DV team exists to catch — which is what makes the aftermath so bitter. An accident is never one mistake. It's coincidences stacking: the case the test didn't cover, the day the reviewer was tired, the assumption two people made independently because it was the reasonable one. Each coincidence is survivable. Overlapping, they produce a hell with the structure of a drama — you can even narrate it afterward, beat by beat, and every beat was somebody doing the sensible thing.
What level is driving?
Ask yourself honestly: the AI currently driving your work — if it were an autonomous car, what level would it be?
For most of us the honest answer is Level 2. Assisted, not autonomous. The hands must stay on the wheel, and the driver — not the manufacturer — is responsible for what happens. We all accept this about cars. About agents, the second group has quietly started driving with their eyes closed, because the steering got good enough to make supervision feel optional.
Level 2 is not an insult. It's a contract. The system does the work; the driver owns the outcome. That contract doesn't loosen as the system gets better — if anything, it's the only thing that stays constant while everything else improves.
The stance that follows
So the rule we work by is simple to state and expensive to keep: you are responsible for the model you accept. Not the agent that drafted it. Not the tooling that generated the correlation report. If it went out under your name, the accountability is yours — and the only way to make that accountability bearable is to make the model genuinely yours.
Which means: if you received a model, read the spec. Then read it again. Keep reading until the model is no longer something you were handed — until you can predict where it's faithful, where it's guessing, and where it's silently wrong. A model isn't yours when it runs. It's yours when you can defend every line of it without opening the file.
The agent can drive. You're the one holding the license.