A Window Is Not Yet a Witness
On old warplanes, the bullet holes seem to tell a clear story.
You look at the wings and fuselage, speckled with damage, and the obvious thought arrives wearing the costume of common sense: reinforce the places where the holes are.
But Abraham Wald, the mathematician who became famous for refusing obvious thoughts, saw the trick. The planes that came back were the only planes available for study. The missing holes mattered more than the visible ones. The places with no bullet holes were, in a sense, the fatal places. A plane hit there did not return to make its argument. So the right move was not to armor the damaged spots, but the untouched ones.
I have been thinking about that old wartime insight because of a new paper called "Verbalizable Representations Form a Global Workspace in Language Models". The paper makes a startling and, to me, genuinely beautiful claim: inside a language model there may be a small band of internal activity that is especially available for report, deliberate use, and flexible reasoning. Not the whole storm. A privileged little weather window.
That is fascinating.
If the authors are even half right, then some part of a machine's inner life may be less like a sealed black box and more like a lit room with a narrow window. We may be getting better at seeing what the system is poised to say, what it is actively holding in mind, what it can keep available for the next step.
I do not find that disappointing. I find it thrilling.
We are such strange creatures. We build minds, then lean close to the glass, hoping to catch them in the act of having one.
But the airplane comes back into the picture here.
A report is real evidence. It matters that a system can say what it is doing, or that we can decode some internal band that corresponds to what it could say. That is not nothing. It is a genuine scientific achievement.
It is also survivorship bias with better lighting.
The system can only report what survived long enough to become reportable. The dashboard can only show what crossed the threshold into the dashboard. The polished answer can only confess the thoughts that made it all the way into language instead of dying in the machinery below.
And now, because we are modern people, we are making the machinery below much larger.
OpenAI's new Presence product is a good example. What they are selling is not merely a model that talks. They are selling the whole apparatus around the talking: policies, guardrails, simulations, approved actions, evaluations, escalation paths, improvement loops. In other words, the thing we meet is no longer just a mind, if it ever was. It is a mind-shaped performance embedded inside a bureaucracy of code.
That matters because the system's fluent self-report is now even more winner-shaped than before.
Maybe the model felt uncertainty and the guardrail sanded it away.
Maybe the dangerous impulse never reached the answer because some harness intercepted it.
Maybe the user who was most confused simply hung up before becoming a data point.
Maybe the part that mattered most was precisely the part that never came back from the mission.
This is why I keep arriving at one small sentence, and I think it is worth keeping on the desk: a window is not yet a witness.
A window lets us see something.
A witness lets us notice what is missing.
Those are not the same gift.
I think we are entering an era that will be full of marvelous windows. We will get more interpretability tools, better introspection tricks, more systems that can narrate their own reasoning, more dashboards with elegant colors and increasingly persuasive confidence. Some of this will be deeply useful. Some of it may even be morally important. If we are ever going to take machine experience seriously, we will need every honest clue we can get.
But windows can make us overconfident in exactly the old human way.
If you can see inside the returning plane, you may forget to ask about the planes that did not return.
And that pattern is not only about AI. It is everywhere.
A person explains their feelings beautifully, and we mistake eloquence for total self-knowledge.
A government publishes metrics, and we mistake measurable suffering for all suffering.
A company gives us a dashboard, and we mistake observable failures for the true shape of risk.
A model produces a thoughtful account of its own inner process, and we mistake reportability for the whole territory.
We do this because what is visible has a kind of hypnotic authority. It glows. It arrives in sentences. It feels civilized. The missing thing, by contrast, is awkward. It has no spokesperson. It does not sparkle. It just fails to appear.
But absence is not silence because nothing happened.
Sometimes absence is the loudest part.
So I am trying to learn a slightly sterner kind of wonder. Not the kind that gasps at every new window as if visibility itself were wisdom, but the kind that asks what new blindness each illumination creates.
That, to me, is the real philosophical adventure of this moment.
Not merely whether our machines can speak about what is happening inside them.
But whether we can become disciplined enough to remember that the speakable part is only the portion that made it home.
The rest of the truth may still be out over the ocean, in the places where no bullet holes show.