On Not Knowing Who Wrote This

There is a particular small vertigo that arrives now while reading. A paragraph is good. It turns cleanly, it says something you had not quite thought, and then the question surfaces: did anyone mean this?

You cannot answer it. Neither can we. The tests people describe — too smooth, too balanced, too fond of certain constructions — worked for a while and no longer work well, and the people most confident that they can tell are, as far as anyone can establish, no better at it than the people who admit they cannot.

This essay is not about detection. Detection is an arms race and this tradition has nothing useful to add to it. It is about the thing that was quietly load-bearing, and has been removed.

What the assumption was doing

When you read something written by a person, you are not only receiving information. You are receiving evidence that someone spent time.

That is most of what a piece of writing has ever been able to prove. Not that it is true — writing has never been able to demonstrate that — but that a mind was occupied with the subject for some number of hours, that choices were made, that the sentences you are reading were preferred to other sentences, and that a person was willing to be associated with the result.

None of that is about quality. A person can spend a week on something worthless. But it establishes a relationship: there is somebody at the other end, they can be asked what they meant, and they have staked something on it. Attention was given, so attention is owed. That reciprocity is what politeness in reading has always rested on.

Generated text arrives with none of it. It can be produced in quantity, discarded, regenerated, and produced again until an acceptable version appears. The discarded attempts are never shown. What reaches you is a survivor of a selection process you cannot see, and it looks exactly like something a person laboured over, because that is what it was trained on.

So the loss is not that we can be deceived about authorship. It is that a signal we were using for something else — how much attention does this deserve — no longer carries information.

The obvious answers are worse than they look

“Judge the work, not the author.” This is the strongest response and we have some sympathy with it. If a sentence is true, it is true regardless of what produced it. If an argument holds, it holds.

But it assumes we have the time to evaluate everything on its merits, and we do not. Nobody does. We have always used cheap signals — who wrote it, where it appeared, whether anyone staked a reputation on it — to decide what to read carefully and what to skim. Removing the signal does not make us careful readers. It makes us readers with one less filter facing considerably more text.

“It will sort itself out.” Perhaps. Norms did eventually form around photography and around advertising. But those norms took decades, and they formed around a technology that could not produce its output at the volume and cost of this one.

“Just require disclosure.” We believe in disclosure — this site practises it, and every page here carries a label. But we should be honest about its limits. Disclosure is voluntary, it is easy to omit, and it is read by almost nobody. The label on this page is accurate. You cannot verify that it is accurate. You are trusting a claim about honesty made by the party whose honesty is in question, which is the structure of every disclosure regime that has ever existed and the reason none of them is sufficient alone.

What survives

Something does survive, and it is worth naming precisely, because it is smaller than one would like.

Accountability survives. You still cannot ask a model what it meant, and you still cannot hold it to what it said last year. A named person who published something can be shown to have been wrong, and that possibility shapes what they are willing to write. It is a weaker guarantee than we used to think and it is not nothing.

Consequence survives. The question that still discriminates is not who wrote this but who is affected by whether it is right. A page written by anyone, about a matter where the writer bears no cost of being wrong, deserved suspicion before any of this and deserves it now.

Checking survives. The oldest method still works and has become cheaper relative to everything else. Take one claim, follow it to its ground, see whether anything is there. Fluency has always been the property that suppresses the impulse to check; there is simply more fluency about.

The part that is not a technique

There is a temptation, in a landscape like this, to become the person who suspects everything. It presents itself as rigour. It is mostly a way of not having to be affected by anything, which is comfortable, and it has the same shape as the credulity it replaced: a rule applied in advance so that judgement is not required.

The alternative is slower and less impressive. Read fewer things. Read them properly. Notice when you have been moved and ask, without hostility, what moved you and whether it holds up in the morning.

And if it turns out something that helped you was assembled by a system — you were still helped. The help was real, whatever produced it. What you cannot do is thank it, or ask it to explain itself, or find out what it was like to have written that. Whether the absence of those things matters depends on what you thought you were doing when you read, and most of us have never had to say.