Every Sentence Needs a Return Address
For most of human history, a piece of writing carried its origin with it. A letter had a sender, and the handwriting on the envelope told you who before you opened it. A newspaper had a masthead, and the masthead had a reputation it could lose. Even a rumor had a shape: a cousin who heard it from a neighbor who heard it at the market. You could weigh a claim partly by weighing who stood behind it, and when the claim turned out to be wrong, you knew whose door to knock on.
That habit of reading was so old and so automatic that most of us never noticed we were doing it. We are about to notice, because the thing it depended on is quietly disappearing.
Smooth, confident and orphaned
A generated paragraph has no one standing behind it. That is not a criticism of the systems that write it. It is a description of how they work. A language model draws on an enormous body of text and produces something new that belongs to no single part of it. The result comes from everywhere, which in practice means it comes from nowhere in particular. It arrives fluent and well organized and confident, and it has no return address.
Then it travels. A summary is drafted by a model, edited by a person, pasted into a report, summarized again by another model for someone too busy to read the report, quoted in a slide, and repeated in a meeting. At every step the sentence gets a little smoother and the trail back to whatever was originally true gets a little fainter. By the end, nobody in the meeting could tell you where the claim began, and the claim sounds more authoritative than it did at the start, because it has been polished by every hand it passed through.
This is new. Not the error, which is as old as writing, but the scale and the smoothness. Bad information used to look bad more often than not. It had typos, odd phrasing, the fingerprints of a person in a hurry. Now the error and the truth wear the same clean clothes.
What payments already learned
I spent a large part of my working life in payments, and payments solved a version of this problem a long time ago, because it had to. Every dollar that moves through a modern payment system carries a trail. Where it started. Which institution touched it, and when. What it was for. Who authorized it. When something goes wrong, you can follow the money back, step by step, to the point where it went wrong.
Nobody in that world trusts money because it looks like money. A transfer that looks perfect and cannot be traced is not reassuring. It is alarming. The look of a thing tells you almost nothing. The trail tells you nearly everything. A payment with no origin is not treated as a payment. It is treated as a problem, held until someone can account for it.
Text is moving in the opposite direction. We are producing more of it than ever, with less provenance than ever, and we are still reading it with the old instincts, which judge by surface. Fluent sounds trustworthy. Confident sounds informed. Those instincts were calibrated for a world where fluency took effort. That world is gone.
The scarce thing
When something becomes free, the value moves to whatever is still scarce next to it. Sentences are now nearly free. What is scarce is knowing where a sentence came from, and knowing that someone will answer for it.
I think this changes what good work looks like, in companies and in public life. A few habits seem to me worth building now, before the trail is lost for good.
Keep the source attached to the claim. Not in a folder, not in a footnote nobody follows, but next to the sentence, so that when the sentence travels, the source travels with it. If a claim cannot carry its origin, it should carry a warning that it has none.
Treat an orphaned fact the way a bank treats an orphaned deposit. Hold it. Do not build on it. Do not repeat it in the next meeting. Wait until someone can vouch for it, and if nobody can, let it go, however good it sounds.
Put a name on what you publish. Not the name of the tool. The name of a person who read it, checked it and will answer for it. I write with AI openly, and I sign everything I publish, because the signature is the part the machine cannot supply. It is a promise that a person stands behind the words and can be found.
Reward the people who say "I do not know where that came from." In most rooms that sentence sounds like weakness. It is about to become one of the most valuable things a colleague can say.
Who answers
It is worth being precise about what the machines are and are not doing here. Today's AI wants nothing, including the truth. It does not set out to deceive anyone, and it does not set out to vouch for anything either. It produces text that is likely, and likely is not the same as true. The job of standing behind a claim has not moved. It is still ours.
Nobody knows what self-improving systems will eventually be able to do, and I would not bet against their capability. But capability is not accountability. A system can become far more capable and still leave the question of who answers for a sentence exactly where it was, with the people who chose to publish it. More capable is never more human, and it is never more answerable.
The letter writers of earlier centuries had one advantage over us that we should try to win back. When they said something, everyone knew who had said it. The return address was on the envelope. It is worth the effort to keep writing it there.