The second audience¶

Where a machine produced something meant for people, a person met it somewhere in the loop. Somebody read the text, looked at the picture, ran the code, and decided whether it was any good. That step was rarely written into the design, and it was doing more work than the design ever credited it with. It was the point at which the world got to disagree.

There is now a second audience in the loop. A generated page is read by a crawler before it is read by anyone, if it is read by anyone at all. Generated code is checked by a generated test. An answer is scored by a model built to score answers. Where the output lands has changed, and the check that used to sit at the landing point has not moved with it.

Who the output is for¶

Sorting by who receives the output is more useful than sorting by capability.

A machine acting on material people made, with a person at the far end: the compiler, the spreadsheet, the search index. Old, unremarkable, and still most of what runs anywhere.

A machine producing material people consume: the first generative wave, and the one the public argument is still largely about. The essays, the pictures, the cloned voices, and the question of whether any of it is any good.

A machine producing material other machines consume: generated pages entering the next training set, generated tests exercising generated code, model answers scored by model judges.

A machine producing material that determines how later machines are built: architectures, training pipelines, evaluation harnesses, and the datasets themselves.

The public argument draws its line between the first two positions and asks whether machine-made things are as good as handmade ones. The line that changes the question sits between the second and the third, where output stops being something a reader judges and becomes part of what trains and evaluates the next system. That line goes largely undiscussed, because nothing about it is visible to a reader.

Made without meaning to¶

Substrate here means the material a system is built out of: text, images and code, and also the benchmarks, the tooling, the datasets and the accumulated write-up of what has already been tried. Every generation of these systems has been assembled out of a substrate that people made over decades, mostly without meaning to, by writing things down in public.

The important transition is not from human to machine authorship. It is from a human-mediated substrate to one increasingly produced by the same kind of system it will be used to build.

Friction was the filter¶

A human-mediated substrate came with costs that behaved like quality control without ever being described as such. An article cost somebody a day. Publishing cost money and a name that could be complained to. Code cost a salary, and the salary bought an argument in review. None of this was efficient and none of it was reliable, which is why the costs were treated as friction to be removed rather than as a filter to be replaced.

When the cost of production falls far enough the filter goes with it, and little in the arrangement looks for a replacement, because nothing had recorded that a filter was present. The consequence is not mainly volume. It is that the next system’s training set, benchmark suite and reference implementations are drawn from a pool whose selection pressure has changed, and which carries no record of the change.

Generation keeps no receipt¶

A copied text carries traces of where it came from: a phrasing, a citation, a file date, sometimes a URL. A generated text carries none of that in itself. Its origin can be recorded, which model, which prompt, which documents were in front of it, and whether that record exists is a decision of the generating system. Nothing in the artefact carries it. Provenance here does not degrade in transit, because it was never attached to the thing that travels.

Platforms have begun adding provenance back by hand, as labels and disclosure requirements. A label recorded at the point of upload is a reasonable thing to ask for, and it is not a provenance system. It depends on the candour of the party with the least interest in disclosing, and it stops at the platform’s edge. Material that never passes a platform, which includes most of what a crawler collects, carries nothing at all.

The model is the part everyone looks at. The part that decides what comes next is the material it leaves lying about.

The clerk’s brief¶

2025: The audience stopped being human¶

Imperva’s bad bot report recorded automated traffic at 51 per cent of all web traffic in 2024, the first time in a decade that it surpassed human activity, with bad bots at 37 per cent. Traffic composition is not readership, and a crawler is not an audience in any sense a publisher would recognise. The clerks note only that this is the closest count available, and that most requests on the web now come from something that cannot be persuaded, bored or annoyed by anything it is sent.

March and May 2024: Provenance added back by hand¶

On 18 March 2024 YouTube began requiring creators to disclose realistic altered or synthetic content, meaning “content a viewer could easily mistake for a real person, place, scene, or event”. Meta said it would start labelling organic AI-generated content in May 2024, following an Oversight Board finding that removal risked restricting expression unnecessarily, and applies its labels where it detects industry-standard indicators or where a user self-discloses. The clerks observe that a record which begins at upload can say nothing about how the material was made, and that material which never passes a platform sits outside the scheme altogether.

The hardest part to put back¶

The substrate was made by people who were not thinking about substrate. It is now being extended by systems whose output is cheap, unlabelled at production, and not reliably separable by whoever collects it next. The clerks’ standing assessment is that the change of audience may prove the hardest part to reverse, because the check that used to sit where the output landed was never written down as a component, and components nobody wrote down do not get reinstated.