Human-Made Art in the Age of AI

For most of the history of art, making something took time, and that time was a kind of proof. Nobody had to ask whether a painting was painted. The evidence was the object.

That has stopped being true for images, and it is on its way to stopping for music, video and prose. Generating a convincing picture now costs a few seconds and almost nothing. The supply of plausible work has become effectively unlimited, and the old assumption underneath it, that someone sat down and made this, no longer arrives attached to the file.

The usual response is to ask how we detect the difference. That turns out to be the wrong question.

Detection does not work, and is not going to

Every automated AI detector shares the same structural weakness: it is a classifier trying to identify the output of systems that are explicitly trained to be indistinguishable from the thing it is looking for. Improving the generator degrades the detector, automatically, as a side effect. It is a race with a built-in result.

The consequences of pretending otherwise are not abstract. Text detectors have repeatedly flagged human writing as machine-generated, and the people most often harmed are the ones least able to argue back: students, non-native speakers, anyone whose style is unusual. A tool that is wrong ten percent of the time in a field of millions of works is not a safeguard. It is a machine for producing false accusations.

Any platform promising reliable detection of AI-generated art is either misunderstanding the problem or selling something. The honest position is that you cannot look at a finished image and know how it was made.

Provenance is the question that can be answered

What you can know is where a file came from, if the information travelled with it.

This is the idea behind C2PA, an open standard developed by a group including Adobe, the BBC, Microsoft and Sony, and its consumer-facing form, Content Credentials. Rather than inspecting the pixels, these attach a signed, tamper-evident record to the file: what device or software produced it, what edits were applied, whether generative tools were involved.

The shift matters. Detection asks a question about the artefact and guesses. Provenance asks a question about the history and reports what it was told. One is a probability, the other is a claim with a signature on it.

It is not a complete answer. Metadata can be stripped, screenshots discard it, and a signed record only tells you what the signer asserted. A camera that says a photograph is a photograph is evidence, not proof. But a claim someone has put their name to is a different kind of object from a statistical guess, and it fails in ways you can inspect.

The law has already drawn a line

Copyright offices have moved faster than most platforms. The US Copyright Office has examined the copyrightability of AI-generated material since 2023 and holds that copyright protects human authorship. Material generated by a machine without human creative control is not registrable, while work where a person contributed the creative expression can be.

For artists this is more than a technicality. It means the question "who made this, and how" is already load-bearing in law. Whatever platforms decide to do, an artist will increasingly be asked to answer it, and the ones who can answer clearly will have an easier time than the ones who cannot.

What this changes for artists

The practical effect is that context stops being decoration around the work and becomes part of what the work is worth.

A finished image on its own is now the least informative version of itself. The same image accompanied by the studio it was made in, the series it belongs to, the exhibition it hung in, the person who made it and what else they have made, is a different proposition. Not because the picture improved, but because the surrounding evidence is the part that cannot be generated in four seconds.

This is an unwelcome shift for anyone whose practice is private by temperament, and it is worth naming that cost rather than pretending it away. Documenting a practice is work, and it is not the work most artists want to be doing. But the alternative is to compete on output alone in a market where output has become free.

There is a quieter consequence too. Attribution has always been unevenly distributed, and the artists whose names were already easiest to lose, the ones without galleries or institutional backing, are the most exposed when provenance becomes the currency. A system that only works for artists who already have representation would repeat the problem it was meant to solve.

Where we stand

Artistivo's position is that human artistic practice belongs at the centre of the platform, and that the useful contribution is not to adjudicate what is real.

Some of this is already running. An artist page ties works to a named person with a location and a discipline, events and classes record what an artist actually did and where, and Artist Visits film people in their studios, which is provenance of a low-tech and fairly persuasive kind.

The rest is intention, and should be read as such. We expect to combine artist declarations with provenance signals like Content Credentials where they exist, and to treat both as claims that can be inspected rather than verdicts. We do not intend to deploy an AI detector and present its output as fact, because we do not believe such a thing works.

The distinction we care about is narrow and worth stating plainly. Technology can support artists without making them invisible. A generated image has no practice behind it, no location, no exhibition history and nobody to ask. As that kind of material becomes the majority of what exists online, the link between an artist, their work and its history stops being background information and becomes the thing itself.