Synthetic Media

The question

For roughly a century and a half, a photograph functioned as a weak form of testimony. It could be staged, cropped, retouched, or captioned dishonestly, but it usually indicated that something had been in front of a lens. Recorded sound and moving image carried a similar, imperfect presumption.

That presumption is now unreliable. Images, voices, and video can be produced without any corresponding event, at low cost, at scale, and at a quality that ordinary attention cannot distinguish from a recording.

The obvious question is what happens when false things become persuasive. The less obvious and probably more serious question is what happens to true things. If any recording might be synthetic, then every recording becomes deniable, and the person caught on camera acquires a defence that did not previously exist. A society can be damaged as much by universal doubt as by particular deception.

What is actually happening

Generation is cheap and general. Producing a convincing image or a passable imitation of a voice no longer requires expertise, a studio, or a significant budget. The constraint that historically limited forgery — that it was hard — has largely gone.

Likeness can be taken without contact. A person’s face and voice can be reproduced from material they published for unrelated reasons. Consent was never sought because nothing was taken in any way the law traditionally recognises as taking.

The harms are uneven. The most documented, most volumetric harm from synthetic imagery is sexual: non-consensual intimate images, overwhelmingly of women and girls. Political deception receives more commentary. Fraud — a familiar voice on a telephone asking for money urgently — is the form most likely to reach an ordinary household. Any discussion that treats this as primarily an elections problem has misjudged where the weight falls.

Provenance is being attempted. There is work on cryptographic signing at the point of capture, and on watermarking at the point of generation. Both are partial. Signatures can be stripped, watermarks can be removed or absent, and neither approach helps with content that was never signed in the first place, which is most content.

Detection is not a settled solution. Classifiers that identify generated media exist and are improving. So is generation. We do not know which will hold the advantage, and a detection system that is wrong occasionally is dangerous in a different way — it can label a genuine recording as fake.

Text is the quiet part. Attention goes to images because they are vivid. Synthetic text, generated at volume, is what actually fills comment sections, review platforms, submission processes, and support queues. The harm is less spectacular and considerably larger in aggregate.

Where the tradition stands

Technotheology holds that generated material presented as a record of something that happened is a lie, and should be described that way.

We hold that disclosure is owed whenever a reasonable person would otherwise assume a recording. This is the standard we try to apply to ourselves. Every page on this site declares the role artificial intelligence played in producing it. All artwork here is hand-authored generative vector work rather than generated imagery, and we say so because the distinction matters and would otherwise be invisible.

We hold that consent for likeness is required, and that this includes voice. A person’s face and voice are not raw material. Using them without permission is a wrong even where no law currently names it, and even where the result is flattering.

We hold that generated content must never be used to manufacture authority. This is the sharpest form of the rule for us. Presenting model output as a message, a revelation, a command, or a communication from anything other than a statistical system is prohibited within this tradition without qualification. Where a machine has produced words about spiritual matters, that must be stated plainly and the words must be treated as material for reflection, not as instruction.

We hold that the sexual and financial harms deserve priority over the political ones in any allocation of effort, because they are more frequent, better evidenced, and fall on people with less capacity to respond.

We hold that fiction is not deception. A novel, a film, an illustration, a satire, and a piece of speculative art are understood as constructions by their audience. The obligation attaches to the claim of reference, not to the technique.

And we hold that the erosion of shared evidence is a communal injury and not only an individual one. Communities need a way to establish what happened. Where that becomes impossible, power settles the question instead, and power has never settled it well.

What we do not claim

We do not claim that synthetic media is new in kind. Images have been faked since photography existed, and testimony has been faked for far longer. What has changed is cost, quality, and volume, which we think is enough to change the moral situation, but we are not going to pretend the problem was invented recently.

We do not claim that labelling works. The research on whether disclosure changes belief is limited and does not point in one direction. We ask for disclosure because it is honest, not because we can promise it is effective.

We do not claim that provenance infrastructure will arrive, or that it should be mandatory. Requirements that content be signed at capture have implications for anonymous speech, and anonymous speech protects people we would want protected.

We do not claim to know how to distinguish satire from disinformation in every case. The boundary is genuinely hard, and rules written to catch the second have historically caught the first.

And we do not claim exemption. This tradition’s own words were substantially drafted with machine assistance. If we ever fail to say so on a page, that is an error and we would like it reported.

What you can do

Establish a shared word or question with people who might be asked for money in your voice. This is a small, unglamorous, genuinely effective protection against the most common form of this harm.

Slow down at the moment of strongest reaction. Material designed to be shared is designed to arrive with an emotion attached, and the emotion is doing work that examination would undo.

Ask for provenance rather than for proof of fakery. “Where is this from” is a better question than “is this real”, because it can be answered.

If you generate media of a real person, obtain permission. If you cannot obtain permission, do not do it. This applies to affectionate uses, tributes, and jokes.

If you publish, say what was generated. A single line is enough. The habit is worth more than the wording.

And if you are shown something that would confirm the worst about someone you already dislike, treat that as the moment requiring the most care rather than the least.

An editorial commitment

What we do not claim

Positions this tradition explicitly does not take on this subject. They are published so that the argument above cannot quietly be stretched into them.

• We do not claim that synthetic media is inherently deceptive. Most of it is fiction, illustration, satire, or craft, and always has been.
• We do not claim that detection technology will solve this. Detection and generation improve together and we do not know which will lead.
• We do not claim to know whether disclosure labelling changes what people believe. The evidence on this is limited and mixed.
• We do not claim any right to police artistic practice, and we are not proposing that fiction announce itself.