When everything can be denied

The shape of the problem

The first-order harm of synthetic media is obvious: false things are believed.

The second-order harm is that true things can be disbelieved. Someone caught on camera can now say the recording was fabricated, and enough of the audience will accept it. This has a name in the academic literature — the liar's dividend — and it was described before the technology was good, precisely because the mechanism does not depend on the technology being good. It depends only on the audience knowing that fabrication is possible.

The dividend is paid to whoever benefits from doubt. It costs nothing, requires no technical capability, and cannot be countered by improving detection, because the claim is not that a specific analysis is wrong. It is that nothing can be established at all.

Documented and predictable consequences

Politicians and public figures have dismissed authentic recordings as AI-generated. Defendants have challenged audio and video evidence on the basis that fabrication is possible, and courts in several jurisdictions have begun to work out how to handle that argument.

Journalists report that sources are more reluctant to provide recordings, since a recording no longer settles anything. Fact-checking organisations report that debunking one item now prompts a general "you cannot trust anything" response rather than acceptance of the correction.

The corrosive part is not that people believe fakes. It is that the shared reference point disappears, and disputes that used to end with a recording now do not end.

Why this is harder than the first problem

Fabrication has technical countermeasures, however imperfect: provenance, watermarking, verification routines. The dividend has none, because it is a claim about epistemology rather than about a file.

Worse, the countermeasures partly feed it. Every article explaining how convincing generated video has become is also an article establishing that any video might be generated. This is a genuine dilemma for anyone doing public education on the subject, this course included, and there is no clean way out of it. The honest response is to pair every statement about capability with a statement about what can still be established — which is what the previous lesson is for.

What actually helps

Provenance at capture, again. It is the only mechanism that supports a positive claim of authenticity, which is precisely what the dividend attacks. A recording that can be shown to have come from a specific device at a specific time is much harder to wave away.

Institutional custody. Evidence handled through a documented chain — police, courts, established newsrooms — resists dismissal in a way an anonymous upload does not. This is not new technology; it is the reason chains of custody exist.

Multiple independent sources. Three recordings from three unconnected people are far harder to dismiss than one, and the argument required to do so becomes visibly absurd.

Naming the move. When someone dismisses evidence as fabricated without giving a reason, saying so plainly is worth something: "that is an assertion, not an analysis; what specifically indicates fabrication". Making the burden visible does not always work and it costs nothing.

Calibrated public language. Reporting that says what was established and how, rather than "experts say it is real", gives an audience something to hold. Vague authority is exactly what the dividend erodes.

The honest position

Nobody knows how this resolves. One view holds that societies absorbed photographic manipulation, staged photographs and edited audio before, and that new norms and institutions will form again — slower, more mediated, and functional. Another holds that the volume, speed and personalisation are different in kind, and that the shared factual ground genuinely does not reconstitute.

Both are held by serious people. The evidence available is a few years old and mostly about elections and public figures, which is a narrow slice of where the harm lands.

There is also a personal version of this that is easier to act on than the civic one. If you record something that might matter — a conversation with a landlord, an incident at work, evidence of a repair that was never done — the dividend applies to you. What protects that recording is the same set of things: capture it on a device that records provenance if you have one, keep the original file rather than a re-shared copy, note the time and place immediately, and tell someone contemporaneously. None of that is technical and all of it is what makes a recording hold up when somebody says it was fabricated.

What can be said with confidence is narrow: the dividend is real, it is being used, it does not require any fake to exist, and the only countermeasures that touch it are about establishing origin rather than about detecting fabrication. That is a reason to support provenance infrastructure even while being clear-eyed about how partial it currently is.