The File and the Event

A perfectly genuine photograph can be used to tell a lie.

Take Tess’s business picture. She took it outside a house after finishing a job. The image shows her standing beside a tool bag. It does not show her making a statement about customers. Put an invented quotation beneath it and nothing inside the photograph has to change.

This is why a conversation about authenticity can become confusing. We may be asking whether a file has changed since it was signed, whether a camera captured the scene, whether the caption identifies the right person, or whether the image supports a claim about that person’s conduct. Those questions overlap. They ask for different kinds of evidence.

A useful answer begins by naming the question. If the dispute concerns the words attached to the picture, showing that the picture is genuine does not settle it. If the dispute concerns the speaker’s identity, checking the spelling in the transcript does not settle it. The error can survive a technically correct answer to a question nobody needed answered.

Four things to keep apart

Imagine a video of a person leaving a building. A caption says that she fled a public hearing after admitting misconduct. We can ask about the file, the identity, the setting and the claim.

The file question concerns the material we have: its origin, known changes, missing sections and relationship to other versions. The identity question asks who is shown. The setting question asks when and where this took place, and what happened around it. The claim question asks whether leaving the building supports the story told in the caption.

Suppose we establish that the clip was recorded at the stated building on the stated day and has not been changed since a particular copy was made. The person may still be misidentified. Suppose the identity is correct. She may have left before the hearing began. Suppose she left during it. The caption’s account of why she left may still be invented.

Each answer narrows the uncertainty. None should quietly answer the next question too.

The same care applies in defense. A person might show that a clip has been shortened and conclude that the allegation is false. Shortening alone does not establish that. Most excerpts omit something. The important question is whether the omission changes the meaning relevant to the claim. A full recording may support the criticism, weaken it or reveal that the excerpt was misleading.

This approach can feel slow beside a confident verdict. It is faster than days of argument about the wrong issue. An investigator, a publisher or an ordinary reader can write the disputed statement in one line, then ask what evidence would make that statement more or less credible. That keeps technology in service of a question rather than turning it into a show of authority.

What a credential can tell us

The Coalition for Content Provenance and Authenticity, or C2PA, offers one approach to recording a file’s history. Its May 2025 version 2.2 specification describes tamper-evident provenance information and validation. The companion explainer makes an important limit clear: provenance does not by itself establish factual truth. A record of a file’s history can be incomplete, and metadata can be removed. Missing credentials therefore do not prove that media is fake. These documents describe a technical system, not a field experiment showing that it resolves every dispute. ¹

There is a simple way to see why that limit matters. Imagine a camera faithfully recording a rehearsal. The scene contains real people and real speech. The file’s history may be well documented. A later caption calls the scene an unplanned encounter. The false claim lies in how the genuine material is described.

Or imagine a genuine photograph of Tess placed beside somebody else’s genuine words. We could know the history of both items and still need to ask why they were joined. Information about each ingredient would help. It would not automatically justify the new sentence created by putting them together.

Use the record for the question it can answer. Continue checking the caption and the attribution.

Think about an old family photograph. Its owner may have no signed digital record of where it came from. That absence does not turn the people in it into invented people. We would look for other evidence if a consequential question arose: who kept it, whether another copy exists, what can be established about the setting. The inquiry would fit the claim.

The opposite mistake is also possible. A badge or certificate can encourage a reader to stop asking about the caption. The symbol seems to cover the whole story because it sits near the whole story. A good interface should explain what has been checked, in language a person can use. “This file’s recorded history validates” and “this allegation is true” should never be allowed to blend into the same reassuring mark.

That is a design recommendation, and a demanding one. It asks for a label precise enough to remain honest when the surrounding account is wrong.

A score needs a question

Another friend sends Tess a website that promises to detect generated images. She uploads nothing. First she asks what she would learn from the result.

If the website declares the picture human-made, the false sentence remains. If it declares the picture artificial, the tool may be wrong about a photograph Tess knows she took. Either result could distract from the unsupported attribution. A test should be chosen because it bears on the disputed claim, not because it is available and produces a number.

NIST’s November 2024 technical overview, AI 100-4, describes limits across approaches to synthetic-content transparency. Detection performance depends on context, and false positives can harm a person’s reputation. The report does not certify a detector as a general judge of authenticity. Its assessment supports treating a result as evidence with limits, rather than a verdict about a person. The report itself is a technical synthesis; it does not supply a success rate for the particular file in front of us. ²

A percentage can make uncertainty look settled. But a number printed beside “AI” needs an explanation. Does it describe the system’s confidence, a measured error rate on a named test set, or something else? Which kinds of material were tested? Does this file resemble them? If the answer is unavailable, the number should not be promoted into certainty merely because it has a decimal point.

For a consequential accusation, I would want the person relying on the test to explain the possible errors in ordinary terms. Could genuine material be labeled artificial? Could artificial material pass as genuine? What independent evidence bears on the same claim? Those questions are fair whether the result favors the person we like or the person we distrust.

We should be especially wary of turning a technical suspicion into a moral diagnosis. A detector result cannot establish why someone posted a file, what they knew, or whether they intended harm. Even a demonstrated alteration may have an innocent purpose: privacy, compression, translation, accessibility or a clearly marked illustration. The purpose and use have to be examined rather than read off the existence of an edit.

That does not make all editing harmless. It means the judgment should concern the actual misleading act. If a cropped frame hides the person who spoke, explain that. If audio has been joined so that the words mean something different, describe the change that can be established. Precision makes an accusation more accountable and gives its subject a fairer chance to answer.

The words around the words

A transcript is a useful aid. It can also conceal a decision.

Someone has chosen how to mark a pause, identify a speaker, or represent an uncertain word. In a translated transcript, somebody has chosen among possible meanings. These choices may be careful and defensible. They still deserve attention when a person’s reputation depends on a sentence.

Suppose a speaker reads a bad proposal aloud, then explains why she opposes it. A transcript that ends before her explanation can assign the proposal to her. Every visible word may be accurate. The account is wrong because the act of quoting has been lost.

Or the transcript turns a question into a statement by removing the punctuation that carries its shape. The recording may settle that point, or it may remain hard to hear. In either case, the uncertainty should travel with the transcript. A reader should not encounter a polished sentence where the source contains a doubtful sound.

These examples do not mean that full context is endless or that no excerpt can be fair. An editor must choose. The duty is to include what changes the reasonable reading of the quoted material. No one needs the entire day to understand a two-minute exchange. They may need the sentence just before it.

A person contesting an excerpt should also be specific. “You took it out of context” is the start of an explanation. Which missing passage changes the meaning, and how? If the missing material merely repeats the same point, the objection may not help. If it reveals a rehearsal, quotation or different speaker, it may be decisive.

The reader is entitled to this clarity from both sides. Context should be something that can be shown, not a word used to make evidence vanish.

Do not demand a perfect past

There is a temptation to answer synthetic media by requiring everyone to maintain a fully documented record of their life. Record every conversation. Sign every image. Keep every message. Then, perhaps, no one can invent you.

Consider what that would ask of an ordinary person. Privacy would become a weakness to be defended. A lost phone could become a gap in innocence. The person without the right tools would begin every dispute at a disadvantage. Someone accused of an invented conversation would have to produce a record of every hour in which it might have happened.

I do not think that is a fair bargain. Better records can be valuable, especially for institutions making public claims. They should not become a demand that private people live under constant recording to deserve belief.

The burden should remain attached to the claim. If an account says Tess made a statement in an interview, it should identify the interview or explain its evidence. Tess may provide information that helps resolve the question. Her inability to produce a lifelong archive does not fill the gap in the accusation.

This protects more than the falsely accused. A genuine witness may also lack a recording. We should listen carefully, ask relevant questions and assess what can be established. An absolute rule that only recorded events count would erase real experiences. An absolute rule that an asserted event counts as proved would erase the need to examine evidence. Fair judgment requires more work than either shortcut.

Technology can make part of that work easier. It can preserve a useful history, reveal an inconsistency or help find an earlier version. Its value grows when we describe exactly what it has done.

Tess closes the detector page. The next message she sends asks for the interview that the image appears to quote. It is a small request. It bears on the actual claim. Whatever tools become available next year, that habit will still deserve a place beside them.

Source notes