AI Governance // Truth Forensics 03
When AI Cannot Determine the Truth
Desmond Otieno Odhiambo · ClearGlass Intelligence · Burlington, Ontario · 30 September 2026
An analyzer can find a two-second gap in a video's frame timing. It cannot tell you whether those frames were dropped by an overloaded encoder or cut by a person. Any system that pretends otherwise is not being careful. It is being confident.
What automated analysis is good at
Machines are excellent at the tedious, exact parts of media examination:
- reading every metadata field and noticing that a modification time is two days after capture;
- walking a container's timing table and finding the one interval that is fifty times the others;
- finding runs of exact-zero samples in audio, which live microphones almost never produce;
- matching thousands of image blocks to find a region copied from elsewhere in the frame;
- checking that a file's hash still matches the one recorded when it was received;
- comparing what ten sources say about time and place, without getting tired.
None of that is intelligence in the interesting sense. It is measurement. It is also the part people do badly.
What it cannot know
| Signal | What it shows | What it cannot determine |
|---|---|---|
| Frame-timing gap | Intervals are missing from the timeline | Encoder drop, pause, or deletion |
| Exact digital silence | Samples were zeroed | Noise gate, codec, mute button, or edit |
| Abrupt waveform step | A sharp discontinuity | Knock, plosive, cable click, or splice |
| Cloned region | Two areas are identical | Which is the original; whether the pattern is genuinely repetitive |
| Editor in the software tag | A program wrote the file | What, if anything, it changed |
| No indicators at all | The available tests found nothing | That the content is authentic |
The last row is the one vendors skip. The absence of indicators is a statement about the tests, not about the event. The honest sentence is: no manipulation indicators detected by the available analyzers; authenticity cannot be established solely from this analysis.
Language is a control
“This video is fake” is a conclusion about the world. “Manipulation indicators detected; human review required” is a statement about analysis. Reports should be written so that the second is the only thing the machine is allowed to say.
The ClearGlass report generator enforces it. Engine-authored text passes a guard that refuses certainty phrasing (“is fake”, “proves authentic”, “100% accurate”, “guaranteed”). Quoted user text, such as the claim under test or a reviewer's note, is exempt, because a report must be able to quote a claim it does not endorse. Every line of the report is typed as an observation, an interpretation or a conclusion.
Governance that holds up
- Record every analyzer run: which analyzer and version, the input hash, the output hash, when, and for whom. If a model provider is involved, record the model and prompt version too.
- Keep external models opt-in. Evidence is often sensitive. The ClearGlass console calls no AI service at all; an adapter would have to be switched on deliberately, and every call would be logged the same way.
- Separate duties. The analyst who produced a finding cannot accept or reject it. A reviewer can request a second opinion, and the requester cannot then decide alone.
- Append decisions; never edit findings. A reviewer rejecting an item excludes it from the claim assessment, and the report says who did so, when, and why.
- Keep demonstrations labelled. Synthetic data stays SIMULATED whatever a reviewer does, so a demo can never be mistaken for a case.
A buyer's checklist
Before relying on any media-authenticity tool, ask:
- What does it say when it finds nothing? If the answer is “authentic”, walk away.
- Does it count copies and reposts as independent confirmation?
- Does every finding state its method, its limitation and another explanation?
- Can it show the hash of the file it analysed, and the hash of its output?
- Where does the evidence go? Who else can see it?
- Is a human decision required before anything is called verified, and is that decision recorded?
See how each of these is handled in the Truth Forensics console, on synthetic data or on your own files, without uploading anything. Companion pieces: Truth Is an Evidence Graph and Why Deepfake Detection Needs Chain of Custody.