Chain of Custody in the Deepfake Era — Why OSINT Evidence Has to Prove Its Own Integrity

2026-07-16 · Philip Choo · AI9OS

For most of investigative history, the integrity of evidence rested on a simple, physical idea: a signed, dated log of who handled an item, when, and what they did with it. Break the chain — an unexplained gap, an unlogged handler — and the evidence is challengeable, sometimes fatally. That discipline was built for a world of physical exhibits: a seized phone, a printed document, a photograph in an envelope.

Open-source intelligence broke the envelope. The "exhibit" is now a screenshot of a profile that may be edited tomorrow, a cached page that may 404 next week, a data-broker record with no author and no date. And in 2026, every one of those artefacts can be fabricated convincingly by anyone with a consumer GPU. A deepfake image, a synthesised voice note, a forged "screenshot" of a message that was never sent — all indistinguishable from the real thing to the naked eye.

So the question a court, a regulator, or an opposing expert will ask is no longer just "is this true?" It is "how do you know this is the artefact you collected, and that nobody — including you — has altered it since?" In the deepfake era, a finding that cannot answer that question is not evidence. It is an assertion.

The three failures a modern chain of custody has to close

Open-source evidence fails integrity in three specific ways, and a serious system has to close all three:

1. Fabrication at source — the artefact was fake before you ever touched it. No custody system can make a forgery true, but it can fix the artefact in time, so that later analysis (and later admissions) attach to exactly what was collected, not a moving target.

2. Alteration after collection — the artefact was genuine when captured but has been edited since, whether by accident, by a later "correction", or by tampering. This is the classic chain-of-custody failure, and it is the one software can close completely.

3. Disappearance — the source changes or is deleted, and the only remaining record is the investigator's word that it once said what they claim. On the open web, this is the normal case, not the exception.

How a tamper-evident locker answers all three

The mechanism is old cryptography applied with discipline. Every item that enters the evidence locker is recorded with the tool that collected it, the source, the time, and a cryptographic hash — a fingerprint of its exact contents. Crucially, each new record's fingerprint is computed to include the previous record's fingerprint. The exhibits are chained.

That single design decision has a powerful consequence. To alter any past exhibit — change a finding, backdate an entry, quietly delete an inconvenient item — you would have to recompute not just that record's fingerprint but every fingerprint after it, because each one depends on the last. The chain makes tampering evident: any break shows exactly where it happened. On our platform the underlying store is append-only by construction — attempts to update or delete a recorded exhibit are rejected outright, so corrections are added as new, superseding records and the original is never silently overwritten. The history is immutable; only the interpretation moves forward.

Against the three failures:

• Alteration after collection is closed absolutely. Re-running the verification re-computes every fingerprint from the case data itself; if a single byte changed, the chain reports the exact exhibit where integrity broke. An independent expert can run the same check without trusting the investigator at all.
• Disappearance is answered by fixing the artefact in time at capture — hash-stamped and time-stamped the moment it enters the locker, ideally alongside an independent archival snapshot — so the record of what the source said survives the source itself.
• Fabrication at source is the one no locker can defeat by itself. That is exactly why integrity of custody is necessary but not sufficient — and why it has to sit underneath a discipline of judgment, not replace it.

Integrity is the floor, not the ceiling

Here is the trap: a beautiful, tamper-evident chain of custody can perfectly preserve a fabricated artefact. It proves the artefact hasn't changed since you collected it. It does not prove the artefact was real when you collected it. Confusing the two is how a rigorous-looking process launders a forgery.

That is why, in our practice, the tamper-evident locker is the floor. On top of it sits the harder discipline: every finding that reaches the highest confidence tier must first survive a structured challenge — the case against it written down, its sources checked for genuine independence, its fabrication risk explicitly assessed — before a human investigator signs it off. The chain of custody proves the exhibit is unaltered. The adjudication proves it was interrogated. The court gets both, bound together in one verifiable pack.

What the deliverable looks like

The output of all this is not a PDF of screenshots. It is a provenance pack: a signed manifest with one entry per exhibit — its chain links, the tool and source, the timestamps, the confidence rating and the adjudication behind it — topped by a full chain-verification result an independent party can reproduce. It opens with the lawful-collection record and the enforced scope of the engagement, and closes with a hash chain anyone can re-check.

In a world where anything can be faked, the persuasive thing is no longer the artefact. It is the artefact plus its provable, unbroken history. The chain of custody was always the quiet backbone of good evidence. In the deepfake era, it moved to the front of the room.

General information for practitioners, not legal advice. AI9OS is an open-source-intelligence technology platform; investigation services are conducted solely by licensed agencies under Singapore's Private Security Industry Act.