Seeing is a starting point.
Verification is the work.

THE AUTHENTICITY QUESTION

What exactly are you trying to establish?

When a disputed video arrives, “is it real?” compresses several questions. Was the file manipulated? Does it show the event claimed? Was it recorded at the stated time and place? Is the speaker the claimed person? Does that person have authority for the requested action?

A technically authentic clip can be old, cropped, or attached to an invented caption. A real employee can appear in a meeting without having approved a payment. Conversely, visible compression artefacts do not, by themselves, establish fabrication. The business needs to identify the proposition on which its decision depends.

CORROBORATION / THE ORIGINAL BACKSTORY

Look beyond a single suspicious-looking clip.

Disputed imagery surrounding the March 2026 Minab school attack illustrates the need for corroboration. Amnesty International’s account described analysis of videos, photographs, satellite imagery, and interviews. It also stated limits on independently verifying reported casualties. The methodological lesson is the combination of different evidence, rather than confidence based on one visual cue.

An investigation should state what each source contributes and where uncertainty remains. Multiple reposts of the same recording are not independent corroboration. A witness, a timestamp, a physical feature, and an independently obtained record may answer different parts of the question.

This matters beyond public investigations. In a business dispute, preserve the source email, relevant transaction record, and known approval history alongside a screenshot or recording. Ask for the best available original material without assuming that the person who raises a concern must become a forensic expert.

WHAT PEOPLE CAN PERCEIVE

More attention helps in some conditions, but it is not a guarantee.

Nightingale and Farid’s 2022 face study found that the selected synthetic faces were difficult for participants to distinguish from real faces and were judged more trustworthy in the study. That finding concerns the study’s materials and tasks; it is not a universal accuracy rate for every face, model, or viewing situation.

The 2025 voice-cloning study likewise found limitations in people’s ability to recognise AI-generated speech. Familiarity and confidence should not be treated as proof that the voice belongs to the authorised speaker. A familiar voice is one signal, not an approval control.

The 2026 video experiment supports giving people time to examine evidence, while showing why confidence should remain bounded. Different media, generators, audiences, and viewing conditions can change performance. The bar lengths share a percentage scale; they do not claim a continuous learning curve between the two viewing conditions.

THE LIAR’S DIVIDEND

A fake can mislead; the possibility of a fake can also conceal.

If an authentic recording can be dismissed with “AI could have made that”, the mere availability of synthetic media creates room for denial. This is the liar’s dividend: uncertainty can benefit someone challenged by real evidence, even when no fabrication occurred.

The right response is not to accept all recordings or reject them all. Preserve the evidence, identify the precise challenge, and seek proportionate corroboration. A bare allegation of fabrication should not automatically erase a well-supported account; a confident assertion of authenticity should not make verification unnecessary.

The debate has reached US evidence rulemaking. The October 2026 Advisory Committee agenda materials discuss a draft Rule 901(c), including a proposed response to a supported challenge that evidence was generated by AI. This is a proposal under discussion, not an enacted rule and not a UK evidential standard.

For a company, the operational point is to prepare records before a dispute. Keep approval history, source documents, access logs, and a clear account of how important evidence was obtained. A future detector cannot reconstruct records that were never retained.

BUSINESS IMPERSONATION

Persuasive media becomes dangerous when it carries authority.

Hong Kong’s June 2024 government account described a fraud involving a pre-recorded conference and transfers of about HK$200 million. This should not be retold as proof that every participant was a live, interactive avatar. The important failure was the path from a convincing representation to consequential payment.

The FBI’s IT-worker warning also describes identity deception in hiring, including AI-enabled interview methods. Its relevance is the need to connect identity evidence with the person doing the work and receiving access; appearance, accent, or nationality is not a reliable risk judgement.

A familiar face on screen can help a fraudster borrow hierarchy, urgency, and social confidence. But the attack still needs the organisation to do something: send money, change a destination, reset access, hire a person, publish a statement, or disclose information. That is where a workflow can interrupt the harm.

DETECTION & PROVENANCE

Useful signals have different limits.

NIST’s review of synthetic-content approaches considers complementary techniques such as detection, provenance, and watermarking. Performance depends on the material and conditions. A detector result should be interpreted with its tested scope and uncertainty, rather than used as a binary finding of truth.

C2PA provenance can provide signed information about a file’s origin and changes. Valid credentials do not establish that the depicted event or accompanying claim is true. Missing credentials do not establish that a file is fake. Adoption, transformations, and the chain through which content reaches you all matter.

Nor is a failed search for manipulation proof that manipulation did not occur. There may be too little information, an unfamiliar technique, or loss of detail through reposting. A useful review records “not established” where the evidence does not support a stronger conclusion.

Detector performance and the spread of social-media claims can change rapidly. An aggregate failure percentage or a dramatic market-loss estimate without a clear method can obscure the point. State what was measured, over which period and material, before using a number to justify a decision.

A WORKFLOW THAT CAN SAY “PAUSE”

Verify the decision through a route the suspicious content does not control.

The independent route matters. Calling a number supplied in the suspicious message, or asking the same unverified conference to confirm itself, can reproduce the deception. Keep verified contact details available when the normal platform is unavailable.

Give staff permission to challenge and a practical escalation contact. A policy that requires verification but penalises delay can fail at the moment it is needed. Rehearse how a junior employee can pause a request apparently coming from a senior executive.

THE NEXT EXERCISE

Measure the quality of the decision, not the realism of the fake.

Choose a plausible request affecting money, access, or reputation. Use a safe simulation with agreed boundaries and no real transfer. Test whether the receiving team identifies the required evidence, uses the independent route, preserves the record, and reaches an authorised decision.

Review the process without blaming the person who found a representation persuasive. Ask whether contact information was usable, the approval policy covered the request, and the escalation owner was available. Assign responsibility for the gap that the exercise exposed.

The business does not need every employee to outperform a forensic laboratory. It needs a decision process that remains dependable when a face, a voice, or a file cannot settle the question on its own.