AI-Enabled Fraud Changes What Counts as Evidence

In early 2024, a Hong Kong employee received what appeared to be an email from the company’s chief financial officer. A group video call seemed to remove the doubt: familiar colleagues and senior executives appeared on screen and reinforced the request. The employee subsequently authorized transfers totaling about HK$200 million. Police later concluded that the conference had been fabricated using publicly available video and voice material, with payment instructions continuing through messaging. The case is memorable because the deepfake was convincing, but its more important audit lesson is this: several pieces of apparently independent evidence can be generated from the same manipulated source. The Hong Kong Government later described the case and its mechanics.

Generative AI does not make traditional evidence principles obsolete. It makes their disciplined application more important. When voice, video, identity documents, invoices, emails, and screenshots can be created or altered cheaply, the question can no longer stop at “Does this look authentic?” Auditors increasingly need to ask: Where did this come from? Who or what actually authorized it? Is the corroborating source genuinely independent? Can the record be traced back to a controlled system?

The problem is not just fake evidence. It is correlated evidence

Auditors have always understood that repeating the same assertion does not make it more reliable. AI-enabled fraud scales that problem across media. One actor can now produce an email, cloned voice, video call, identity document, and supporting attachment that all tell the same story while creating the illusion of independent corroboration.

Standard 14.1 of the Global Internal Audit Standards requires internal auditors to gather information that is relevant, reliable, and sufficient. In explaining what increases reliability, the Standards point to information obtained directly by the internal auditor or from an independent source, corroborated information, and information drawn from systems with effective governance, risk management, and control processes. None of those reliability factors depends on how realistic a file looks.

That distinction changes the way corroboration should be evaluated. Different media are not the same thing as independent sources. If an apparent executive requests a transfer by email and then appears on video to confirm it, both signals may still be controlled by the same fraudster. Independence must exist in the origin or governance of the evidence, not merely in its format.

Four assumptions auditors can no longer make

Apparent evidence	AI-enabled risk	Stronger audit response
Voice or video of a known person	Voice cloning, avatars, or prerecorded or real-time deepfakes.	Use an out-of-band confirmation through trusted contact data and inspect corporate identity and authentication records.
Professional-looking document or ID	Synthetic or altered invoices, letters, PDFs, or identity documents that are internally consistent.	Verify the original in the system of record, the issuer, document history, and where relevant its provenance and metadata.
Approval in email, chat, or a meeting	Impersonation, account compromise, or an approval fabricated outside the official workflow.	Inspect workflow authorization, segregation of duties, authenticated-session evidence, and transaction traceability.
Several artifacts that tell the same story	The same actor or model generates all artifacts and creates false corroboration.	Cross-check against separately governed sources that do not depend on the compromised channel.

This does not mean video, PDFs, or messages no longer count as evidence. It means their evidential value depends much more heavily on provenance, control context, and their relationship to other sources.

The same shift matters for fraud controls. The COSO/ACFE Fraud Risk Management Guide frames fraud risk management around governance, risk assessment, preventive and detective controls, investigation, and monitoring. AI does not replace that architecture. It adds scenarios that must now be designed into it: executive impersonation, synthetic identities, fabricated documents, and approval channels that previously carried more implicit trust.

A practical evidence test: provenance, independence, integrity, and traceability

For engagements where digital manipulation could matter, auditors can organize their challenge around four questions.

Provenance. Do we know who or what created the artifact, when it was created, and how it reached us? The NIST AI RMF Generative AI Profile emphasizes content provenance, sources, timestamps, and other metadata as useful risk-management considerations. ISO/IEC 42001 similarly brings provenance and documented information into the management of AI-related data and resources. ISO/IEC 42001:2023 does not certify a particular incoming video or file as authentic, but its management-system logic reinforces disciplined risk assessment, documentation, monitoring, and traceability around AI systems.

Independence. Could the second source be affected by the same actor, account, device, channel, or provider? A confirmation is not truly independent if the fraudster controls both ends of the conversation.

Integrity. Can we establish that the record being examined is the original, or a controlled copy, and that its history is understood? Depending on the risk, that may involve the source system, version history, original files, session logs, or other technical records rather than a screenshot alone.

Traceability. Could another competent person repeat the procedure and understand why the same conclusion was reached? That question connects directly to sufficiency under the Global Internal Audit Standards and helps prevent conclusions that rest on visual impressions that cannot be reproduced.

Detection technology can support this work, but it should not become the next shortcut. NIST’s Reducing Risks Posed by Synthetic Content notes that detection, watermarking, and provenance techniques all have limitations and none offers a comprehensive solution on its own. For high-consequence use cases, a defense-in-depth approach is stronger than relying on a single “deepfake detector.”

Corroboration is stronger when it comes from genuinely independent sources

What Internal Audit and fraud teams should test

The most important impact of AI-enabled fraud is not the addition of one more technology risk. It is the need to revisit business controls that assumed identity and authenticity were easy to infer. Useful audit procedures include:

Update the fraud risk assessment for synthetic impersonation, fabricated identities, generated documents, and manipulation of approval channels, especially in payments, bank-account changes, treasury, recruitment, and customer support.
Identify high-consequence decisions where the channel used to request an action should not also be the channel used to verify it. A transfer requested in a video call, for example, may require confirmation through a separately controlled corporate channel.
Test the authorization system, not just the approval message. Determine who had authority, how the approver authenticated, whether segregation of duties operated, and what evidence the system retained.
Inspect what the organization preserves as evidence. Screenshots are weak substitutes when the original system contains transaction history, version history, identity events, or better-governed logs.
Evaluate third-party identity and synthetic-content controls, including known assumptions, service changes, escalation paths, and the consequences of false positives and false negatives.
Run AI-enabled fraud scenarios through control testing. Do not ask only whether a procedure exists. Test what employees would actually do when an urgent request arrives with a convincing voice, video, and supporting documentation.

FinCEN has reported increased suspicious activity involving deepfake media and described the use of fraudulent identity documents to bypass verification processes. Its November 2024 alert also links deepfakes to business email compromise, phishing, and executive impersonation. The audit implication is straightforward: identity and authorization controls need to be tested under adversarial conditions, not only under the assumption that people and artifacts are what they appear to be.

Future evidence will be less visual and more verifiable

Digital transformation has steadily moved audit evidence into files, systems, screenshots, remote meetings, and electronic communications. Generative AI does not reverse that trend. It forces the profession to mature it.

Convincing content will remain useful, but appearance will carry less weight. Source systems, provenance, authentication, segregation of duties, technical logs, independent corroboration, and reproducibility will matter more. The practical shift is from trusting what looks authentic to demonstrating what can be verified.

Different formats do not equal independent sources.

That principle extends beyond deepfakes. As AI becomes capable of generating coordinated documents, conversations, images, and decisions, audit quality will increasingly depend on our ability to separate persuasive content from reliable evidence.