Deepfake examples: documented incidents with primary records

This page sits in the deepfake detection hub beside the deepfake statistics with a primary source page, which applies the same rule to numbers. Here the rule applies to incidents: a case appears only if a government, regulator or court has described it in a document anyone can read, and the record is linked in the same paragraph.

01Why list only incidents with an official record?

Most lists of deepfake examples are built from news reports that cite other news reports. Figures change between retellings, and the technical facts that matter to detection, whether the video was live or pre-recorded, whether a face or a voice was altered, how the clip reached the target, are the first details to drop out. An official record fixes those facts at one point in time.

The cost of the rule is coverage. Many widely reported incidents have no fetchable record from a court, regulator or government, and they are not here. Nothing on this page names a private individual, and companies appear only where the record itself names them. The records below describe how each fake was made and delivered, which is what a reader of a detection benchmark needs: the kind of deepfake a detector would have had to catch, and under what conditions.

02What does the Hong Kong record say about the HK$200 million case?

The primary record is a written reply by Hong Kong's Secretary for Security to the Legislative Council on 26 June 2024 (Hong Kong Government press release LCQ9). It states that the Police kept separate deepfake fraud statistics from 2023 and that, up to 31 May 2024, they had received three cases, in August 2023, January 2024 and May 2024.

The January 2024 case began with a phishing email from a fraudster impersonating the Chief Financial Officer of the informant's head office in the United Kingdom. The informant was invited to a group video conference "for some alleged confidential transactions", and "eventually gave authorisation to transfer funds to five local bank accounts and suffered a loss of about HK$200 million". The Police's finding on the mechanism is the part that matters for detection: "the pre-recorded video conference was generated using downloaded online public video clips and voices of the impersonated officer", and "since the meeting was pre-recorded, there was actually no interaction between the informant and the fraudster". The fraudster ended the meeting under a pretext and continued the payment instructions over instant messaging. The reply records the case as still under investigation; it does not name the company, and neither does this page.

The May 2024 case followed the same shape with a different first contact. On 20 May 2024 a staff member of a multinational trade company received a WhatsApp message impersonating the head office CFO, "took part in a video conference for nearly 30 minutes", and was instructed to transfer nearly HK$4 million to a local account. The Police believed the suspect had altered public videos with deepfake technology and "played the altered content at the video conference". The August 2023 case is different in kind: a syndicate used an AI face-swapping programme to apply for loans online, involving HK$200,000, and nine people were arrested. The same reply counts 21 online deepfake clips impersonating officials or celebrities identified or reported between 1 November 2023 and 31 May 2024, all removed at the Police's request, with no report of anyone defrauded by them.

03What did the FCC rule about AI-generated voices?

On 8 February 2024 the US Federal Communications Commission released a Declaratory Ruling, adopted on 2 February, in docket CG 23-362 (FCC 24-17). It states: "we confirm that the TCPA's restrictions on the use of 'artificial or prerecorded voice' encompass current AI technologies that generate human voices". Calls that use such voices therefore need the prior express consent of the called party under the Telephone Consumer Protection Act, absent an emergency or an exemption.

The reasoning is short and useful as a definition. Cloned voices "are 'artificial' voice messages because a person is not speaking them", and "AI technologies such as 'voice cloning' fall within the TCPA's existing prohibition on artificial or prerecorded voice messages". The Commission notes that "no commenter opposes these conclusions" and that the ruling took effect on release. It also names the harm it is aimed at: voice cloning "can convince a called party that a trusted person, or someone they care about such as a family member, wants or needs them to take some action".

04Which incident has a court or enforcement record?

The FCC's Enforcement Bureau order and consent decree of 21 August 2024 with Lingo Telecom, LLC is an enforcement record of a specific deepfake incident (DA 24-790). The order states that on 21 January 2024, two days before the New Hampshire Democratic primary, potential voters "received thousands of spoofed robocalls carrying a deepfake generative artificial intelligence (AI) voice message" purporting to be from the sitting US president, and that "the prerecorded AI message, using voice cloning technology, told recipients not to vote in the upcoming Primary Election".

The company's liability concerned caller ID attestation under the STIR/SHAKEN rules (47 CFR 64.6301(a)), not the making of the fake itself. A Notice of Apparent Liability of 23 May 2024 proposed a US$2,000,000 penalty; the consent decree settles at a US$1,000,000 civil penalty and a compliance plan. The person who commissioned the calls is the subject of separate proceedings and is not named here.

05What does the FBI's alert add?

The FBI's Internet Crime Complaint Center issued Public Service Announcement I-120324-PSA on 3 December 2024, "Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud" (IC3 PSA). It is a catalogue of methods, not of cases: it lists AI-generated text, images, audio ("vocal cloning") and video, and under video states that "criminals generate videos for real time video chats with alleged company executives, law enforcement, or other authority figures". Under audio it describes "short audio clips containing a loved one's voice to impersonate a close relative in a crisis situation".

Two sentences in the alert are worth keeping. It defines synthetic content as "the class of media generated or manipulated by machine-learning-based techniques", and it states that "the creation or distribution of synthetic content is not inherently illegal". The alert gives no counts and no dollar figures, which matches what the statistics page found in the Bureau's 2024 annual report: the FBI publishes no deepfake loss total.

06What do the records have in common?

Three features repeat. The source material was public: in both Hong Kong conference cases the Police believed the fakes were built from downloaded public clips of the impersonated person, which is the same raw material the kinds of face deepfake page describes as the input to a swap or reenactment. The fake was not interactive: the January 2024 conference was pre-recorded, and the robocall was a prerecorded message. And the instruction arrived through a second, ordinary channel once the fake had done its work. No record among these describes a live, interactive video deepfake, and none describes a detector being run. In the four-type vocabulary of the DF40 dataset, the two conference cases are reenactments or swaps of a real person and the robocalls are voice clones, a kind no face benchmark covers; a detector score on any of them would have to be read against the ROC curve for deepfake detection, at an operating point chosen before the call.

That last point is why the Hong Kong reply's advice is to request "certain movements in front of the camera, such as turning their head up, down, left or right", and why the FBI's advice is to "create a secret word or phrase with your family" and to verify "by hanging up the phone" and calling back. Both are checks of liveness and provenance rather than of pixels, the order of checks the checking a face video page follows. The in-the-wild results explain the ordering: detectors tested on circulated fakes scored far below their benchmark figures, so a detector score would have been weak evidence in any of these cases even if one had been run. The incidents also show why the deepfake meaning and origin page treats voice as part of the word: two of the four records concern cloned voices with no face at all.