Deepfake Fraud Statistics 2026
By Axis Intelligence Research
Co-author: Marcus Chen, Cybersecurity | Last updated: September 4, 2026 | License: CC BY 4.0
The FBI’s Internet Crime Complaint Center logged $893,346,472 across 22,364 AI-related complaints in 2025 — the first time in the bureau’s 25 years of cybercrime reporting that artificial intelligence appears as its own descriptor. That single pool is the only audited, nationally-scoped deepfake fraud number in existence. Axis Intelligence Research decomposed it by crime type. What comes out is not the story the surge percentages tell.
Quick Answer
Deepfake fraud statistics rest on one measured federal figure and a great deal of vendor telemetry that is often mistaken for it. In 2025 the FBI IC3 attributed $893.3 million in adjusted losses to 22,364 AI-nexus complaints — 4.28% of the $20.877 billion in total reported cybercrime losses, from just 2.22% of complaints. According to Axis Intelligence Research, that asymmetry means an AI-nexus complaint carried 1.93 times the loss of the average IC3 complaint. Separately, Entrust’s 2026 Identity Fraud Report found deepfakes in one in five biometric fraud attempts across more than a billion identity verifications.
Key Findings
- According to Axis Intelligence Research’s decomposition of the FBI’s 2025 Internet Crime Report, AI-nexus complaints represented 2.22% of IC3 complaint volume but 4.28% of reported losses — a severity multiple of 1.93x.
- Axis Intelligence Research finds that phishing and spoofing carries the highest loss amplification of any high-volume category: an AI-nexus phishing complaint averaged $12,806.64 against $1,126.76 for phishing complaints overall, a Synthetic Loss Amplification Ratio of 11.37.
- Investment fraud absorbed $632,041,188 — 70.75% of the entire federal AI-nexus loss pool — from 4,356 complaints, per Axis Intelligence Research’s calculation from IC3 crime-type tables.
- Axis Intelligence Research finds that complainants aged 60 and over accounted for 14.05% of AI-nexus complaints but 39.46% of AI-nexus losses, at an average of $112,152.79 per complaint — 2.81 times the all-ages AI-nexus average.
- 83.02% of the federal AI-nexus loss pool carried a cryptocurrency nexus ($741,639,787 of $893,346,472), making crypto rails the dominant settlement layer for synthetic-media fraud, per Axis Intelligence Research’s cross-table calculation.
What Do Deepfake Fraud Statistics Actually Measure?
Six figures circulate as “the” deepfake fraud number. They measure six different things, and blending them is how a $893 million problem gets reported as a $442 billion one.
The FBI counts complaints filed by victims who mentioned AI, with losses adjusted by IC3 analysts. Entrust counts fraud attempts detected inside its own verification network. Resemble AI counts incidents that appeared in news reports. Gartner counts security leaders answering a survey question. None of these is a census. Each has a denominator, and the denominator is the whole argument.
Marcus Chen: No zero-day here, and no global census either. Every “deepfake fraud surged 3,000%” headline you have read this year is a ratio of two numbers from inside one vendor’s funnel, published without the funnel. That is not a criticism of the vendors — telemetry is genuinely useful, and Entrust’s billion verifications are a real observation about a real population. It becomes misinformation the moment somebody strips the population off and calls it global. The federal number is smaller, slower, and less exciting. It is also the only one with an audit trail.
The Five Measurement Boundaries
| Boundary | What it counts | 2025–2026 anchor | Source |
|---|---|---|---|
| Victim-reported loss | Adjusted USD loss on complaints referencing AI | $893,346,472 | FBI IC3 2025 Internet Crime Report |
| Complaint volume | Complaints with an AI descriptor | 22,364 | FBI IC3 2025 Internet Crime Report |
| Detected attempts | Deepfake share of biometric fraud attempts | 1 in 5 (20%) | Entrust 2026 Identity Fraud Report |
| Documented incidents | Attacks verified from news reporting, H1 2026 | 821 attacks, 15,736 victims | Resemble AI H1 2026 Deepfake Threat Report |
| Regulatory signal | SAR filings citing deepfake media | Increase observed from 2023 onward (no count published) | FinCEN FIN-2024-Alert004 |
How Much Money Did Deepfake Fraud Cost in 2025?
The audited answer for the United States is $893,346,472, from 22,364 complaints, at an average adjusted loss of $39,945.74. Against IC3’s 2025 totals of $20.877 billion and 1,008,597 complaints — themselves up 25.77% and 17.34% on 2024 — the AI-nexus pool is 4.28% of the money and 2.22% of the volume.
That gap is the finding. The all-complaint average loss in 2025 was $20,699.05. The AI-nexus average was 1.93 times higher.
The Three Original Axis Metrics
Every number in this section is computed from the IC3 crime-type tables and reproducible from them. The formulas:
SFPR — Synthetic Fraud Penetration Rate. The AI-nexus share of a crime type’s total reported loss. SFPR = AI-nexus loss ÷ total category loss
SLAR — Synthetic Loss Amplification Ratio. How much more (or less) an AI-nexus complaint costs than a typical complaint in the same category. A reading of 1.00 means no amplification. SLAR = (AI-nexus loss ÷ AI-nexus complaints) ÷ (total category loss ÷ total category complaints)
DFEI — Deepfake Fraud Exposure Index. A 0–100 composite ranking crime types by synthetic-media exposure, combining penetration, amplification, and absolute materiality. Geometric weighting is used so that a category weak on any one dimension cannot rank highly on the strength of another.
DFEI = 100 × (SFPR ÷ SFPR_max)^0.40 × (SLAR ÷ SLAR_max)^0.35 × (pool share ÷ pool share_max)^0.25
Weights: penetration 40%, amplification 35%, materiality 25%. Eligibility floor: a category must carry at least $1,000,000 in AI-nexus loss and at least 50 AI-nexus complaints, which admits 19 of IC3’s 26 crime types. Maxima across the eligible set are SFPR 37.889%, SLAR 14.350, pool share 70.750%. First published by Axis Intelligence Research, September 4, 2026. Snapshot date for all readings: IC3 calendar year 2025 data, retrieved 2026-09-03.
Which Fraud Types Are Most Exposed to Synthetic Media?
Deepfake Fraud Exposure Index — 2025 baseline readings
| Rank | Crime type | DFEI | SFPR | SLAR | AI-nexus loss | AI-nexus complaints | Source |
|---|---|---|---|---|---|---|---|
| 1 | IPR/Copyright and Counterfeit | 35.6 | 37.889% | 14.350 | $10,103,789 | 63 | Axis calc. from FBI IC3 2025 |
| 2 | Investment | 21.9 | 7.308% | 1.224 | $632,041,188 | 4,356 | Axis calc. from FBI IC3 2025 |
| 3 | Phishing/Spoofing | 14.4 | 4.764% | 11.366 | $10,283,732 | 803 | Axis calc. from FBI IC3 2025 |
| 4 | Employment | 6.1 | 3.458% | 1.235 | $12,550,185 | 691 | Axis calc. from FBI IC3 2025 |
| 5 | Business Email Compromise | 5.3 | 0.993% | 1.822 | $30,256,592 | 135 | Axis calc. from FBI IC3 2025 |
| 6 | Lottery/Sweepstakes/Inheritance | 5.1 | 2.311% | 2.407 | $4,486,965 | 54 | Axis calc. from FBI IC3 2025 |
| 7 | Confidence/Romance | 4.6 | 2.049% | 0.758 | $19,041,653 | 626 | Axis calc. from FBI IC3 2025 |
| 8 | Harassment/Stalking | 4.5 | 5.217% | 1.474 | $1,445,378 | 763 | Axis calc. from FBI IC3 2025 |
| 9 | Personal Data Breach | 4.1 | 1.427% | 0.800 | $18,767,964 | 1,204 | Axis calc. from FBI IC3 2025 |
| 10 | Extortion | 3.6 | 2.401% | 1.213 | $2,940,642 | 1,764 | Axis calc. from FBI IC3 2025 |
| 11 | Tech/Customer Support | 3.4 | 0.911% | 0.759 | $19,457,078 | 574 | Axis calc. from FBI IC3 2025 |
| 12 | Government Impersonation | 2.9 | 0.885% | 1.104 | $7,061,628 | 260 | Axis calc. from FBI IC3 2025 |
| 13 | Real Estate | 2.4 | 0.981% | 1.055 | $2,699,085 | 115 | Axis calc. from FBI IC3 2025 |
| 14 | Data Breach | 2.3 | 0.992% | 0.655 | $4,319,380 | 60 | Axis calc. from FBI IC3 2025 |
| 15 | Non-Payment/Non-Delivery | 2.0 | 0.740% | 0.687 | $3,726,777 | 609 | Axis calc. from FBI IC3 2025 |
| 16 | Advanced Fee | 1.9 | 1.054% | 0.779 | $1,642,712 | 105 | Axis calc. from FBI IC3 2025 |
| 17 | Credit Card/Check Fraud | 1.7 | 0.650% | 0.877 | $1,836,105 | 139 | Axis calc. from FBI IC3 2025 |
| 18 | Identity Theft | 1.7 | 0.884% | 0.609 | $1,643,308 | 460 | Axis calc. from FBI IC3 2025 |
| 19 | Other | 4.6 | 2.294% | 0.912 | $11,750,591 | 504 | Axis calc. from FBI IC3 2025 |
“Other” is IC3’s residual category and is shown for completeness rather than as an interpretable sector.
Ransomware is the notable zero. Across 3,611 ransomware complaints and $32,320,105 in reported losses, IC3 attributed $0 to an AI nexus in 2025, from 16 AI-flagged complaints. Synthetic media is a social-engineering instrument, not an encryption one, and the federal data draws that line cleanly.
Where the Money Actually Is
Investment fraud is 70.75% of the pool and everything else is noise by comparison. Strip investment out and the remaining 9,140 crime-typed AI-nexus complaints average $18,218.58 — less than half the headline figure. The AI-nexus average of $39,945.74 is an artifact of one category.
Marcus Chen: This is where the deepfake conversation and the actual loss data stop agreeing. The industry narrative is the boardroom — a CFO’s face on a Teams call, a wire out the door. That attack is real and it is expensive per event, but BEC accounts for 3.39% of the federal AI pool. Investment fraud accounts for 70.75%, and the mechanism there is not a real-time video call. It is a fabricated celebrity endorsement, an AI-voiced “portfolio manager,” and a scripted relationship that runs for eleven weeks before the first deposit. Dwell time, not deception speed, is what makes that category expensive.
Deepfake Fraud Statistics by Attack Channel
IC3’s own narrative section names the mechanisms behind the money. In 2025 the bureau recorded over $30 million in BEC losses involving AI, over $19 million in confidence and romance fraud with a likely AI nexus — of which over $5 million came from distress scams using voice cloning to imitate a family member — and almost $13 million in employment fraud, where the tell is voice-audio desynchronised from lip movement during remote interviews. IC3 notes that employment cases carry little direct dollar loss because the objective is network access rather than payment.
The bureau also states plainly that overall investment losses exceeded $8 billion against $632 million flagged as AI-related, “demonstrating that many victims do not realize the extent AI may be involved.”
Why Is Phishing the Most Amplified Deepfake Fraud Vector?
Phishing and spoofing produced the third-highest DFEI reading and by far the most violent amplification outside a rounding-error category. SLAR 11.37. An AI-nexus phishing complaint cost 11.4 times an ordinary phishing complaint.
The category-level trend corroborates it. Phishing and spoofing complaints fell 0.95% year over year — 191,561 in 2025 against 193,407 in 2024 — while reported losses rose 208.29%, from $70,013,036 to $215,843,126. Fewer attempts, three times the money.
Marcus Chen: Flat volume with tripled losses is the signature of a targeting change, not a volume change. The old economics of phishing were spray-and-pray: cheap to send, low conversion, small per-victim take. Synthetic media inverts that. When the follow-up to the email is a cloned voice on the phone confirming the request, conversion rises and the ask can be larger, because the victim’s out-of-band verification path has been closed. The attacker did not get better at writing emails. They got a second channel that the victim was trained to trust.
That is precisely the failure mode Hong Kong police documented at Arup in January 2024, when a finance employee in the firm’s Hong Kong office executed 15 transfers totalling HK$200 million — roughly US$25.6 million — to five local accounts after joining a video conference in which every other participant was synthetic. Arup confirmed to CNN in May 2024 that fake voices and images were used, that it had notified Hong Kong police in January, and that no internal systems were compromised. The employee had been suspicious of the initial email. The video call was the verification step, and the verification step was the attack.
Where the Data Contradicts the Narrative
Four categories returned SLAR readings below 1.00, meaning AI-nexus complaints cost less than the category average:
| Crime type | SLAR | AI-nexus avg. loss | Category avg. loss | Source |
|---|---|---|---|---|
| Confidence/Romance | 0.758 | $30,417.98 | $40,126.41 | Axis calc. from FBI IC3 2025 |
| Tech/Customer Support | 0.759 | $33,897.35 | $44,664.10 | Axis calc. from FBI IC3 2025 |
| Personal Data Breach | 0.800 | $15,588.01 | $19,493.06 | Axis calc. from FBI IC3 2025 |
| Identity Theft | 0.609 | $3,572.41 | $5,866.86 | Axis calc. from FBI IC3 2025 |
Romance fraud is the one worth pausing on, because it is the category most often invoked in deepfake coverage. Two readings are consistent with the data: AI is being deployed against a broader, shallower victim pool in romance fraud, or the highest-value long-cycle romance operations still rely on human operators and AI is concentrated in the cheaper end. Axis Intelligence Research does not have the evidence to choose between them, and says so rather than picking the more quotable one.
Who Loses the Most Money to Deepfake Fraud?
By Age
| Segment | AI-nexus complaints | Share of AI complaints | AI-nexus loss | Share of AI loss pool | Avg. loss | Source |
|---|---|---|---|---|---|---|
| 60+ | 3,143 | 14.05% | $352,496,231 | 39.46% | $112,152.79 | FBI IC3 2025 |
| 17 and under | 355 | 1.59% | $126,391 | 0.01% | $356.03 | FBI IC3 2025 |
| All ages | 22,364 | 100% | $893,346,472 | 100% | $39,945.74 | FBI IC3 2025 |
Complainants aged 60 and over carried 2.81 times the AI-nexus loss per complaint of the all-ages average. AI-nexus losses were 4.55% of all reported 60+ losses of $7.748 billion — a cohort whose total losses rose 59% year over year against a 37% rise in complaints.
By Settlement Rail
$741,639,787 of the AI-nexus pool — 83.02% — carried a cryptocurrency nexus, across 8,712 complaints averaging $85,128.53. Crypto-nexus complaints were 38.96% of AI-nexus complaints but 83.02% of the losses.
Marcus Chen: The rail is the reason the losses are permanent. Wire fraud has a recall window; IC3’s Recovery Asset Team froze $679,013,183 of $1,163,919,846 in attempted theft last year, a 58% success rate on the cases it could reach. Crypto has no equivalent. Eighty-three percent of the synthetic-media loss pool moved on rails with no reversal mechanism, which is why the effective control is upstream verification, not downstream recovery.
What Do the Regulators Require?
FinCEN: deepfake fraud has had a SAR keyword since November 2024
FinCEN issued FIN-2024-Alert004 on November 13, 2024, instructing financial institutions to flag suspected deepfake-enabled fraud with the key term FIN-2024-DEEPFAKEFRAUD in SAR field 2. The alert states that from 2023 onward FinCEN observed rising SAR volume describing deepfake media, most often fraudulent identity documents used to defeat customer identification and customer due diligence controls — and that criminals have opened accounts with GenAI-produced identity documents and used them as funnel accounts to launder proceeds of other fraud.
The alert lists nine red-flag indicators. Three are operationally distinctive and rarely reproduced accurately:
- A customer uses a third-party webcam plugin during a live verification check, or requests a change of communication method citing repeated technical glitches.
- A reverse-image lookup of an identity photo matches an image in an online gallery of GenAI-produced faces.
- A customer declines multifactor authentication to verify their identity.
FinCEN does not publish a count of FIN-2024-DEEPFAKEFRAUD SAR filings. Axis Intelligence Research will not estimate one.
FTC: impersonation is now the largest reported fraud category
The FTC reported that consumers lost $3.5 billion to imposter scams in 2025, the most-reported fraud category, with nearly one in three fraud reports concerning impersonation. Business impersonators took nearly $1 billion — up from $866 million in 2024, a rise of about 15.5% — and government impersonators about $920 million, up from $789 million, about 16.6%. Total reported fraud losses reached about $16 billion, roughly 25% above 2024. The FTC’s Impersonation Rule, finalised in 2024, has since produced a dozen enforcement actions and over $70 million in consumer redress.
EU: labelling deepfakes became mandatory on 2 August 2026
Article 50 of Regulation (EU) 2024/1689 has applied since 2 August 2026. Deployers must inform individuals exposed to deepfakes; providers must add machine-readable marks enabling detection of AI-generated or manipulated content. The European Commission published final guidelines on the Article 50 transparency obligations on 20 July 2026, alongside a Code of Practice on Transparency of AI-generated Content. Non-compliance carries penalties of up to €15 million or 3% of worldwide annual turnover under Article 99.
Marcus Chen: Article 50 governs disclosure by lawful deployers. It does not govern the Southeast Asian scam compound generating an investment-pitch video, and nobody at the Commission claims otherwise. The realistic compliance effect is provenance infrastructure — if legitimate generators mark output reliably, unmarked synthetic content becomes a weak signal rather than no signal. That is worth something. It is not a control.
What Does Enterprise Detection Data Show?
Vendor telemetry answers a question the federal data cannot: how often synthetic media is attempted against identity systems, as opposed to how often it succeeds and gets reported.
| Metric | Value | Population | Source |
|---|---|---|---|
| Deepfake share of biometric fraud attempts | 1 in 5 (20%) | 1bn+ identity verifications, 195 countries, 30+ industries | Entrust 2026 Identity Fraud Report |
| Deepfaked selfie attempts, 2025 change | +58% | Same | Entrust 2026 Identity Fraud Report |
| Injection attacks, year-over-year | +40% | Same | Entrust 2026 Identity Fraud Report |
| Digital forgeries as share of document fraud | 35% (vs 29% avg. 2022–2024) | Same | Entrust 2026 Identity Fraud Report |
| National ID cards as share of fraudulent documents | 46% | Same | Entrust 2026 Identity Fraud Report |
| Verified attacks, H1 2026 | 821, from 1,760 news reports | Publicly reported incidents | Resemble AI H1 2026 Deepfake Threat Report |
| Documented victims, H1 2026 | 15,736 minimum | Same | Resemble AI H1 2026 Deepfake Threat Report |
| Synthetic files linked to those attacks | ~3.46 million | Same | Resemble AI H1 2026 Deepfake Threat Report |
The Entrust one-in-five figure is a share of biometric fraud attempts inside one verification network. It does not mean one in five logins, one in five users, or one in five fraud events. That distinction is the difference between a defensible citation and a retraction. The same figure anchors Axis Intelligence Research’s ADIFSI index in our digital identity fraud statistics, where the downstream account-takeover mechanics are broken out.
Injection attacks deserve their own line because they bypass the human entirely. Rather than presenting a synthetic face to a camera, the attacker feeds manipulated video directly into the verification data stream — which is why liveness detection alone, without injection resistance, closes a door that attackers have stopped using.
Figures Axis Intelligence Research Refused to Publish
Verification is only credible if the failures are visible. Six widely-circulated deepfake fraud statistics did not survive our checks, and we are naming them so that anyone who has cited them can reconsider.
| Claim in circulation | Why it was refused |
|---|---|
| “$442 billion in AI fraud in 2025” | Traces to a total-financial-fraud figure relabelled as AI-specific in secondary coverage. Against the measured $893.3M federal AI figure, the mislabel inflates by roughly 500x. |
| “Global deepfake fraud losses of $3.7bn / $2.19bn / $1.56bn” | Three incompatible totals attributed to overlapping incident-database compilations, published within months of each other. No reconciliation is available. We report the components, not a synthesised total. |
| “Deepfakes involved in ~40% of BEC incidents by Q1 2026” | No named dataset, sample, or publisher. Federal data puts AI-nexus BEC at 0.99% of BEC losses in 2025. |
| “Deepfake fraud surged 1,633% / 3,892% / 3,000%” | Percentage changes without a baseline period, a measured population, or a disclosed methodology. Unusable as evidence. |
| “$40 billion in US AI-enabled fraud by 2027” | A projection, not a measurement. Legitimate as a forecast; not a statistic, and not citable as one. |
| A single corporate-vs-consumer loss split from one 2025 incident compilation | The published consumer sub-total exceeds the published grand total. Internally inconsistent as issued. |
Marcus Chen: Every one of those figures outranks the real ones. That is the actual state of this vertical: the most-cited deepfake fraud statistics are the least checkable, because a four-digit surge percentage travels and a $12,806 conditional mean does not. We would rather publish the smaller true number and show our work.
Methodology
Collection. All quantitative claims derive from documents fetched and read on 2026-09-03: the FBI IC3 2025 Internet Crime Report (published April 2026), FinCEN alert FIN-2024-Alert004 and its accompanying release, the FTC press release of June 15, 2026 on 2025 Consumer Sentinel imposter-scam data, the European Commission’s Article 50 transparency guidelines page (last updated 6 August 2026), Entrust’s 2026 Identity Fraud Report announcement of November 18, 2025, Resemble AI’s H1 2026 Deepfake Threat Report page (August 12, 2026), and CNN’s May 16, 2024 report carrying Arup’s confirmation. Retrieval dates and URLs for every row appear in the accompanying CSV.
Derivation. SFPR, SLAR and DFEI are computed exclusively from the two IC3 crime-type tables (“AI References by Complaint Count” and “AI References by Complaint Loss”) and the corresponding all-complaint tables in the same report. Arithmetic was verified twice — once in floating point, once in exact rational arithmetic using Python’s fractions.Fraction — and every prose figure was matched against a CSV row in an adversarial token audit.
Scope of the AI-nexus breakdown. IC3’s published crime-type detail accounts for $798,559,009 of the $893,346,472 AI-nexus loss pool (89.39%) and 13,496 of 22,364 AI-nexus complaints (60.35%). The remainder is not broken out in the report. All SFPR, SLAR and DFEI readings are therefore computed on the crime-typed subset; pool-share figures use the published $893,346,472 denominator and are labelled as such. IC3’s all-complaint crime-type tables similarly account for $20,295,363,088 of the $20.877 billion headline and 768,452 of 1,008,597 complaints. Ratios within a category use numerator and denominator drawn from the same table pair, so the coverage gap does not distort them.
What the index does not capture. DFEI measures reported, adjusted, US-only, victim-initiated loss. It cannot see unreported corporate losses suppressed for reputational reasons, attacks stopped before payment, non-financial harm, or any jurisdiction outside IC3’s reporting population. Complaint-based data also carries a self-selection bias toward victims who recognised AI involvement — which, given IC3’s own note that many investment-fraud victims did not, means the AI-nexus pool is a floor.
Versioning. DFEI methodology is versioned internally; the 2025 readings published here are the baseline. They do not describe themselves relative to a history that has not been computed. Any future change to components, weights or the eligibility floor will be disclosed and re-baselined transparently, and historical readings will not be silently rewritten.
About This Dataset
deepfake-fraud-statistics.csv contains every figure in this article as a row, with value, unit, as-of date, geography, segment, source organisation, source document, source URL, retrieval date, primary-source flag, Axis-calculated flag, and — for calculated rows — the formula in method_note.
Licence: CC BY 4.0. Free to reuse, including commercially, with attribution.
Distribution: Hugging Face, Kaggle, GitHub.
Citation
APA — Axis Intelligence Research. (2026). Deepfake fraud statistics 2026: Federal AI-nexus loss decomposition and the Deepfake Fraud Exposure Index. https://axis-intelligence.com/deepfake-fraud-statistics/
MLA — Axis Intelligence Research. “Deepfake Fraud Statistics 2026: Federal AI-Nexus Loss Decomposition and the Deepfake Fraud Exposure Index.” Axis Intelligence, 3 Sept. 2026, axis-intelligence.com/deepfake-fraud-statistics/.
Chicago — Axis Intelligence Research. “Deepfake Fraud Statistics 2026: Federal AI-Nexus Loss Decomposition and the Deepfake Fraud Exposure Index.” Axis Intelligence, September 4, 2026. https://axis-intelligence.com/deepfake-fraud-statistics/.
Frequently Asked Questions
If a deepfake video call is the verification step, what actually stops the transfer?
A payment control that does not depend on recognising a face or a voice. The three that held in documented 2024–2025 cases were a pre-agreed callback to a number held in the finance system rather than one supplied in the request, a codeword shared out of band, and a mandatory second authoriser for any first-time beneficiary above a threshold. Note what these have in common: none of them asks an employee to detect a fake. FinCEN’s own guidance points the same direction, listing avoidance of multifactor authentication and requests to switch communication channels as red flags — the attacker’s tell is evading the verification path, not failing it.
Our bank’s SAR team asks: is there a filing obligation specific to deepfakes?
There is no separate obligation, but there is a required tag. FinCEN alert FIN-2024-Alert004 asks institutions to include the key term FIN-2024-DEEPFAKEFRAUD in SAR field 2 and in the narrative where suspected deepfake media is connected to the activity, plus any key term for the underlying typology. Ordinary BSA thresholds and timelines apply; the keyword is what lets FinCEN aggregate the typology.
Does liveness detection defeat deepfake account-opening fraud?
Not on its own, and increasingly not at all against the fastest-growing vector. Injection attacks feed manipulated video directly into the verification data stream rather than presenting it to a camera, which means live-capture checks never see the physical scene they are designed to assess. Entrust recorded injection attacks up 40% year over year. Effective stacks pair liveness with injection resistance, device and session intelligence, document checks, and behavioural risk — with manual review reserved for high-risk exceptions.
Which sector should worry about voice cloning versus video?
The channel follows the workflow, not the sector. Any process where a voice is the final authorisation — contact-centre account recovery, insurance claims intake, wire callback verification — is exposed to voice cloning. Any process where a face on a screen closes a decision is exposed to video. IC3’s employment-fraud data shows the third pattern: synthetic candidates in remote interviews, where the objective is network access rather than payment, and where the reported tell is audio desynchronised from lip movement.
Why do federal figures look so much smaller than vendor numbers?
They count different events. IC3 counts victim-reported, analyst-adjusted losses on US complaints where the complainant mentioned AI. Vendors count attempts detected inside their own networks, most of which are blocked and none of which produce a loss figure. A blocked attempt and a completed wire are both real; they are not the same unit, and dividing one by the other produces nothing.
Is the $893 million figure the true cost of deepfake fraud in the US?
No — it is a floor, and IC3 says as much implicitly. The report notes that overall investment-fraud losses exceeded $8 billion against $632 million flagged as AI-related, because many victims never learn AI was involved. Add corporate losses absorbed quietly for reputational reasons and the underreporting that affects all complaint-based data, and $893.3 million is the visible portion of a larger number that nobody can currently size honestly.
What does EU AI Act Article 50 require of a company that publishes AI-generated marketing video?
Since 2 August 2026, a deployer publishing deepfake content — synthetic media resembling a real or plausibly real person, object or event, presented as authentic — must disclose that it is artificially generated or manipulated. Providers of the generating system must apply machine-readable marks. Exceptions exist for standard editing and for human-reviewed editorial content. Penalties reach €15 million or 3% of worldwide annual turnover.
Does deepfake fraud show up in ransomware incidents?
Not in the 2025 federal data. Across 3,611 ransomware complaints and $32.3 million in reported losses, IC3 attributed zero dollars to an AI nexus. Synthetic media is an instrument of social engineering and identity fraud; it does not encrypt anything. Where the two connect is upstream — a cloned voice used against a help desk to obtain the credentials that later enable intrusion.
What is the single most reliable deepfake fraud statistic to cite in 2026?
$893,346,472 across 22,364 AI-related complaints, FBI IC3 2025 Internet Crime Report, published April 2026. It is measured rather than surveyed or modelled, nationally scoped, methodologically documented, and it carries a named federal publisher. Pair it with a denominator — 4.28% of $20.877 billion in total reported losses — so that readers can see the proportion rather than only the headline.
Corrections and data contributions: [email protected]. Axis Intelligence Research publishes dated corrections at /corrections/ and does not silently amend published figures.
Related Axis Intelligence Research datasets: digital identity fraud statistics · credential attack statistics · AI scam statistics · online fraud statistics · social engineering statistics · AI in banking statistics · Insider Threat Statistics