Voice Cloning Scam Statistics 2026
By Axis Intelligence Research
Co-author: Marcus Chen, Cybersecurity | Last updated: September 12, 2026 | License: CC BY 4.0
Federal data recorded 22,364 AI-related fraud complaints in 2025 with $893,346,472 in losses, the first year the FBI tracked artificial intelligence as a formal descriptor. Within that total, voice cloning appears across seven distinct crime types — and Axis Intelligence Research finds the two most exposed are not the ones consumer coverage focuses on.
Quick Answer
Axis Intelligence Research finds that an AI-nexus fraud complaint cost a U.S. victim an average of $39,946 in 2025 — 1.93 times the $20,699 average across all 1,008,597 complaints filed with the FBI’s Internet Crime Complaint Center. Voice cloning is not a standalone crime category in federal reporting; it is an attack technique distributed across investment fraud, business email compromise, distress scams and employment fraud. On the Voice Clone Exposure Index (VCEI) built for this report, investment fraud scores 78.8 and business email compromise 63.7 out of 100 — far above the family-emergency scams that dominate public awareness campaigns.
Key Findings
- Axis Intelligence Research finds that AI-related complaints made up 2.22% of all 2025 IC3 complaints but 4.28% of all reported losses, a loss-density ratio of 1.93 to 1 (FBI IC3, 2025 Internet Crime Report, April 2026).
- Axis Intelligence Research finds that complainants aged 60 and over filed 14.1% of AI-related complaints but absorbed 39.5% of AI-related losses — $112,153 per complaint, 2.81 times the all-AI average.
- Axis Intelligence Research finds that 83.0% of AI-related fraud losses in 2025 involved cryptocurrency ($741,639,787 of $893,346,472), even though only 39.0% of AI-related complaints carried a crypto nexus.
- AI-nexus business email compromise carried the highest loss density of any crime type at $224,123 per complaint — 135 complaints, $30,256,592 in losses (FBI IC3, 2025).
- In a peer-reviewed Berkeley study of 604 listeners, participants correctly identified an AI-generated voice clone only 60.8% of the time, and judged a real speaker and their clone to be the same person on a median 83.3% of trials (Barrington, Cooper & Farid, Scientific Reports, March 2025).
How Much Money Do Voice Cloning Scams Actually Steal?
The honest answer is that nobody knows the exact figure, and any publication quoting a precise “voice cloning losses” number for the United States is inventing it. Voice cloning has no line in the federal ledger. What exists instead is a descriptor.
The FBI’s Internet Crime Complaint Center added an AI Related descriptor to its taxonomy for the 2025 reporting year — the first time in the center’s 25-year history that artificial intelligence appeared as a tracked field. That descriptor caught 22,364 complaints and $893,346,472 in adjusted losses, against a 2025 total of 1,008,597 complaints and $20.877 billion.
The narrowest, most defensible voice-cloning figure in U.S. federal data is smaller and more specific. Within the confidence and romance category, the IC3 report states that victims claimed losses over $5 million in 2025 to distress scams — the scheme in which a cloned voice imitates a relative in trouble. That is the number. It is not the ceiling, and the FBI’s own framing makes clear why: the descriptor only fires when the complainant recognised AI was involved.
| Metric, 2025 | Value | Source |
|---|---|---|
| AI-related complaints | 22,364 | FBI IC3, 2025 Internet Crime Report |
| AI-related losses | $893,346,472 | FBI IC3, 2025 Internet Crime Report |
| Distress-scam losses (voice-clone “family in trouble”) | over $5,000,000 | FBI IC3, 2025 Internet Crime Report |
| All IC3 complaints | 1,008,597 | FBI IC3, 2025 Internet Crime Report |
| All IC3 losses | $20,877,000,000 | FBI IC3, 2025 Internet Crime Report |
| Average loss, AI-related complaint | $39,946 | Axis Intelligence Research calculation |
| Average loss, all complaints | $20,699 | FBI IC3, 2025 Internet Crime Report |
Marcus Chen: The gap between “$893 million in AI-related fraud” and “$5 million in distress scams” is not a contradiction. It is a measurement artifact, and it is the single most important thing to understand about this data. A descriptor that depends on victim recognition measures awareness of the attack surface, not the attack surface. When someone loses $40,000 to a tech-support caller who sounded exactly like a bank investigator, the complaint gets coded as tech support. Nobody asks whether the voice was synthetic, and the victim has no way to know. Every figure on this page is a floor.
Source : axis-intelligence.com/voice-cloning-scam-satellite-statistics/
Which Fraud Types Are Most Exposed to Voice Cloning?
This is where compilation ends and analysis starts. The IC3 report documents, in its own narrative, exactly where voice cloning and voice spoofing appear as tactics: business email compromise (voice clones requesting wire payment), confidence and romance fraud (distress scams), employment fraud (voice spoofing during remote interviews), investment fraud (AI-generated voices of executives and celebrities), and the impersonation and support categories that run through call centres.
Axis Intelligence Research built the Voice Clone Exposure Index (VCEI) to rank those categories. The index answers one question: if an attacker adds a synthetic voice to this crime type, how much leverage does it buy?
Voice Clone Exposure Index — 2025 reading
| Crime type | AI-nexus complaints | AI-nexus losses | Loss per complaint | AI share of crime type | VCEI |
|---|---|---|---|---|---|
| Investment | 4,356 | $632,041,188 | $145,097 | 5.97% | 78.8 |
| Business Email Compromise | 135 | $30,256,592 | $224,123 | 0.55% | 63.7 |
| Confidence/Romance (incl. distress) | 626 | $19,041,653 | $30,418 | 2.70% | 26.3 |
| Employment | 691 | $12,550,185 | $18,162 | 2.80% | 23.6 |
| Tech/Customer Support | 574 | $19,457,078 | $33,897 | 1.20% | 17.1 |
| Government Impersonation | 260 | $7,061,628 | $27,160 | 0.80% | 12.6 |
| Phishing/Spoofing | 803 | $10,283,732 | $12,807 | 0.42% | 6.2 |
Source: Axis Intelligence Research, computed from FBI IC3 2025 Internet Crime Report. Formula disclosed in Methodology. Snapshot date: September 12, 2026.
Two readings stand out.
Business email compromise is the highest-value target per attempt. At $224,123 per AI-nexus complaint, AI-assisted BEC is roughly 12 times more damaging per incident than distress scams and 17 times more damaging than AI-flavoured phishing. The volume is tiny — 135 complaints out of 24,768 BEC reports — which is precisely why it scores 63.7 rather than 90. The technique has not saturated the category yet. When it does, the loss-per-complaint figure is the number that will move.
Investment fraud is where voice cloning has already scaled. Nearly 6% of all investment-fraud complaints in 2025 carried an AI reference, and the dollar exposure is $632 million — 71% of all AI-nexus losses in the dataset. This is the celebrity-endorsement and fake-advisor machinery: synthetic voices of recognisable figures in video clips, and AI-run “investment club” calls. The FBI notes separately that total investment losses exceeded $8 billion, which means the AI-flagged portion is a small visible tip of a category where victims rarely diagnose the tooling.
Marcus Chen: The grandparent scam gets the airtime because it is legible — a panicked voice, a grandmother, a wire transfer. The exposure data says something less comfortable. The highest-leverage use of a cloned voice in 2025 was not fooling a family member into sending $8,000. It was clearing a payment control inside a company, where a single call authorises a six-figure transfer, and where the only verification in the chain was recognising a familiar voice on a phone.
Who Loses the Most to AI-Enabled Fraud?
The age distribution in AI-related fraud is more skewed than in fraud generally, and by a wide margin.
| Cohort, 2025 | AI-related complaints | AI-related losses | Loss per complaint | Source |
|---|---|---|---|---|
| Aged 60+ | 3,143 | $352,496,231 | $112,153 | FBI IC3 / Axis calculation |
| All complainants | 22,364 | $893,346,472 | $39,946 | FBI IC3 / Axis calculation |
| Aged 17 and under | 355 | $126,391 | $356 | FBI IC3 / Axis calculation |
Axis Intelligence Research finds that people aged 60 and over accounted for 14.1% of AI-related complaints but 39.5% of AI-related losses. Their average loss per AI-nexus complaint was 2.81 times the all-cohort figure.
That sits inside a broader 2025 shift. IC3 recorded 201,266 complaints from people aged 60 and over, up 37% year over year, with $7.748 billion in losses, up 59%, and 12,444 complainants in that cohort lost more than $100,000 each.
The channel data reinforces the point. Federal Trade Commission reporting for 2025 shows that phone calls accounted for 11% of fraud reports indicating a loss, well behind websites and apps (31%) and social media (28%). But for people aged 80 and over, the FTC found phone calls were the costliest contact method in aggregate reported losses — the only age band where that is true, and the only band where social media ranked third.
In 2024, the last year for which the FTC published full contact-method tables, phone calls produced 284,659 fraud reports, $948 million in aggregate losses and a $1,500 median loss — the highest median of any contact method, against $600 for email across 371,651 reports. Axis Intelligence Research calculates phone contact at 18.9% of the 1,509,002 reports that specified a channel.
Marcus Chen: Volume moved to the browser years ago. Value never left the phone. The per-report median is the tell: a scam that starts with a call is a scam that has already selected for the victim most likely to move real money, and that selection now runs on a channel where the primary authentication mechanism is “I recognise this voice.” Synthetic speech does not attack a system. It attacks a control that was never a control.
Can People Tell a Cloned Voice From a Real One?
The evidence says no, and the best measurement comes from a peer-reviewed study rather than a vendor survey.
Researchers at the University of California, Berkeley cloned 220 speakers using a commercial instant-voice-cloning service and tested 604 listeners across two tasks, publishing in Scientific Reports in March 2025.
| Measurement | Result |
|---|---|
| Correctly identified an AI clip as AI-generated | 60.8% of the time |
| Correctly identified a real clip as real | 67.4% of the time |
| Participants at or below chance on AI clips | 21.0% |
| Median “same person” judgement, real voice vs. its clone | 83.3% of trials |
| Median “same person” judgement, two different real speakers | 11.1% of trials |
Source: Barrington, Cooper & Farid, Scientific Reports 15:11004 (2025).
The identity result is the operationally significant one. Listeners were not biased toward saying “same person” — they correctly separated two genuine but acoustically similar speakers on 89% of trials. They simply could not separate a person from that person’s clone.
What actually improves detection
The same study found detection accuracy varied sharply with what the speaker was saying. Median accuracy ran 56.7% on scripted audio, 76.7% on unscripted audio and 83.3% on combined clips where a speaker read a question aloud and then answered it spontaneously. Longer clips also helped.
That is a defensive control anyone can apply mid-call: keep the caller talking, off-script, for longer. Ask something the script cannot anticipate. Synthetic speech degrades under improvisation in a way it does not under recitation.
What Have U.S. Regulators Actually Done?
Three actions carry legal weight.
The FCC ruled cloned voices illegal in robocalls. On February 8, 2024, the Federal Communications Commission adopted a unanimous Declaratory Ruling holding that AI-generated voices are “artificial” under the Telephone Consumer Protection Act. The ruling took effect immediately and gave state attorneys general a direct enforcement route against voice-cloning robocall operations.
The FTC finalised the Impersonation Rule. Since the rule took effect in 2024, the Commission reports a dozen enforcement actions and over $70 million in consumer redress under it, including 2025 cases against IRS-imposter and government-imposter operations.
The FTC ran the Voice Cloning Challenge. The Commission awarded four prizes for detection and authentication approaches, including real-time liveness scoring of incoming calls in two-second windows and inaudible watermarking designed to break cloning models. The FTC’s own conclusion from the exercise was that no single solution exists.
Meanwhile the underlying impersonation economy grew. FTC data show consumers reported $3.5 billion lost to imposter scams in 2025, the most-reported fraud category, with about $920 million to government impersonators and nearly $1 billion to business impersonators — up from $789 million and $866 million respectively in 2024. Total reported fraud losses reached about $16 billion, the highest on record.
How Fast Is AI-Enabled Fraud Growing?
Year-over-year comparison for the AI descriptor itself is not possible — 2025 is the baseline year. What can be compared are the categories where voice cloning concentrates.
| Crime type | 2023 complaints | 2024 complaints | 2025 complaints | Source |
|---|---|---|---|---|
| Government Impersonation | 14,190 | 17,367 | 32,424 | FBI IC3 |
| Investment | 39,570 | 47,919 | 72,984 | FBI IC3 |
| Tech/Customer Support | 37,560 | 36,002 | 47,794 | FBI IC3 |
| Confidence/Romance | 17,823 | 17,910 | 23,159 | FBI IC3 |
| Business Email Compromise | 21,489 | 21,442 | 24,768 | FBI IC3 |
Government impersonation complaints rose 87% in a single year. Tech and customer support — the other classic call-centre category — rose 33%. Both are categories where the entire attack runs through a voice channel.
Marcus Chen: Compare the same figure three years back and the direction is unambiguous: the fraud types that require a human on a phone are the ones growing fastest, in a year when everyone assumed fraud was migrating to chat. Cheap synthetic speech removes the last real cost constraint on call-centre fraud — fluent, accent-matched, tireless operators. The IC3 recorded FBI-supported operations producing roughly 175 arrests across 13 joint actions with Indian law enforcement in 2025 against exactly this infrastructure.
Methodology
Data collection. All figures were retrieved from primary sources between September 10 and September 12, 2026: the FBI Internet Crime Complaint Center’s 2025 Internet Crime Report (published April 6, 2026); Federal Trade Commission press releases and Consumer Sentinel Network publications; the FTC’s April 2026 Data Spotlight on social media fraud; the FCC’s February 2024 Declaratory Ruling; and Barrington, Cooper & Farid, Scientific Reports 15:11004 (2025). No figure on this page originates from secondary aggregation.
Voice Clone Exposure Index formula. VCEI is computed for each IC3 crime type in which the 2025 report’s own artificial-intelligence section documents voice cloning or voice spoofing as an attack technique. Two components, both derived from published IC3 counts:
- Loss density (L) = AI-nexus losses ÷ AI-nexus complaints for that crime type, normalised against the highest value in the set (Business Email Compromise, $224,123) × 100.
- Penetration (C) = AI-nexus complaints ÷ total complaints for that crime type, normalised against the highest value in the set (Investment, 5.968%) × 100.
VCEI = (0.60 × L) + (0.40 × C)
Loss density carries the heavier weight because dollar exposure per successful attack is what determines whether a technique is worth an attacker’s marginal effort; penetration is weighted lower because it measures adoption to date rather than potential. Worked example for Business Email Compromise: L = ($224,123 ÷ $224,123) × 100 = 100.0; C = (0.545% ÷ 5.968%) × 100 = 9.13; VCEI = (0.60 × 100.0) + (0.40 × 9.13) = 63.7.
Scope of the index. VCEI scores exposure to synthetic-speech attack, not the prevalence of voice cloning. The IC3 AI descriptor covers all artificial-intelligence references, including text generation and synthetic video, so a crime type’s AI-nexus figures include non-voice techniques. The index is therefore an upper-bound ranking of exposure, and it is consistent across crime types because the same descriptor governs all of them. Any change to components or weights will be disclosed on this page with a re-baselined historical series.
Reporting caveat carried through all FTC figures. The FTC states that reports were not collected during the 2025 government shutdown, and that research indicates roughly 5% of people who experience mass-market consumer fraud report it to a government entity or Better Business Bureau.
Axis-calculated figures. Loss-per-complaint values, cohort shares, channel shares and all VCEI scores are Axis Intelligence Research calculations from the published counts above, marked axis_calculated = yes in the accompanying dataset.
About This Dataset
The complete fact table behind this page — every figure in the prose and every chart datapoint, with source organisation, source document, URL, retrieval date and primary-source flag — ships as voice-cloning-scam-statistics.csv under a Creative Commons Attribution 4.0 licence. Reuse is permitted with attribution.
Contents: 2025 AI-nexus fraud complaint counts and losses by crime type, age cohort and cryptocurrency nexus; FTC contact-method and imposter-scam series; peer-reviewed voice-clone detection accuracy measurements; and the full VCEI component and score table.
Citation
APA: Axis Intelligence Research. (2026). Voice cloning scam statistics 2026: Losses, complaint data and the fraud types most exposed to synthetic speech. Axis Intelligence.
MLA: Axis Intelligence Research. “Voice Cloning Scam Statistics 2026: Losses, Complaint Data and the Fraud Types Most Exposed to Synthetic Speech.” Axis Intelligence, 2026.
Chicago: Axis Intelligence Research. “Voice Cloning Scam Statistics 2026: Losses, Complaint Data and the Fraud Types Most Exposed to Synthetic Speech.” Axis Intelligence, 2026.
Frequently Asked Questions
Does the FBI track voice cloning as a separate crime type?
No. The IC3 uses an AI Related descriptor, not a crime type — it is applied after a crime type is selected, and it records any artificial-intelligence reference in the complaint. In 2025 it covered 22,364 complaints and $893,346,472 in losses. The closest voice-specific figure the FBI publishes is losses over $5 million to distress scams within the confidence and romance category.
Why is the official voice-cloning loss figure so much smaller than press coverage suggests?
Because the descriptor requires the victim to have recognised AI involvement. A cloned voice that works leaves no evidence it was cloned. Federal counts therefore measure detected synthetic speech, which is a subset of used synthetic speech — and the Berkeley detection data suggests that subset is well under half.
Which types of fraud are most exposed to voice cloning?
On the Voice Clone Exposure Index, investment fraud (78.8) and business email compromise (63.7) score highest, ahead of distress and romance scams (26.3), employment fraud (23.6), tech support (17.1), government impersonation (12.6) and phishing (6.2). BEC has the highest loss per incident at $224,123; investment fraud has the widest AI penetration at 5.97% of complaints.
How accurately can a person detect a cloned voice on a call?
In a 2025 Scientific Reports study of 604 listeners, participants correctly flagged an AI-generated clip as synthetic 60.8% of the time, and 21.0% of participants performed at or below chance. When asked whether a real speaker and their clone were the same person, listeners said yes on a median 83.3% of trials.
Does keeping the caller talking help?
Yes, measurably. In the same study, median detection accuracy was 56.7% on scripted speech, 76.7% on unscripted speech and 83.3% on combined clips where the speaker improvised. Asking open-ended questions the attacker cannot have scripted is the single highest-yield consumer-side control in the published evidence.
Is using a cloned voice in a phone call illegal in the United States?
For robocalls, yes. The FCC’s February 8, 2024 Declaratory Ruling classified AI-generated voices as “artificial” under the Telephone Consumer Protection Act, making unconsented AI-voice robocalls unlawful and giving state attorneys general enforcement authority. Fraudulent impersonation is separately actionable under the FTC’s Impersonation Rule, which has produced over $70 million in consumer redress across a dozen actions.
Who loses the most money to AI-enabled fraud?
People aged 60 and over. They filed 14.1% of AI-related complaints in 2025 but absorbed 39.5% of the losses — $112,153 per complaint, 2.81 times the all-cohort average. Across all fraud types, that cohort reported $7.748 billion in losses, up 59% year over year, with 12,444 individuals losing more than $100,000 each.
Why do so many AI fraud losses involve cryptocurrency?
Because irreversibility is the point. 83.0% of AI-related losses in 2025 involved cryptocurrency ($741,639,787), against only 39.0% of AI-related complaints. A synthetic voice buys the attacker authorisation; a crypto rail removes the recovery window that would otherwise undo it. The two techniques are complementary, not coincidental.
What verification control actually stops a voice-cloned payment request?
Callback on a known-good number, and a second authorisation that does not travel over the same channel as the request. The FTC’s standing guidance for family-emergency calls is to hang up and dial the person directly on a number you already have. For corporate payments, the equivalent is an out-of-band approval step — the control has to be structural, because the AI-BEC loss density of $224,123 per incident is what a single unchecked voice authorisation costs.
Where should a voice cloning scam be reported?
To the FBI at IC3.gov and to the FTC at ReportFraud.ftc.gov. Both feed the datasets this page is built from, and IC3’s Recovery Asset Team froze $679 million of $1.16 billion in attempted theft across 3,900 incidents in 2025 — a 58% success rate that depends entirely on reports arriving fast.