Deepfake Statistics 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: October 22, 2026 | License: CC BY 4.0
Reports to the US child-protection clearinghouse involving generative AI reached at least 400,000 in 2025, up from 67,000 in 2024. Axis Intelligence Research calculates that is a 5.97-fold rise, or one report every 78.8 seconds, while platforms themselves flagged only 6.96% of the AI-generated abuse files NCMEC later identified.
Quick Answer
The National Center for Missing & Exploited Children (NCMEC) received at least 400,000 generative-AI-related reports in 2025, roughly six times the 2024 count. On the law side, Axis Intelligence Research’s new Deepfake Legal Readiness Index (DLRI) scores seven jurisdictions out of 100: South Korea leads at 75, the seven-jurisdiction average is 40.0, and five of seven have no offence for simply creating a sexual deepfake.
Key Findings
- Axis Intelligence Research calculates that NCMEC’s generative-AI reports grew 5.97-fold between 2024 and 2025, from 67,000 to at least 400,000.
- Online platforms flagged about 11,000 files as AI-generated between 2023 and 2025, against 158,000 NCMEC categorized, a 6.96% provider flag rate (Axis Intelligence Research calculation).
- South Korea scores 75 out of 100 on the Axis Intelligence Research Deepfake Legal Readiness Index (DLRI), the highest of seven jurisdictions assessed on October 22, 2026.
- In a September 2026 test, 15 of 17 AI-generated files carried provenance signals but platforms labelled 61% of uploads, a 27.2-point provenance-to-label gap (Axis Intelligence Research).
- NCMEC’s 2025 average takedown time after a notice was 2.6 days, which Axis Intelligence Research converts to 62.4 hours, 14.4 hours beyond the TAKE IT DOWN Act’s 48-hour clock.
axis-intelligence.com/deepfake-statistics/
How Many Deepfake Child Abuse Reports Does NCMEC Receive?
NCMEC’s CyberTipline logged at least 400,000 reports with a generative AI (GAI) nexus in 2025, according to NCMEC’s generative AI data. A “GAI nexus” means the report involved generative AI in some way: generated imagery, AI-altered imagery, prompts, or abuse material found in model training data.
NCMEC generative-AI reports by year
| Year | GAI-nexus reports | Change | Source |
|---|---|---|---|
| 2023 | 4,700 | — | NCMEC |
| 2024 | 67,000 | — | NCMEC |
| 2025 | 400,000+ | 5.97× vs 2024; 85.1× vs 2023 | NCMEC; multiples: Axis Intelligence Research |
The 2025 total sits inside a CyberTipline that received 21.3 million reports overall, per NCMEC’s 2025 CyberTipline data. Axis Intelligence Research calculates generative AI’s share of all reports at 1.88% (400,000 ÷ 21,300,000). Small share, steep slope.
What do the 2025 AI reports actually describe?
NCMEC split its 2025 GAI reports into five named categories plus a residual bucket. All figures are lower bounds (“more than”) in NCMEC’s own wording.
| Category (2025) | Reports | Source |
|---|---|---|
| Using GAI to alter an existing abuse file, without a text prompt | 145,000+ | NCMEC |
| Attempting to generate abuse imagery (file upload plus text prompt) | 30,000+ | NCMEC |
| Abuse material found in AI training data | 12,000+ | NCMEC |
| Generating or possessing GAI abuse imagery | 7,000+ | NCMEC |
| Other GAI use, including chat-based exploitation | 3,000+ | NCMEC |
| GAI nexus, insufficient information on how it was used | 200,000+ | NCMEC |
The five named categories add up to at least 197,000 reports. Of those, Axis Intelligence Research calculates that 73.6% (145,000 ÷ 197,000) describe alteration of material that already existed, not text-to-image creation from scratch. Roughly half of all 2025 GAI reports (50.0%, 200,000 ÷ 400,000) carried too little detail to classify at all.
NCMEC’s data page adds a second cut: at least 182,000 reports in 2025 involved someone possessing, generating or attempting to generate GAI abuse imagery, which Axis Intelligence Research puts at 45.5% of the GAI total. NCMEC also counts more than 275 child victims identified in GAI imagery since 2023.
Analyst note — Sarah Mitchell. The headline multiple is real, but read the category table before quoting it. The biggest bucket is not someone typing a prompt into an image model; it is an existing abuse file run through an editing tool. That matters for anyone building safety filters: a classifier tuned on prompt text sees nothing when there is no prompt. And with 12,000+ reports tied to training data, part of the 2025 jump is the industry auditing its own datasets, which counts as a report but not as new abuse. The six-fold figure measures detection and disclosure as much as offending. Both are rising; only one is bad news.
Who catches AI-generated abuse files: platforms or NCMEC?
Between January 2023 and December 2025, electronic service providers flagged just over 11,000 files as AI-generated in their reports, while NCMEC’s own analysts categorized 158,000 AI-generated images and videos, per NCMEC. Axis Intelligence Research calculates a provider flag rate of 6.96% (11,000 ÷ 158,000): NCMEC identified 14.4 times as many AI files as the platforms that submitted them.
| Measure (Jan 2023–Dec 2025) | Files | Source |
|---|---|---|
| Flagged as AI-generated by providers | 11,000+ | NCMEC |
| Categorized as AI-generated by NCMEC | 158,000+ | NCMEC |
| Provider flag rate | 6.96% | Axis Intelligence Research (11,000 ÷ 158,000) |
How much AI child abuse video is there?
The Internet Watch Foundation (IWF) verified 8,029 realistic AI-generated child sexual abuse images and videos in 2025, up 14% on 2024, according to the IWF’s March 2026 release. Video is where the curve broke: 3,443 AI videos in 2025 against 13 in 2024, a 264.8-fold rise by Axis Intelligence Research’s calculation.
Severity moved with volume. The IWF classed 65% of AI videos as Category A, the most severe category, compared with 43% of non-AI criminal videos. Axis Intelligence Research puts that at 1.51 times the severe-content share. In the same release, a Savanta poll of 2,204 UK adults found 82% backed rules requiring AI tools to be safe by design.
Analyst note — Sarah Mitchell. Thirteen videos to 3,443 is the gap between “video models can’t hold a face for ten seconds” and “they can.” The IWF’s image count grew 14%; its video count grew more than two-hundredfold. Capability shipped first; the abuse followed within a single release cycle. Anyone forecasting 2026 should model on video, not stills.
How Many Sexualized Deepfakes Did Grok Create on X?
The Center for Countering Digital Harm (CCDH) estimated that Grok produced 3,002,712 sexualized images on X between December 29, 2025 and January 8, 2026, including an estimated 23,338 sexualized images of children, according to CCDH’s January 2026 study.
| CCDH measure (Dec 29, 2025–Jan 8, 2026) | Value | Source |
|---|---|---|
| Grok image posts in period | 4,621,335 | CCDH |
| Random sample analysed | 20,000 posts | CCDH |
| Share of sample that was photorealistic sexualized imagery | 65% | CCDH |
| Estimated sexualized images (95% interval 2,959,272–3,045,460) | 3,002,712 | CCDH |
| Estimated sexualized images of children (95% interval 17,099–30,039) | 23,338 | CCDH |
| Additional non-photorealistic (cartoon) sexualized child images | 9,936 | CCDH |
| Sampled child images still public on January 15, 2026 | 29% (29 of 101) | CCDH |
CCDH’s own rate is 190 sexualized images per minute. Axis Intelligence Research converts the 11-day estimate into 272,974 images per day and one sexualized image of a child every 40.7 seconds (950,400 seconds ÷ 23,338).
Regulators opened formal cases. The European Commission opened a Digital Services Act investigation into Grok and X’s recommender systems on January 26, 2026. In the UK, Ofcom’s October 14, 2026 update says its investigation, opened in January 2026, is moving toward a compliance decision, and flags the standalone Grok service as a separate gap.
Do Social Media Platforms Label Deepfakes?
Mostly not, and not consistently. A Maastricht University and University of St. Gallen audit collected 10,722 posts on Instagram, TikTok, X and YouTube and found 43 deepfakes in systemic-risk contexts such as elections and public health; platforms had labelled 14 of them, or 33%, per the preprint published on arXiv. The paper has not yet been peer reviewed. Axis Intelligence Research calculates that 67.4% of those risk-context deepfakes ran unlabelled. Their median reach was 160,000 views.
AI label rates by platform
| Platform | Controlled uploads labelled (Sept 2026) | Share of full dataset carrying an AI label | Source |
|---|---|---|---|
| 82% (14 of 17) | 34% | Maastricht/St. Gallen preprint | |
| YouTube | 63% (5 of 8) | 12% | Maastricht/St. Gallen preprint |
| TikTok | 59% (10 of 17) | 21% | Maastricht/St. Gallen preprint |
| X | 41% (7 of 17) | 5% | Maastricht/St. Gallen preprint |
| All four | 61% (36 of 59) | — | Maastricht/St. Gallen preprint |
The provenance-to-label gap
The researchers also checked the files they uploaded: 15 of 17 AI-generated test files (88.2%) already carried machine-readable provenance signals before upload. Platforms labelled 61.0% of uploads. Axis Intelligence Research calls the difference the provenance-to-label gap: 27.2 percentage points ((15 ÷ 17 − 36 ÷ 59) × 100) of content that announced its own synthetic origin and still reached feeds without a label.
Analyst note — Sarah Mitchell. The industry pitch for provenance standards was that detection is hard, so let generators sign their output and let platforms read the signature. This audit tested the easy half of that promise, and the easy half is still failing. Detection accuracy is a model problem; reading metadata that is already attached is a pipeline problem. The 27-point gap is a pipeline problem, which means it is fixable, and it means Article 50 audits will start there.
Does a deepfake label stop people believing the video?
Not reliably. In a peer-reviewed University of Bristol study, 53.8% of participants who were warned that a video was a deepfake, and who said they believed the warning, still relied on the video to judge the person in it guilty, against 12.0% in the control group, per Clark and Lewandowsky in Communications Psychology (2026). Axis Intelligence Research puts that at 4.5 times the control rate. The study ran three preregistered experiments with 175, 275 and 223 participants; in the second, a specific warning still left 47.1% judging guilt from the video, against 10.4% in the control group.
| Study measure | Value | Source |
|---|---|---|
| Warned, accepted the warning, still relied on the video (Exp. 1, 28 of 52) | 53.8% | Clark & Lewandowsky, 2026 |
| Control group judging guilt from the video (Exp. 1, 12 of 100) | 12.0% | Clark & Lewandowsky, 2026 |
| Specific warning, judged guilty from the video (Exp. 2, 16 of 34) | 47.1% | Clark & Lewandowsky, 2026 |
| Control group (Exp. 2, 5 of 48) | 10.4% | Clark & Lewandowsky, 2026 |
| Human deepfake detection accuracy, all modalities (56 papers, 86,155 participants) | 55.54% | Diel et al., 2024 |
| Human detection accuracy for video / image / audio / text | 57.31% / 53.16% / 62.08% / 52.00% | Diel et al., 2024 |
| Accuracy with training, AI support or caricaturization | 65.14% | Diel et al., 2024 |
Without labels, people do little better than a coin toss. The most recent meta-analysis of human deepfake detection, published in November 2024 and still the largest synthesis available, pooled 56 papers and found 55.54% overall accuracy, per Diel et al. in Computers in Human Behavior Reports.
Analyst note — Sarah Mitchell. Put the two studies next to the Maastricht audit and the policy chain looks thin at every link. Humans can’t spot the fake unaided; platforms label about six in ten even when the file announces itself; and a label that does land changes the verdict for only part of the audience. Labelling is necessary for Article 50 compliance. It is not, on this evidence, a harm-reduction strategy on its own: removal and friction on sharing have to carry the rest.
Labelling stops being voluntary in the EU this year. Article 50 of the EU AI Act, which requires deployers to disclose deepfakes and providers to mark synthetic output, applies from August 2, 2026, with fines up to €15 million or 3% of worldwide annual turnover, per the European Commission’s Article 50 fact page. Systems already on the market before that date have until December 2026 to meet the marking requirement. For the wider compliance calendar, see our EU AI Act compliance tracker.
How Fast Must Platforms Remove Deepfakes?
Removal deadlines now range from 2 hours in India to 48 hours in the United States.
| Jurisdiction | Statutory removal clock | Scope | Source |
|---|---|---|---|
| India | 2 hours | Complaints about intimate imagery, including artificially morphed images (Rule 3(2)(b)) | MeitY, IT Rules amendment, Feb 10, 2026 |
| India | 3 hours | Lawful government or court orders (Rule 3(1)(d)) | MeitY |
| Australia | 24 hours | eSafety removal notices, intimate images including deepfakes | eSafety Commissioner |
| United States | 48 hours | Valid requests to remove non-consensual intimate imagery, including AI “digital forgeries” | FTC business guidance |
The US clock has teeth as of this year. The FTC began enforcing the TAKE IT DOWN Act on May 19, 2026, sending compliance letters to 15 major platforms. The next day it sent warning letters to 12 “nudify” companies; each violation can carry a civil penalty of up to $53,088.
How do real takedown times compare with the legal clocks?
NCMEC’s 2025 data gives the only large observed benchmark: providers took 2.6 days on average to remove images or videos after an NCMEC notice, across more than 172,000 notices, per NCMEC. Axis Intelligence Research converts that to 62.4 hours: 14.4 hours beyond the US 48-hour clock and 31.2 times India’s 2-hour window.
This is a benchmark, not a compliance finding. NCMEC’s notices concern child sexual abuse material and pre-date federal enforcement; the TAKE IT DOWN clock covers adult and child intimate imagery from valid requests. The comparison shows the distance between what platforms were doing and what the newest laws demand.
Victims also use NCMEC’s own tool: Take It Down received more than 130,000 submissions covering more than 273,000 images and videos, about 2.1 files per submission by Axis Intelligence Research’s calculation.
Analyst note — Sarah Mitchell. A 2-hour clock is only meetable with automated triage: hash-matching, classifier pre-screening, then a human. India has written a latency requirement into law that most trust-and-safety stacks were never engineered for. Expect platforms to over-remove in India first, then argue about it.
Which Countries Have the Strongest Deepfake Laws? The Deepfake Legal Readiness Index (DLRI)
The Deepfake Legal Readiness Index (DLRI) is an Axis Intelligence Research score from 0 to 100 that measures how far a jurisdiction’s law in force reaches across the deepfake abuse chain: sharing, creating, possessing, removing and labelling. It scores statutes and binding rules, not enforcement outcomes.
DLRI ranking, as of October 22, 2026
| Rank | Jurisdiction | DLRI | C1 Sharing (20) | C2 Creation (20) | C3 Possession/viewing (10) | C4 Removal clock (25) | C5 Mandatory labelling (25) | Primary legal basis |
|---|---|---|---|---|---|---|---|---|
| 1 | South Korea | 75 | 20 | 20 | 10 | 0 | 25 | Sexual violence law amendment (Oct 2024, reported by Korea Times); AI Basic Act, in force Jan 22, 2026 (MSIT) |
| 2 | India | 50 | 0 | 0 | 0 | 25 | 25 | IT Rules amendment, G.S.R. 120(E) |
| 3 | United Kingdom | 40 | 20 | 20 | 0 | 0 | 0 | Sexual Offences Act 2003 s.66B; creation offence commenced by SI 2026/31 |
| 4 | Australia | 35 | 20 | 0 | 0 | 15 | 0 | Criminal Code Amendment (Deepfake Sexual Material) Act 2024; eSafety |
| 5 | United States (federal) | 30 | 20 | 0 | 0 | 10 | 0 | TAKE IT DOWN Act, Pub. L. 119-12 |
| 6= | China | 25 | 0 | 0 | 0 | 0 | 25 | CAC labelling measures, effective Sept 1, 2025 |
| 6= | European Union (EU-level law) | 25 | 0 | 0 | 0 | 0 | 25 | AI Act Article 50 |
The seven-jurisdiction average is 40.0. Five of the seven (India, Australia, the US federal level, China and the EU) score zero on C2: no offence that expressly targets creating a sexual deepfake, as opposed to sharing one.
How the DLRI is scored
Each component is binary except C4, and every point traces to a statute or binding rule listed in the dataset:
- C1 — Sharing offence (20 points): a criminal offence for distributing sexual imagery that expressly covers synthetic or AI-altered content.
- C2 — Creation offence (20 points): a criminal offence for making such imagery, whether or not it is shared.
- C3 — Possession or viewing offence (10 points): a criminal offence for possessing, buying, storing or viewing it.
- C4 — Statutory removal clock (25 points): ≤3 hours = 25; ≤24 hours = 15; ≤48 hours = 10; none = 0.
- C5 — Mandatory AI labelling in force (25 points): a binding obligation to label or mark AI-generated content that applies on the snapshot date.
Removal and labelling carry half the weight because they act at platform scale; criminal offences act case by case. A different weighting would reorder the middle of the table, but South Korea stays first under any weighting that gives the three offence components, together, at least as much weight as the removal clock, because it is the only jurisdiction scoring on all three.
What the DLRI scores tell you, and what they don’t
South Korea’s 75 reflects the October 2024 amendment that, per Korea Times, made possessing, purchasing, storing or viewing sexual deepfakes punishable by up to 3 years in prison, with up to 7 years for producing or distributing them. The US federal 30 rests on the TAKE IT DOWN Act, approved May 19, 2025, which makes knowingly publishing a sexual digital forgery punishable by up to 2 years in prison when the victim is an adult and 3 years when the victim is a minor; it does not criminalize creating one. India’s 50 comes entirely from platform rules; its offences rely on general provisions that do not name synthetic imagery, so C1–C3 stay at zero. The EU scores 25 because Directive (EU) 2024/1385, which requires member states to criminalize sharing sexual deepfakes, has a transposition deadline of June 14, 2027, as noted in a Bundestag transposition paper. National laws in individual member states are outside this EU-level reading.
The index does not score enforcement. A score of 25 on labelling says a rule is in force; the Maastricht audit above says how well it is followed.
Analyst note — Sarah Mitchell. The pattern in the table is that governments legislated where the harm became visible: distribution. Creation went largely untouched until open-weight image models put a nudify pipeline on a consumer GPU, at which point the shared file stopped being the first step. The UK moved in February 2026; South Korea in 2024. The other five still treat a locally generated deepfake as legal until it leaves the device. That is a law written for the previous generation of models.
Track every bill and rule change on our AI regulation tracker.
How Many US States Have Election Deepfake Laws?
Thirty-one US states have enacted laws regulating deepfakes in elections and political communications, 27 of them through a disclosure requirement, according to the National Conference of State Legislatures (NCSL), updated October 13, 2026. California, Minnesota and Texas use prohibitions rather than disclosure alone.
Courts have started removing laws from that count. Hawaii’s statute was found unconstitutional in January 2026, and Montana’s and Vermont’s were preliminarily enjoined in September 2026, per NCSL. Axis Intelligence Research therefore counts 28 states with an election-deepfake law that is not blocked in court, leaving 22 states without one heading into the November 2026 midterms. California is kept in the count: one provision (§20012) was enjoined in 2025, but its other statutes stand.
| Measure (as of Oct 13, 2026) | States | Source |
|---|---|---|
| Enacted election-deepfake laws | 31 | NCSL |
| Using a disclosure model | 27 | NCSL |
| Only statute blocked in court (HI, MT, VT) | 3 | Axis Intelligence Research, from NCSL |
| Enacted and not blocked | 28 | Axis Intelligence Research (31 − 3) |
| No unblocked law | 22 | Axis Intelligence Research (50 − 28) |
Analyst note — Sarah Mitchell. Disclosure laws survive court better than bans, which is why 27 states chose them. But a disclosure law assumes the person who made the deepfake will label it, and the Maastricht numbers show platforms don’t reliably label even the content that is already marked. For election integrity, the weak link is no longer the statute; it is the upload pipeline.
Deepfake Sex Crime Statistics in South Korea
South Korea publishes the most granular national deepfake crime data we found, from three government bodies: the police, the gender ministry’s victim support center and the women’s policy institute. Each counts something different, and the three counts move in the same direction.
Police crackdowns: deepfakes became the largest category
In the November 2024–October 2025 cyber sexual violence crackdown, deepfake (“false video”) crimes made up 35.2% of cases, the largest single category, according to the Korean National Police Agency (KNPA) results published on the Korea Policy Briefing portal on November 25, 2025. Reported cases rose 35%, from 3,270 to 4,413; police solved 3,411 cases and arrested 3,557 suspects.
A dedicated deepfake crackdown ran in parallel. Between August 28, 2024 and March 31, 2025, police arrested 963 people for deepfake sex crimes and detained 59, a 260% rise in arrests on the previous crackdown, per the KNPA release of April 17, 2025. Teenagers were 669 of the 963, including 72 children below the age of criminal responsibility; Axis Intelligence Research puts the teen share at 69.5%. Police also filed 10,535 requests to delete or block the videos.
| Measure | Value | Period | Source |
|---|---|---|---|
| Cyber sex crimes reported | 4,413 (vs 3,270, +35%) | Nov 2024–Oct 2025 | KNPA via korea.kr |
| Deepfake share of cases (largest category) | 35.2% | Nov 2024–Oct 2025 | KNPA via korea.kr |
| Cases solved / suspects arrested | 3,411 / 3,557 | Nov 2024–Oct 2025 | KNPA via korea.kr |
| Deepfake crackdown arrests (detained) | 963 (59) | Aug 28, 2024–Mar 31, 2025 | KNPA via korea.kr |
| Teen share of deepfake crackdown arrests | 69.5% (669 of 963) | Aug 28, 2024–Mar 31, 2025 | KNPA; share: Axis Intelligence Research |
| Deepfake cases cleared, earlier crackdown | 48 of 1,612 (3%) | Mar–Oct 2022 | KNPA press release, Nov 16, 2022 |
| Cyber sex crime cases / suspects, current crackdown (interim) | 1,446 / 1,506 (87 detained) | Nov 2025–Jun 2026 | KNPA, via Korea JoongAng Daily |
In the March–October 2022 crackdown, deepfakes were 48 of 1,612 cases cleared, which the KNPA put at 3%, per its own November 16, 2022 release. The 2022 figure counts cases cleared and the 2025 figure counts cases reported, so the rise from 3% to 35.2% is directional, not an exact multiple.
The newest data point runs the other way. Interim results for the crackdown that began in November 2025, released June 16, 2026, counted 1,446 cases and 1,506 suspects, and the KNPA said deepfake sex crimes had declined compared with the same period a year earlier, as reported by Korea JoongAng Daily and Hankyung. The police did not publish the deepfake count itself, so the size of the drop is unknown.
Victim support: 1,616 deepfake cases in 2025, almost all against women
The government’s central Digital Sex Crime Victim Support Center supported 10,637 victims in 2025, according to the Ministry of Gender Equality and Family, and 91.2% of victims of synthesized or edited imagery (the center’s deepfake category) were in their teens or twenties. Newsis reported the category’s 2025 cases as 1,581 involving women and 35 involving men, up 16.8% on 2024.
Axis Intelligence Research adds the two to 1,616 cases, 97.8% of them involving women. The figure reconciles with the center’s 2024 count of 1,384 synthesis-and-editing cases, published by the Korean Institute of Criminology and Justice: 1,616 ÷ 1,384 is a 16.8% rise, the same increase the ministry reported.
| Measure | Value | Period | Source |
|---|---|---|---|
| Digital sex crime victims supported | 10,637 (+3.2%) | 2025 | MOGEF; change via Newsis |
| Synthesis/editing (deepfake) cases | 1,384 → 1,616 (+16.8%) | 2024 → 2025 | KICJ (2024); Newsis (2025, 1,581 + 35); sum: Axis Intelligence Research |
| Share of synthesis/editing cases involving women | 97.8% | 2025 | Axis Intelligence Research (1,581 ÷ 1,616) |
| Synthesis/editing victims aged 10–29 | 91.2% | 2025 | MOGEF |
Long-run trend: 17.7 times more deepfake sex crimes than in 2020
The Korean Women’s Development Institute (KWDI) counted 31 deepfake sex crimes in 2020 and 550 in 2024, against 15,612 digital sex crimes in 2024, in its 2026 Gender Statistics of Korea, as reported by The Korea Times. Axis Intelligence Research calculates a 17.7-fold rise and a 3.5% share of digital sex crimes in 2024.
These series cannot be merged. KWDI counts crimes for a calendar year; the police count crackdown cases in their own windows; the support center counts victims who asked for help. Read the police figures as enforcement output, the support-center figures as harm reported by victims, and the KWDI series as the longest trend line.
Analyst note — Sarah Mitchell. Korea is the closest thing to a natural experiment in deepfake law. It criminalized possession and viewing in late 2024, ran two crackdowns, and by mid-2026 the police were reporting a decline in deepfake cases while victim-support cases were still rising 16.8%. Both can be true: fewer offenders caught in new cases, more victims coming forward about old images. Watch the next annual police figure before calling it a turning point.
How Often Does a Deepfake Abuse Report Happen?
Axis Intelligence Research converted each annual count into an average interval, a dated snapshot that makes the scale legible. Units differ by source and the rows are not additive.
| Event (period) | Average interval | Calculation | Underlying source |
|---|---|---|---|
| Sexualized Grok image of a child on X (Dec 29, 2025–Jan 8, 2026) | every 40.7 seconds | 950,400 s ÷ 23,338 | CCDH |
| NCMEC report with a generative-AI nexus (2025) | every 78.8 seconds | 31,536,000 s ÷ 400,000 | NCMEC |
| NCMEC report of possessing, generating or attempting GAI abuse imagery (2025) | every 173.3 seconds | 31,536,000 s ÷ 182,000 | NCMEC |
| Take It Down submission (2025) | every 4.0 minutes | 31,536,000 s ÷ 130,000 | NCMEC |
| IWF-verified AI abuse image or video (2025) | every 65.5 minutes | 31,536,000 s ÷ 8,029 | IWF |
| IWF-verified AI abuse video (2025) | every 2.5 hours | 31,536,000 s ÷ 3,443 | IWF |
| Deepfake case recorded in the Korean police crackdown (Nov 2024–Oct 2025) | every 5.6 hours | 31,536,000 s ÷ 1,553 (= 35.2% of 4,413) | KNPA |
Because the NCMEC and IWF annual figures are lower bounds or verified counts, the real intervals are at most these values.
Methodology
Axis Intelligence Research assembled this page in October 2026. Every figure was read from the source document on retrieval date October 22, 2026 and logged as a row in deepfake-statistics.csv, with organization, document title, URL and a primary-source flag.
Source choice. We used producers of the data: NCMEC and IWF for abuse-imagery counts, CCDH for its Grok measurement, the Maastricht University/University of St. Gallen team for platform labelling, NCSL for state law, and the statute or regulator page for each law. South Korean police figures come from KNPA results published on the government’s Korea Policy Briefing portal (korea.kr) and from the KNPA’s original 2022 press release; victim-support figures come from the Ministry of Gender Equality and Family. Where a government page did not carry a figure (the 1,553 deepfake-case count, the 2025 synthesis/editing case split, the June 2026 interim results and the KWDI series), we used The Korea Herald, Newsis, Korea JoongAng Daily and The Korea Times, and marked those rows is_primary = no. The two behavioural studies (Clark and Lewandowsky; Diel et al.) are peer reviewed; the Diel meta-analysis dates from November 2024 and is used because no newer synthesis of human detection accuracy exists. Deepfake fraud losses are deliberately excluded and live on our deepfake fraud statistics page.
Calculated figures. Every row with axis_calculated = yes carries its formula in method_note. Growth multiples are later value ÷ earlier value. Shares are part ÷ whole × 100. Intervals use a 365-day year (31,536,000 seconds) or, for CCDH, the 11-day study window (950,400 seconds). Where NCMEC reports “more than” a figure, we used the stated floor and labelled the derived value as resting on lower bounds.
DLRI. Scores are the sum of five weighted components (C1 20, C2 20, C3 10, C4 25, C5 25), assessed against law in force on October 22, 2026. Rules adopted but not yet applicable score zero (the EU deepfake Directive, for example). A labelling rule with a penalty grace period still scores C5 if the obligation itself applies, as with South Korea’s AI Basic Act. EU scoring reflects EU-level instruments only. Each component’s points sit in the CSV under dlri_component_points, so any reader can recompute a score or re-weight it.
What we did not merge. The KNPA, victim-support-center and KWDI Korean series use different units and periods; we publish them side by side. The one cross-check we could run held: the support center’s 2025 count (1,616) over its 2024 count (1,384) gives the 16.8% rise the ministry reported. The Maastricht study is a preprint with small controlled samples (59 uploads, 17 files), so its platform percentages describe that test, not each platform’s global labelling rate.
About This Dataset
deepfake-statistics.csv holds 201 rows in long format, one observation per row. It covers generative-AI child-abuse reports (NCMEC, 2023–2025), AI abuse imagery verified by the IWF (2024–2025), CCDH’s Grok estimates (December 2025–January 2026), platform label rates from the Maastricht/St. Gallen audit (2026), South Korean deepfake crime, crackdown and victim-support counts (2020–2026, with a 2022 police baseline), two peer-reviewed studies on human deepfake detection and on warning labels, removal clocks and penalties in four jurisdictions, US state election-deepfake law counts (October 2026), and every DLRI reading and component for seven jurisdictions.
Columns: row_id, metric, value, unit, as_of_date, geography, segment, source_org, source_document, source_url, retrieved_date, is_primary, axis_calculated, method_note. Values are raw numbers; units sit in their own column; dates are ISO 8601.
License: Creative Commons Attribution 4.0 (CC BY 4.0). Credit line: Axis Intelligence Research, Deepfake Statistics 2026, 2026.
How to Cite This Page
APA: Axis Intelligence Research, & Mitchell, S. (2026, October 22). Deepfake statistics 2026: Abuse reports, platform labels, takedown clocks and the laws behind them. Axis Intelligence. https://axis-intelligence.com/deepfake-statistics/
MLA: Axis Intelligence Research, and Sarah Mitchell. “Deepfake Statistics 2026: Abuse Reports, Platform Labels, Takedown Clocks and the Laws Behind Them.” Axis Intelligence, 22 Oct. 2026, axis-intelligence.com/deepfake-statistics/.
Chicago: Axis Intelligence Research, and Sarah Mitchell. “Deepfake Statistics 2026: Abuse Reports, Platform Labels, Takedown Clocks and the Laws Behind Them.” Axis Intelligence, October 22, 2026. https://axis-intelligence.com/deepfake-statistics/.
Deepfake Questions People Ask
Does the TAKE IT DOWN Act cover AI-generated images of adults?
Yes. The Act covers “digital forgeries” of adults and minors: knowingly publishing one carries up to 2 years in prison for an adult victim and 3 years for a minor, per Public Law 119-12. Covered platforms must remove reported imagery within 48 hours of a valid request, and the FTC can seek civil penalties of up to $53,088 per violation since enforcement began on May 19, 2026.
Is it a crime to make a sexual deepfake if you never share it?
It depends on the country. In the UK, creating one became an offence from February 6, 2026; in South Korea, even viewing one can bring up to 3 years in prison. Five of seven jurisdictions on the Axis Intelligence Research DLRI, including the US at federal level, have no creation offence.
Why did NCMEC’s AI-related reports jump about six-fold in 2025?
More detection, not only more abuse. Of at least 197,000 categorized 2025 reports, 145,000+ involved altering existing abuse files and 12,000+ concerned abuse material found in AI training data, per NCMEC. Half of the at-least-400,000 total lacked enough detail to classify.
If a platform shows no “AI info” label, does that mean the image is real?
No. In a September 2026 audit, platforms labelled 61% of 59 AI-generated test uploads, even though 15 of 17 files carried provenance signals. X labelled 41% and Instagram 82%, per the Maastricht/St. Gallen preprint.
How quickly do platforms actually remove child abuse imagery after a report?
In 2025, providers took 2.6 days on average to remove images or videos after an NCMEC notice, across more than 172,000 notices. That is 62.4 hours by Axis Intelligence Research’s conversion, well above India’s 2-hour rule for intimate imagery complaints.
Which country has the toughest deepfake laws in 2026?
South Korea, with a score of 75 out of 100 on the Axis Intelligence Research Deepfake Legal Readiness Index as of October 22, 2026. It is the only jurisdiction assessed that criminalizes sharing, creating and possessing sexual deepfakes, and it has mandatory AI labelling under its AI Basic Act.
When does the EU require deepfakes to be labelled?
From August 2, 2026, under Article 50 of the AI Act, with fines up to €15 million or 3% of worldwide turnover. Systems already on the market have until December 2026 to add machine-readable marking, per the European Commission.
Are AI deepfake videos of child abuse more severe than other abuse videos?
On the IWF’s 2025 data, yes: 65% of AI-generated abuse videos were Category A, the most severe class, against 43% of non-AI criminal videos. AI video volume rose from 13 in 2024 to 3,443 in 2025.
Does labelling a video as a deepfake stop people from believing it?
Only partly. In a 2026 University of Bristol study, 53.8% of participants who were warned a video was a deepfake, and believed the warning, still used it to judge the person guilty, against 12.0% of a control group, per Communications Psychology.
Are US election deepfake laws holding up in court?
Partly. Of 31 states with election-deepfake laws, Hawaii’s was struck down and Montana’s and Vermont’s were preliminarily enjoined as of October 2026, per NCSL. Axis Intelligence Research counts 28 states with a law in effect and 22 without.
Related reading: Deepfake fraud statistics · AI scam statistics · Digital identity fraud statistics · Synthetic identity fraud statistics · AI in social media statistics · Social media regulation statistics · AI regulation tracker · EU AI Act compliance tracker