detection

Best Deepfake Detectors

Compare the best deepfake detectors as of September 2026 — Resemble Detect, Whispeak, Pindrop Pulse, Aurigin AI, GetReal Security, Sensity, Hive, DuckDuckGoose and Reality Defender — with independent benchmark accuracy and false-positive rates for video and audio, published pricing, and which to use for fraud defence, journalism or a one-off check.

Updated September 2026

Quick answer: For audio — the modality that actually costs organisations money — Resemble Detect leads on the only public benchmark where vendors cannot see the labels, at 99.47% accuracy with a 0.7% false-positive rate on Podonos’s August 2026 leaderboard. Whispeak is second at 97.70% with the most balanced error profile in the field (2.9% false positives, 1.7% false negatives), and Pindrop Pulse is the pick for contact centres at 95.05% and roughly 282 milliseconds per file, because a detector that is slower than the call is useless. For video and executive impersonation, GetReal Security and Reality Defender were the two startups named Market Shapers in Gartner’s 25 June 2026 Emerging Market Quadrant for deepfake detection — though Reality Defender’s one independent audio result is poor, and we set out why below. Two caveats govern everything on this page. On video, the best commercial detector managed 78% accuracy on Deepfake-Eval-2024, the largest in-the-wild benchmark, against roughly 90% for human forensic analysts. And false-positive rates across tested systems run from 0.7% to 53.7%, so the number that decides a purchase is not headline accuracy but how often a tool calls a real voice fake.

This page ranks the tools that detect deepfaked video and audio — a different problem, with different vendors and different economics, from AI-generated still images (best AI image detectors) and AI-written text (best AI detectors). Three things make this category unlike the other two. First, the buyer is usually an organisation defending against fraud rather than an individual checking a file, which is why most vendors sell APIs and enterprise contracts rather than $10 monthly plans. Second, real-time performance matters in a way it never does for a static file: one tested system needs 7.6 seconds to score a 5-second clip, which rules it out of any live use. Third, the independent evidence is thin, recent and unflattering, and it disagrees sharply with vendor marketing. If you are here because of a specific incident — a voice on a call, a video in a feed — read how to verify without a detector before you read the rankings.

The current state of deepfake detection: September 2026

Five shifts define where this market stands, and audio and video are moving in opposite directions.

1. Audio detection got genuinely good in 2026, and it is now measurable. Podonos has run the only public audio benchmark with private gold-standard labels, first published on 5 May 2026 with eight systems and expanded to sixteen by August 2026. It uses a fixed set of 4,524 clips at a 50/50 real-to-fake split, synthetic audio from around 25 modern text-to-speech systems including ElevenLabs, F5-TTS and Chatterbox, and six file formats — mp3, wav, flac, ogg, m4a and webm — because production audio does not arrive as clean 16 kHz wav. Critically, Podonos scores each system at the threshold the vendor actually ships, not the threshold that flatters it. On the August 2026 leaderboard, Resemble AI’s DETECT-World reached 99.47%, Whispeak 97.70%, Aurigin AI 96.75% and Pindrop 95.05%.

2. Video detection lags badly, and the gap to marketing is enormous. Deepfake-Eval-2024 assembled 45 hours of video, 56.5 hours of audio and 1,975 images of real deepfakes circulated in 2024, drawn from 88 websites in 52 languages. Against it, the best commercial detector reached 0.78 accuracy on video (AUC 0.79), 0.89 on audio (AUC 0.93) and 0.82 on images (AUC 0.90) — while human forensic analysts were estimated at roughly 90%. Open-source models fared far worse, with AUC falling by roughly 50% for video, 48% for audio and 45% for image models relative to the academic benchmarks they were tuned on. The paper’s own conclusion is that academic benchmarks are out of date and unrepresentative of real deepfakes.

3. Training recency, not architecture, is the variable that predicts performance. All four open-source models in Podonos’s first round had been trained on ASVspoof 2019 LA, a dataset predating every generator in the test set. They scored between 48% and 63% — barely better than random — and one, LCNN, simply predicted “fake” for every input. The single most useful procurement question in this category is when the model last saw new generators, not what it claims to score.

4. The attack side industrialised faster than the defence side. iProov’s 2026 Threat Intelligence Report, published 8 April 2026, recorded a 741% rise in iOS injection attacks across 2025, with a 1,151% increase in the second half alone, and a 720% spike in Southeast Asia during the third quarter as attackers tested virtual-camera techniques before scaling them elsewhere. Injection attacks bypass the camera entirely, feeding synthetic video straight into the verification pipeline — so a detector that only inspects delivered frames is watching the wrong layer. iProov also cites Ponemon Institute research finding 41% of organisations have experienced deepfake attacks targeting executives, and Gartner finding 37% of cybersecurity leaders encountered a deepfake during a video call.

5. Provenance arrived for video, but the schemes do not read each other. OpenAI attaches C2PA Content Credentials to Sora 2 assets and applies a visible moving watermark to downloads from the consumer app; Google applies the invisible SynthID watermark across Veo, and DeepMind reported in May 2026 that more than 100 billion images, videos and audio files have carried a SynthID mark since 2023. EU AI Act Article 50 became applicable on 2 August 2026, requiring machine-readable marking of generative output and visible labelling of deepfakes, with penalties up to €15 million or 3% of worldwide annual turnover. The catch is that a SynthID lookup returns nothing on a Sora video and a C2PA read returns nothing once a platform has stripped the metadata. See our AI watermarking guide for how the marks are built.

How deepfake detectors actually work

Four distinct technologies are sold as “deepfake detection”, and buying the wrong one is the most common and most expensive mistake in this category.

Media classifiers guess from the signal. A classifier is a model trained on genuine and synthetic media that returns a probability the file is fake. On video it looks for blending artefacts at the face boundary, temporal inconsistency between frames, unnatural blink and head-pose dynamics, and traces left by the generation pipeline. On audio it looks for vocoder artefacts, spectral discontinuities and unnatural prosody. This is what Resemble Detect, Whispeak, Pindrop, Aurigin, Sensity, Hive, DuckDuckGoose, Reality Defender and Deepware do. It works on any file from any source with no cooperation from the generator, and it degrades against generators released after its training data.

Biological-signal detectors look for evidence of a living body. Intel’s FakeCatcher is the best-known example: it uses photoplethysmography, reading minute colour shifts across 32 facial locations caused by blood flow, on the reasoning that a generated face has no pulse to render. Intel reports 96% accuracy. Announced by Intel Labs in November 2022, it remains research and demonstration technology rather than a product you can buy, and it needs enough clean facial resolution to see the signal at all.

Liveness and injection detection defend the capture path, not the file. iProov, Sumsub and Incode are not primarily asking “is this video fake?” — they are asking “did this video come from a real camera pointed at a real person right now?” They do that with challenge-response, controlled illumination and device attestation. Given that injection is the fastest-growing vector in iProov’s data, this is the correct control for identity verification and account recovery, and a media classifier is the wrong tool for that job.

Provenance readers check for a signature or watermark. C2PA Content Credentials attach cryptographically signed metadata describing what produced a file and how it was edited; SynthID embeds an invisible statistical watermark in the media itself. When a signal is present the answer is close to definitive. When it is absent you have learned nothing — most open-weight video and voice models add no mark, and metadata is routinely stripped when a file is re-encoded on upload. The absence of a Content Credential is not evidence that a video is real.

The practical consequence: a newsroom verifying a viral clip needs a classifier plus a provenance reader plus ordinary reporting; a bank stopping account takeover needs liveness and injection detection; a contact centre stopping voice fraud needs a fast audio classifier on the call path. These are three different purchases.

Top deepfake detectors compared (September 2026)

Two sets of numbers exist and they disagree sharply, so we show both. Vendor accuracy is measured on the vendor’s own test set at the vendor’s own threshold and is effectively a ceiling. Independent results come from Podonos and Deepfake-Eval-2024, the two public tests with published methodology.

What the vendors claim

None of these figures has been independently reproduced. Where a vendor publishes no number, we have said so rather than repeat a percentage circulating on review sites. Reality Defender and GetReal declining to publish a headline accuracy figure is arguably the more defensible position, given how little those figures predict.

What independent testing found

Flag the conflict. Resemble AI leads the Podonos benchmark and promotes the result heavily on its own site. The benchmark itself was run by Podonos with private labels and a public repository, which is a materially better design than a vendor self-test, but the amplification comes from the winner. Read the methodology, not the press release. Note also that the Podonos benchmark is audio-only: it says nothing about any vendor’s video performance, in either direction.

The best deepfake detectors, reviewed

1. Resemble Detect — best evidenced accuracy on audio

Best for: Organisations that need the highest measured audio accuracy, with on-premises deployment available.

Resemble Detect holds the top position on the only public audio benchmark with private labels. Its current DETECT-World model reached 99.47% accuracy with a 0.7% false-positive rate and a 0.4% false-negative rate at 399 milliseconds per file — the rare result that is strong in both error directions at once, which matters more than the headline number. Resemble ships Detect as an API with real-time operation, explainable per-segment output and on-premises deployment, and reports testing against more than 160 generative systems. Resemble also sells voice cloning, which cuts both ways: the detection team sees new synthesis techniques early, and the company profits from both sides of the problem. See our best AI voice cloning page for the generation side.

The caveat: this is one benchmark, on audio only, run by a company whose result Resemble markets aggressively.

2. Whispeak — most balanced error profile

Best for: Deployments where a false accusation costs as much as a missed fake.

Whispeak entered the Podonos benchmark at second place with 97.70% accuracy, 2.9% false positives and 1.7% false negatives — described in that write-up as the most balanced error profile in the field. In production that balance is often worth more than a percentage point of accuracy, because it means the tool is not systematically biased towards either wrongly flagging genuine callers or waving fakes through.

The caveat: a newer entrant with a thinner deployment record than Pindrop or Resemble, and no independent video result.

3. Pindrop Pulse — best for contact centres and live calls

Best for: Banks, insurers and any organisation with a phone channel where fraud arrives by voice.

Pindrop Pulse is built for a constraint others largely ignore: the answer has to arrive while the caller is still talking. On the Podonos benchmark it returned results in roughly 282 milliseconds per file at a real-time factor of 0.076, described there as the fastest accurate detector by a wide margin, while scoring 95.05% accuracy. Pindrop advertises 99.2% accuracy on its own analysis and 93% against previously unseen deepfakes, and pairs detection with device, behaviour and carrier-metadata signals rather than relying on audio alone — which matters, because in a real fraud call the voice is only one of several things that can be checked. Pindrop took a $100 million five-year loan from Hercules Capital in July 2024 and was named to TIME’s 2026 list of most influential software companies.

The caveat: its 6.2% false-positive rate is the highest among the strong performers, which at contact-centre volumes is a lot of wrongly-suspected genuine customers. It is also a contact-centre platform, not a file-checking tool.

4. Aurigin AI — lowest false positives among the early field

Best for: Teams that would rather miss a fake than wrongly flag a real customer.

Aurigin AI scored 96.75% accuracy in the May 2026 round with a 1.5% false-positive rate — the lowest of the four commercial systems tested at that point — against a 5.0% false-negative rate. That shape is the mirror image of Resemble’s early DETECT-3B result and suits any workflow where a flagged customer triggers expensive manual review.

The caveat: one in twenty fakes gets through at that threshold.

5. GetReal Security — best for executive impersonation and live meetings

Best for: Enterprises whose threat model is an impersonated executive on a video call.

GetReal Security was co-founded by Hany Farid, the UC Berkeley professor who is among the most-cited academics in image forensics, and it is the most explicitly security-shaped vendor in the category — combining multimodal detection with threat intelligence, attack-surface visibility and automated response rather than selling detection as a standalone score. On 18 May 2026 it made continuous identity verification generally available in GetReal Protect, so a participant is re-verified throughout a call rather than only at the point they join, citing more than $1.6 billion in deepfake-driven financial losses in 2025. It was named a Market Shaper in the 25 June 2026 Gartner quadrant, positioned furthest right on the potential-for-disruption axis, and raised a $17.5 million Series A led by Forgepoint Capital with participation from Capital One Ventures in March 2025.

The caveat: a $17.5 million Series A is early-stage, and GetReal has no result in either independent benchmark. This is a bet on the team and the research.

6. Sensity AI — best forensic report for investigations

Best for: Investigators, legal teams and platform trust-and-safety functions that need documentation, not just a score.

Sensity runs a multi-layer analysis across visual artefacts, voice spectral characteristics, file structure, container metadata and cross-modal inconsistency, and returns a forensic report rather than a bare probability — which is what you need if the output has to survive scrutiny. It also runs monitoring and takedown workflows, making it closer to a service than an API.

The caveat: Sensity does not publish a headline accuracy figure against a named test set. The 95–98% range widely attributed to it traces to review sites rather than to Sensity, and it appears in neither independent benchmark.

7. Hive AI — best video and image classifier sold as a metered API

Best for: Platforms and moderation pipelines that want detection as a simple per-unit API call.

Hive sells detection the way infrastructure is normally sold — per unit, self-serve, with a free browser demo — covering AI-generated and manipulated video, images and audio. Its strongest independent credential is on images: Hive was one of only two tools that never mislabelled an authentic news photograph in NewsGuard’s May 2026 audit, the best result any vendor holds on that side; see best AI image detectors for that test in full. Hive’s audio is weaker, at 83.53% on the Podonos benchmark — though notably with a low 2.4% false-positive rate, its failure mode being missed fakes (30.5%) rather than wrongly flagged real audio.

The caveat: the 30.5% audio false-negative rate rules it out as a voice-fraud control. Hive video is listed at $0.07 per second on third-party comparisons, but Hive publishes credit-based pricing rather than a video rate card, so confirm at quote.

8. DuckDuckGoose — European option for regulated deployments

Best for: Banks, government agencies and identity-verification providers in the EU that need on-premises deployment and explainability.

DuckDuckGoose is a Dutch vendor emphasising explainable output showing which regions of a frame drove a decision, which matters for both audit and dispute, with on-premises deployment for regulated customers.

The caveat: its 99% accuracy claim has no published basis and DuckDuckGoose appears in neither independent benchmark. Treat it as unevaluated.

9. Reality Defender — broadest coverage and the only real free API, with a serious caveat

Best for: Teams that need video, audio, image and text behind one API and want to evaluate before committing.

Reality Defender covers four modalities with an ensemble of models per modality, offers hosted SaaS, private cloud and on-premises deployment, and on 31 July 2025 opened a public API with a free tier of 50 audio or image scans per month — unusual in a market that otherwise routes everyone through sales. It was named a Market Shaper in the 25 June 2026 Gartner quadrant.

The caveat, and it is a large one: on the Podonos audio benchmark Reality Defender scored 71.27% accuracy with a 53.7% false-positive rate — it flagged more than half of all genuine audio as fake — alongside a 17.2% rejection rate on clips under about 1.5 seconds and 5,718 milliseconds per file, making it the only system tested that runs slower than real time. A 5-second clip takes roughly 7.6 seconds to score. On that evidence it should not be deployed on a live voice channel or any high-volume authentication flow. The result is audio-only and says nothing about Reality Defender’s video or image models, and Reality Defender publicly disputes the value of lab benchmarks as a predictor of production performance. We report it because it is the only independent measurement that exists, and because a false-positive rate of that magnitude is the single most consequential number a buyer could be missing.

10. Free and low-cost tools worth using

These are the right choice for a one-off check on a single file, and the wrong choice for anything with money or a reputation attached.

• DeepFake-o-Meter — free, from the University at Buffalo Media Forensics Lab under Siwei Lyu. It runs several published research detectors across image, video and audio and shows each one’s probability separately, so you see disagreement between models rather than a single confident number. That transparency makes it the most honest free tool available, and it is the one journalists have used on cases from the Biden robocall to Zelensky footage. It is an academic platform, so throughput and uptime are not guaranteed.
• Deepware Scanner — free web upload for video, returning a probability in minutes, oriented towards face-swap manipulation of real footage rather than fully generated video. See best AI face swap apps for the generation side, and what the May 2026 safety audit found about it.
• Reality Defender free tier — 50 audio or image scans per month. Useful for evaluation; read the caveat above before trusting an audio verdict.
• Hiya Deepfake Voice Detector — free Chrome extension, launched October 2024, that checks audio playing in the browser and needs about one second of speech. Hiya reports over 99% accuracy on in-the-wild data and placed top-tier on the Hugging Face Speech Deepfake Arena in February 2026, but that leaderboard is community-run on public data.
• ElevenLabs Audio Detector — free, and the right tool for one narrow question: did this audio come from ElevenLabs? The current detector combines a SynthID watermark check with a classifier. The older AI Speech Classifier is a legacy statistical tool that only works reliably on output from ElevenLabs’ earlier models and does not reliably classify Eleven v3. Neither tells you anything about audio from another generator.
• Attestiv — five free video scans a month, aimed at insurance claims and document verification rather than media analysis.

Identity-verification specialists: a different purchase

If your problem is someone impersonating a customer during onboarding, login or account recovery, the vendors above are largely the wrong shape. iProov, Sumsub and Incode defend the capture path with liveness challenges and injection detection rather than classifying a delivered file. Given the 741% rise in iOS injection attacks iProov recorded across 2025, a control that verifies the camera as well as the face is the one that matches the threat.

Provenance: the checks that prove something

Provenance is the only part of this field that gives near-definitive answers, and it is worth checking first because it costs nothing.

• Content Credentials Verify reads C2PA manifests on any file, including video. C2PA published Content Credentials 2.3 on 9 February 2026, adding live video support, and reports more than 6,000 members and affiliates with live applications.
• SynthID finds Google’s invisible watermark in output from Veo, Imagen, Lyria and Gemini. Google DeepMind reported in May 2026 that more than 100 billion files carry the mark, with OpenAI, Nvidia and ElevenLabs among newer adopters.
• Sora 2 video carries C2PA Content Credentials on all assets plus a visible moving watermark on downloads from the consumer app. OpenAI has not documented an invisible watermark for Sora video; its SynthID adoption covers images, announced 19 May 2026, and audio, announced 31 July 2026.

The trap is assuming these are one check. A SynthID lookup returns “not detected” on a Sora video, a C2PA read returns nothing on a Veo clip whose metadata a social platform has stripped, and both return nothing on video from an open-weight model — which is most of what circulates. Run all of them, treat a positive as strong evidence, and treat a negative as no evidence at all.

Feature comparison: the full matrix

“Not tested” means no result in the Podonos benchmark, not that a tool performs badly. Most of this market has never been independently measured on any modality.

Which deepfake detector should you use?

Best overall for audio

Resemble Detect, at 99.47% accuracy with a 0.7% false-positive rate and 0.4% false-negative rate — strong in both error directions, which almost nothing else in the field manages.

Best when a false accusation is costly

Whispeak, for the most balanced error profile measured: 2.9% false positives against 1.7% false negatives.

Best for contact centres and live phone fraud

Pindrop Pulse, because roughly 282 milliseconds per file is fast enough to act on mid-call and 95.05% is close enough to the leaders that speed decides it. Budget for review capacity to absorb its 6.2% false-positive rate.

Best for executive impersonation on video calls

GetReal Security, because continuous identity verification re-checks a participant throughout a meeting rather than only at the point they join.

Best for investigations and evidence

Sensity AI, for multi-layer forensic reporting that documents why a file was flagged rather than returning a bare score.

Best for platforms and moderation at volume

Hive AI, for metered per-unit pricing and the strongest independent image result in the category — but use something else for audio.

Best free deepfake detector

DeepFake-o-Meter, because running several research models and showing their disagreement is more informative than one confident probability from a black box.

Best for identity verification and account recovery

iProov or a comparable liveness vendor, because injection attacks bypass the file layer entirely and a media classifier cannot see them.

What to be most careful with

Any tool whose false-positive rate you have not seen. Reality Defender’s 53.7% on the one independent audio test it has, and Hive’s 30.5% audio false-negative rate, are both invisible in marketing material and both decisive for a production deployment. Ask for both error rates at the shipping threshold before you sign anything.

The false-positive problem: why a detection score is not proof

Detection errors run in two directions and both cause harm.

A false positive means authentic media is called fake. In the deepfake era this has a name — the liar’s dividend — and it is the more corrosive failure: once “it was AI” is an available defence, a detector wrongly flagging real evidence supplies the alibi. The measured range across the Podonos benchmark is wide enough to change a business case entirely: 0.7% for Resemble DETECT-World, 1.5% for Aurigin, 2.9% for Whispeak, 6.2% for Pindrop and 53.7% for Reality Defender. At contact-centre volumes, the difference between 0.7% and 6.2% is thousands of wrongly-suspected customers a month, and 53.7% is not a deployable control at all. Threshold behaviour, not headline accuracy, is the real procurement question.

A false negative means a deepfake passes. On in-the-wild video, where the best commercial detector reached 78%, roughly one in five fakes gets through. Deepfake-Eval-2024 identified where detectors specifically struggle: recent diffusion-based synthesis, manipulations that are partial or non-facial, and audio in underrepresented languages or padded with music and silence.

The correct posture is that a detector output is an input to a decision made by a person with other information, never the decision itself. That is also how forensic practitioners use these tools — and human analysts, at roughly 90% on the same benchmark, still outperform the automated systems.

Where deepfake detectors break

• Generators newer than the training data. The four open-source models trained on ASVspoof 2019 LA scored 48–63% against 2026 voice clones. Every detector is fitted to the artefacts of the generators it has seen, and the fit expires.
• Re-encoding and platform compression. Media that has passed through two or three platforms has been transcoded each time, degrading exactly the fine-grained traces classifiers key on. Podonos tests six container formats for this reason.
• Short clips. Reality Defender declined to score 17.2% of clips under roughly 1.5 seconds — and short audio is exactly what a fraud call gives you.
• Screen recordings and virtual cameras. A deepfake replayed through a virtual camera into a video call arrives as a clean, plausibly-encoded stream; the injection is invisible to a classifier looking only at pixels.
• Partial and non-facial manipulation. Most video detectors are face-centric. Change the background, an object, a document held to camera, or a few seconds of a long clip, and detection rates fall sharply.
• Languages and voices outside the training distribution. Deepfake-Eval-2024 drew from 52 languages precisely because performance is uneven across them.

Pricing comparison: what you will actually pay

This is the least transparent pricing landscape of any category we cover. Most vendors quote rather than publish, and several list no rate card at all.

For a personal or one-off check, the free tools cover it entirely and paying is unnecessary. Paid contracts here are bought for a fraud-loss reason and should be evaluated against your own recorded incidents rather than a vendor benchmark. Ask any vendor for false-positive and false-negative rates at the threshold they propose to run in production, the real-time factor, and when the model last saw new generators. Any vendor unwilling to give you all four has answered the question.

How to verify a video or voice call without a detector

Detectors are one step of five, not the answer.

1. Check provenance first, because it is free and definitive when it fires. Run the file through Content Credentials Verify and a SynthID check. A confirmed signature settles the question; an absent one settles nothing.
2. Find the earliest version. Reverse image search on keyframes, and search for the claimed event. A clip that appears first on an anonymous account and nowhere in any newsroom’s footage is usually answered by that fact alone.
3. Run several detectors and read the spread. Use DeepFake-o-Meter to see multiple models’ scores at once. Where they disagree, you have learned something a single score would have hidden.
4. Break the channel, for anything live. If a call asks for money or credentials, hang up and call back on a number you already had. Agree a spoken pass phrase in advance with family and with your finance team. This single control defeats the entire category of voice-clone fraud regardless of clone quality, and it costs nothing — see our AI voice cloning page for how little source audio a convincing clone now needs.
5. Verify the context, not just the media. Does the claimed location, language, weather, uniform or background match? Authentic footage can be entirely real and captioned to mislead, and no detector in existence catches that.

Recent developments (2025–2026)

• 31 July 2025 — Reality Defender opens a public API with a free tier, at 50 audio or image scans a month (PR Newswire).
• 9 February 2026 — C2PA publishes Content Credentials 2.3, adding live video support and reporting more than 6,000 members and affiliates with live applications (C2PA).
• 8 April 2026 — iProov publishes its 2026 Threat Intelligence Report, recording a 741% rise in iOS injection attacks across 2025, a 1,151% rise in the second half alone, and a 720% Southeast Asian spike in Q3 (Business Wire).
• 5 May 2026 — Podonos publishes the first private-label audio deepfake benchmark, testing eight systems across 4,524 clips at the thresholds vendors actually ship (Podonos). It expanded to sixteen systems by August 2026, when Resemble’s DETECT-World took the lead at 99.47%.
• 18 May 2026 — GetReal Security makes continuous identity verification generally available, citing more than $1.6 billion in deepfake-driven losses in 2025 (PR Newswire).
• 19 May 2026 — OpenAI becomes a C2PA Conforming Generator Product and adopts SynthID for images, extending the same treatment to audio on 31 July 2026 (OpenAI). Google DeepMind reports more than 100 billion files watermarked with SynthID since 2023.
• May 2026 — FTC enforcement of the TAKE IT DOWN Act begins. The Act was signed on 19 May 2025, but covered platforms were given a year to stand up notice-and-removal procedures for non-consensual intimate deepfakes, so the removal obligation became enforceable in May 2026.
• 25 June 2026 — Gartner publishes its first Emerging Market Quadrant for Deepfake Detection — Startup Vendors, naming Reality Defender and GetReal Security as Market Shapers (Reality Defender).
• 2 August 2026 — EU AI Act Article 50 becomes applicable, requiring machine-readable marking of AI output and visible labelling of deepfakes, with penalties up to €15 million or 3% of worldwide annual turnover (EU AI Act). The Commission confirmed the Code of Practice on Transparency of AI-generated Content as an adequate compliance route on 8 July 2026, with the AI Board following on 9 July.
• Denmark’s likeness law remains in progress. A copyright-style amendment granting individuals rights over their face and voice in AI-generated content was notified to the EU in late 2025 and was expected to take effect during 2026, but we have found no confirmation of a final Folketing vote. Treat it as pending rather than in force, and check before relying on it.
• 1 September 2025 — China’s Measures for Labelling AI-Generated Synthetic Content take effect, requiring both visible and machine-readable labels on AI-generated image, video, audio and text across Chinese platforms.

Frequently asked questions

What is the best deepfake detector in 2026?

It depends on the modality, and no single tool leads both. For audio, Resemble Detect leads Podonos’s private-label benchmark at 99.47% accuracy with a 0.7% false-positive rate, followed by Whispeak at 97.70% and Pindrop at 95.05%. For video, no tool has a strong independent result: the best commercial detector managed 78% on Deepfake-Eval-2024, and GetReal Security and Reality Defender are the vendors Gartner named Market Shapers in June 2026 rather than the vendors with the best measured scores. For a free one-off check, use the University at Buffalo’s DeepFake-o-Meter.

Are deepfake detectors accurate?

On audio, the best ones now are: several systems cleared 95% on a private-label benchmark in 2026, with the leader at 99.47%. On video, no. The best commercial video detector reached 78% accuracy on Deepfake-Eval-2024, the largest in-the-wild benchmark, against roughly 90% for human forensic analysts. Vendor claims of 95–99% are measured on the vendor’s own test set at the vendor’s own threshold and should be read as a ceiling, not a prediction.

Can a deepfake detector be wrong about a real video or voice?

Yes, and false positives are the more damaging error. Measured false-positive rates on the Podonos audio benchmark range from 0.7% for the leader to 53.7% for Reality Defender, which flagged more than half of all genuine audio as fake. Wrongly labelling authentic media as synthetic also feeds the liar’s dividend, where a real recording can be dismissed as AI. Never treat a detection score as proof on its own.

What is the best free deepfake detector?

DeepFake-o-Meter from the University at Buffalo Media Forensics Lab, because it runs several published research models and shows each result separately, so you can see where they disagree. Deepware Scanner is free for video face-swap checks, the Hiya Deepfake Voice Detector is a free Chrome extension for audio playing in the browser, and Reality Defender gives 50 free scans a month.

How can I tell if a video was made by Sora or Veo?

Check its provenance rather than its pixels. Sora 2 video from OpenAI carries C2PA Content Credentials, readable at Content Credentials Verify, and downloads from the consumer app carry a visible moving watermark. Veo video from Google carries an invisible SynthID watermark, readable with Google’s SynthID tooling. The two schemes do not read each other, so run both. A negative result from either proves nothing, because metadata may have been stripped in re-encoding or the video may have come from a model that adds no mark at all.

Can deepfake detectors detect AI-generated video from open-source models?

Less reliably, and with no provenance backstop. Open-weight video models generally add neither C2PA credentials nor a watermark, so first-party verification tools cannot help and you are left with pixel and temporal classifiers alone. Those classifiers key on artefacts from the generators they were trained against, so accuracy degrades against anything released after a detector’s training cut-off — the effect measured directly on the audio side, where models trained on a 2019 dataset scored 48–63% against 2026 voice clones.

How do I protect myself from a deepfake voice scam?

Break the channel rather than trying to judge the voice. Hang up on any urgent request for money or credentials and call back on a number you already have. Agree a spoken pass phrase in advance with family members and with anyone in your organisation authorised to move money. Limit the amount of your voice posted publicly, since convincing clones now need only seconds of audio. These controls work regardless of clone quality, which is more than any detector can claim.

What is an injection attack, and can a deepfake detector stop one?

An injection attack feeds synthetic video or audio directly into a verification system through a virtual camera or intercepted data stream, rather than presenting a fake to a real camera. A media classifier inspecting the delivered frames often sees a clean, plausible stream and passes it. Stopping injection requires controls on the capture path itself — device attestation, challenge-response and liveness — which is what identity vendors such as iProov sell. iProov recorded a 741% rise in iOS injection attacks across 2025.

Is there a legal requirement to label deepfakes?

In the European Union, yes. EU AI Act Article 50 became applicable on 2 August 2026 and requires providers to mark AI output in a machine-readable format and deployers to visibly label deepfakes, with penalties up to €15 million or 3% of worldwide annual turnover. China’s labelling Measures have required visible and machine-readable labels on AI-generated media since 1 September 2025. In the United States there is no general federal labelling law, though FTC enforcement of the TAKE IT DOWN Act’s removal obligation for non-consensual intimate deepfakes began in May 2026, and state laws are expanding.

How much do deepfakes actually cost organisations?

Reported deepfake-fraud losses total at least $3.7 billion cumulatively, with roughly 89% recorded during 2025 and the first half of 2026, according to a Surfshark analysis also cited by GetReal Security. The FBI’s Internet Crime Complaint Center logged $893 million across 22,364 AI-related complaints in 2025, the first year AI appeared as its own descriptor, and Deloitte projects US generative-AI-enabled fraud losses rising from $12.3 billion in 2023 to $40 billion by 2027. The largest publicly reported single incident remains the $25.6 million taken from engineering firm Arup in January 2024, when a finance employee in its Hong Kong office made 15 transfers after a video call on which every other participant was a deepfake (CNN). Under-reporting is severe, so these figures are a floor.

Why do vendor accuracy claims and benchmark results differ so much?

Because vendors compute their own figure on their own test set at a threshold they choose, and Podonos does neither. Podonos uses private gold-standard labels — a vendor cannot tune against a test set it cannot read — scores every system at the threshold it actually ships, and tests six audio container formats rather than clean studio wav files. That is why a vendor advertising 99% can land at 71% on the same task. Ask which test set, at which threshold, and when the model was last retrained.

What is the difference between a deepfake detector and an AI image or text detector?

They target different media with different technology and different accuracy. A deepfake detector analyses video and audio for synthesis and manipulation artefacts, or reads provenance marks, and increasingly has to run in real time on a live call. An AI image detector classifies still images or reads C2PA and SynthID signals. An AI text detector estimates how statistically predictable a passage of writing is, and has no watermark to fall back on for most models. The tools are not interchangeable and the vendors barely overlap.

Conclusion: how to check a video or voice in September 2026

Deepfake detection has split into three layers moving at different speeds, and knowing which one your problem sits in decides everything.

• Audio detection works, if the vendor retrained recently. Several systems cleared 95% on a private-label benchmark in 2026 and the leader reached 99.47%, while open-source models trained on a 2019 dataset scored 48–63%. Training recency is the number to ask about.
• Video detection lags badly. The best commercial detector managed 78% on real circulated deepfakes, below human forensic analysts and far below the marketing. Treat every video score as a prompt to investigate rather than a finding.
• Provenance is the only near-definitive layer, and it is free. C2PA credentials and SynthID watermarks now cover output from OpenAI and Google, EU law has required machine-readable marking since 2 August 2026, and the readers take seconds. But the schemes do not read each other, marks get stripped in re-encoding, and open-weight models add none — so a positive is strong evidence and a negative is no evidence.

If you are an individual checking one file, use the free tools, check provenance first, and stop there. If you are an organisation, the honest procurement question is not which vendor advertises the highest accuracy — the spread between 99% claimed and 71% measured on the same task is the whole lesson of this year’s testing. Ask what your false-positive rate will be at the threshold you actually run, how recently the model saw the generators your attackers are using, and whether the control matches the attack path. And for voice fraud, the cheapest effective control is still a call-back and a pass phrase, which does not degrade when a new generator ships.