Deepfakes Are Harder to Spot. What to Know.

On October 7, 2026, Google opened its SynthID Detector to everyone, worldwide and in English. Anyone can now check whether an image, video or audio file carries the hidden watermark added by AI tools from Google and partners including OpenAI, NVIDIA and Kakao; Google says Apple support is coming. That helps in one direction only: a match points to AI involvement, but no match proves nothing. Recent research shows people miss about half of the newest fake videos and most detection software tested does little better than chance, so for a surprising call, video or payment request, the most reliable check is still to reach the person or source through a channel you already trust.

Key takeaways

In a September 2026 preprint, viewers correctly flagged the newest AI-generated videos only 52.6% of the time, against 74.5% on a set of face-swap and talking-head fakes, and most of the nine detection models tested scored close to chance.
In a 2026 Scientific Reports study, people classified faces correctly 65.48% of the time but cloned voices only 53.95%, and their confidence did not track their accuracy for voices.
SynthID Detector and Content Credentials can confirm AI involvement or a file's history, but a missing watermark or missing credentials is common and is not evidence that a file is real.
Since a February 2024 FCC ruling, robocalls using AI-generated voices need the called person's prior express consent unless an emergency purpose or exemption applies; a federal NO FAKES Act on digital replicas cleared a Senate committee in June 2026 but has not become law.
The FBI recommends a secret family word or phrase, and the FTC advises calling the person back on a number you know is theirs.

What it means for you

People getting urgent calls or videos from "family": a familiar voice or face is no longer proof of identity; call back on a known number or ask for a family code word.
Older adults: requests for wire transfers, gift cards or cryptocurrency are a warning sign on their own (see the FAQ on awareness by age).
Social media users seeing viral clips: labels and checkers can confirm AI use but cannot clear a clip (see "Labels and watermarks: the limits").
Parents: make sure children know the family code word, and limit public photos and voice recordings of the family; the FBI advises a family code word and private social media accounts.
People whose images were faked: report the post to the platform; for intimate images, federal law now sets a 48-hour removal duty (see "What the law covers in 2026").
Employees and small business owners approving payments: a video call with a familiar face does not authorize a transfer; confirm through a separate, known channel before money moves.

What the detection research shows

The premise holds for the newest material, with nuance. A deepfake is audio, an image or a video made or altered with AI so that a real person appears to say or do something they did not. Studies report the share of items people label correctly; guessing scores 50%.

Study (date)	What was tested	Result
Köbis et al., iScience (Oct 2021)	210 people judging real and deepfake videos	Could not reliably detect fakes, leaned toward calling fakes authentic, and overestimated their skill; warnings and cash incentives did not help
Nightingale and Farid, PNAS (Feb 2022)	StyleGAN2 faces vs real photos	48% accuracy (315 people); 59% after training (219 people); synthetic faces rated 7.7% more trustworthy
Mai et al., PLOS ONE (Aug 2023)	529 listeners, English and Mandarin speech	Fakes spotted 73% of the time when clips were judged one by one; 85.59% when comparing a pair
iProov, vendor survey (Feb 2025)	2,000 UK and US consumers, images and video	0.1% got every item right; fake videos were 36% less likely to be identified than fake images
Ivory et al., Scientific Reports (2026)	172 people; Adobe Firefly faces, ElevenLabs voice clones	Faces 65.48%, voices 53.95%; skill at one did not predict skill at the other
DF26 preprint, arXiv (Sep 2026)	2,420 videos from seven current video generators, 232 labeling sessions	52.6% of fakes caught vs 74.5% on a face-swap and talking-head benchmark (CelebDF++); best detector AUROC 69.7%

Newer fakes beat people and software

The 2021 Köbis study found a "seeing is believing" bias. The 2022 Nightingale and Farid experiments (published in PNAS) showed that faces from a 2020-era generator were already at chance level.

The clearest test of "harder to spot" comes from the DF26 preprint published on September 7, 2026, which has not yet been peer reviewed. Its fakes are fully generated videos of people speaking, made with seven current text-to-video and image-to-video tools. Viewers, who watched the clips without sound, caught 74.5% of fakes from an established face-swap and talking-head benchmark but only 52.6% of the new fakes, while still labeling 76.0% of real videos correctly. Most of the nine detection models tested also scored near chance on the new set; the best reached an AUROC of 69.7%.

A separate benchmark of fakes that circulated in 2024, Deepfake-Eval-2024, reported that open-source detectors' AUC fell by 50% for video, 48% for audio and 45% for images compared with older test sets. AUC (or AUROC) is a 0-to-100% score where 50% means the model does no better than chance.

Voices are the weak spot

In the 2026 Scientific Reports study, voice-clone accuracy sat just above chance, and people were about as confident about voices (3.79 on a 5-point scale) as about faces (3.74), even though they were much less accurate. The 2023 PLOS ONE speech study found that hearing five example deepfakes first improved accuracy only slightly, by 3.84% on average.

Illustrative example: at the voice accuracy measured in 2026 (53.95%), a listener judging 20 short clips would misclassify about 9 of them (20 × 0.4605 = 9.2). At the face accuracy from the same study (65.48%), about 7 of 20 would be wrong (20 × 0.3452 = 6.9). Either error rate is too high to rely on for a decision about money.

What does help

Two findings point to practical defenses. Comparison works: in the 2023 speech study, accuracy rose from 70.35% when clips were heard alone to 85.59% when a real and a fake version were played as a pair. Outside a lab you rarely get that side-by-side, which is why calling the person back on a number you already know works better than listening harder: it moves the check to a channel the caller does not control.

Training helps somewhat for faces: a PNAS study announced by the University of Victoria in June 2026 reported that under an hour of training raised face-detection accuracy by "nearly 30 per cent," though it was tested on StyleGAN faces and the researchers say it is not yet known whether it carries over to other generators. The visual tells the FBI lists, such as distorted hands, odd teeth or eyes, wrong shadows and lag, can still expose a sloppy fake, but their absence proves nothing.

One caveat on the table: iProov sells biometric identity checks, and its participants were primed to look for fakes.

Labels and watermarks: the limits

Provenance tools work differently from detectors. A watermark is a signal hidden in the content by the AI tool that made it. Content Credentials, built on the C2PA standard, are signed metadata recording how a file was captured and edited. Both can confirm something positive; neither can clear a file.

Check	A positive result means	A negative result means
SynthID Detector (synthid.com, free, sign-in required)	A watermark from Google, OpenAI, NVIDIA or Kakao tools was found; it cannot tell AI-made from AI-edited	Nothing: other AI tools do not use SynthID, and watermarks can be stripped in some cases
Content Credentials (contentcredentials.org/verify)	A signed record of capture, edits and any declared AI use	Nothing: screenshots, file conversions, and uploading or sharing can strip the metadata
Platform AI label (YouTube, TikTok and others)	The creator disclosed AI use, or the platform read metadata or detected it	Nothing: labels depend on disclosure or surviving metadata
Reverse image search (Google Lens, TinEye)	Earlier copies that can reveal the original context	Nothing: brand-new fakes have no earlier copies

The sign-in requirement and the AI-made versus AI-edited limit are reported by MacRumors; free access and watermark stripping by The Next Web. Google's own announcement does not mention any of these points. The Gemini app has offered SynthID image checks since November 2025.

Content Credentials are spreading at the camera. On Pixel 10 phones, Google says Pixel Camera attaches Content Credentials to every JPEG photo. But OpenAI's help page on C2PA says Content Credentials "can be lost through everyday actions, such as taking a screenshot or converting a file," and that content with no OpenAI signal could still have been made with its tools, for example if metadata was stripped during upload or sharing. Upload the original file, not a screenshot, to Content Credentials Verify.

Platform labels have similar gaps. YouTube requires creators to disclose realistic AI content that makes a real person appear to say or do something they did not, and may add a label itself when it detects C2PA metadata. TikTok began in May 2024 to auto-label AI-generated images and videos made on some other platforms when they carry Content Credentials, and said labeling depends on that metadata being present.

What the law covers in 2026

US rules target specific harms rather than deepfakes in general. This section is general information, not legal advice.

AI voices in robocalls: in a declaratory ruling released February 8, 2024, the FCC confirmed that AI-generated voices count as "artificial" under the Telephone Consumer Protection Act, so such calls need the called party's prior express consent unless there is an emergency purpose or exemption, and prior express written consent if the call carries an advertisement or telemarketing. Callers must also identify who is responsible for the call. The rule makes cloned-voice robocalls illegal without consent; it does not stop a scammer from placing one, but state attorneys general as well as the FCC can pursue those behind them.
Impersonation scams: the FTC's impersonation rule took effect April 1, 2024, letting the agency seek civil penalties and refunds from scammers posing as government agencies or businesses. A proposed extension to impersonation of individuals, which would cover a cloned relative's voice, has not been adopted: the FTC's September 2026 notice says it held a hearing on that proposal in January 2025 and "is still evaluating its options." The same notice, announced on September 24, 2026, asks for comment on whether platforms whose ad tools help impersonators of businesses and government should be covered.
Intimate images: since May 19, 2026, covered platforms must remove nonconsensual intimate images, and known identical copies, within 48 hours of a valid request; the FTC takes complaints about platforms that do not.
Voice and likeness: the NO FAKES Act, which would create a federal right over unauthorized digital replicas of a person's voice and likeness, passed the Senate Judiciary Committee unanimously in June 2026 and was placed on the Senate calendar on June 24, but Congress.gov showed no floor vote in either chamber as of October 14, 2026. Tennessee's ELVIS Act, in effect since July 1, 2024, already adds voice to the state's publicity rights.
Elections: Public Citizen counted 30 states with election deepfake laws as of May 2026. Courts have trimmed some: a federal judge blocked California's AB 2839 in October 2024 on First Amendment grounds and struck down AB 2655, its platform takedown law, in August 2025 as preempted by Section 230.

Frequently asked questions

How much of my voice does a scammer need?

Not much. The FTC warns that a scammer can clone a relative's voice from a short audio clip, such as one posted online. The FBI advises limiting public images and voice recordings and setting social accounts to private.

Are paid deepfake detection apps reliable?

Not reliable enough to decide on their own. In the Deepfake-Eval-2024 benchmark, built from fakes that circulated online in 2024, commercial detection models did better than off-the-shelf open-source ones, yet still did not reach the accuracy of human deepfake forensic analysts. That is one benchmark, and it does not rate individual consumer apps, so treat any score as one input, not a verdict.

Does an "AI info" label mean a post is fake?

Not necessarily. Meta renamed its "Made with AI" label to "AI info" in July 2024 after photographers' lightly retouched pictures were tagged. A label can mean the whole image was generated or only that an AI editing tool was used, so read the details behind it.

Are older adults less aware of deepfakes?

In iProov's 2025 survey, 39% of people aged 65 and over and 30% of those aged 55 to 64 said they had never heard of deepfakes, against 22% of all participants. The survey was run by a company that sells identity verification. Someone who has never heard of voice cloning has little reason to doubt a familiar voice, so a family code word agreed in advance gives them a check that does not depend on spotting the fake.

Do the EU's new AI rules change anything in the US?

They apply to AI systems offered or used in the EU. Under Article 50 of the EU AI Act, applying from August 2, 2026, those deploying AI to create deepfakes must disclose that the content is artificial (with a lighter duty for evidently artistic, satirical or fictional work), and AI providers must mark generated output in a machine-readable way. Law firm summaries of a July 2026 EU amendment, such as Cuatrecasas's, report that tools already on the market before August 2 have until December 2, 2026 to add the marking. There is no federal US equivalent, so US readers may see labels on some content and not on other content.

Before you act on a clip or call

Pause on urgency. Pressure to act now, keep a secret or pay by wire, gift card or cryptocurrency is a warning sign whatever the voice sounds like.
Verify through a second channel. Hang up and call the person or company on a number you already have, or check with another relative or a colleague.
Use the code word. Agree on a secret word or phrase with family now, and ask for it on any unexpected request.
Check the file, then the source. Look for an AI label, run the original file through synthid.com and Content Credentials Verify, and do a reverse image search; treat any negative result as unproven.
Confirm payments out of band. At work, require a callback to a known number before changing payment details or sending a transfer requested on a call or video meeting.
Report it. Report the post to the platform, scams to the FTC at ReportFraud.ftc.gov, and losses to the FBI at ic3.gov, with phone numbers, payment details and how contact began.