Can you still tell what's real?

The full story

The episode's narration, word for word. Numbers link to the sources below; timestamps play that moment.

The question 0:00

Last time, we followed what AI uses. This time, it's what AI makes, and whether you can trust your own eyes and ears. Early in 2024, a finance worker in Hong Kong joined a video call with his company's chief financial officer and several colleagues ^1,2. They looked and sounded like people he knew ². Police said every one of them, except him, was an impersonation ¹. So, can you still tell what's real?

Fifteen transfers 0:31

It started with an email, apparently from the company's UK office, asking for a secret transaction ². He suspected phishing, a fake message fishing for money ². Then came the video call, and his doubts went away ². He made fifteen transfers, two hundred million Hong Kong dollars, about twenty-six million US dollars ^1,2. It was reported to police on the twenty-ninth of January ¹.

Built from public videos 0:59

Police said the scammers had used video and audio of the real executives that was publicly available online ¹. The fake call was pre-recorded, so nobody on it actually spoke with him ¹. In May 2024, the engineering firm Arup confirmed it was the victim ². It's the firm that consulted on the structural engineering of the Sydney Opera House ². Arup said fake voices and images were used ².

Don't trust the voice 1:28

Most people will never sit through a fake board meeting. The common version rings your phone. In March 2023, the US Federal Trade Commission warned of panicked calls from a grandchild in trouble, in a cloned voice ³. A scammer needs only a short clip of that person's voice, perhaps from something posted online, it said ³. In 2026, Australia's Scamwatch said a few seconds of recording can be enough ⁴. The Commission's advice was blunt: don't trust the voice ³.

What changed, and when 2:02

So what changed, and when? In 2014, researchers found a way to train one program to fool another, and it became an engine for realistic fakes ^5,6. In a 2022 study, people couldn't tell faces made by one such program from real ones ⁷. In 2023, Microsoft researchers imitated a voice from a three-second recording, in a research paper ⁸. In 2024, OpenAI previewed one needing fifteen seconds, and chose not to release it widely ⁹. And in 2025, Google said its Veo 3 could make video with sound ¹⁰.

Forger and detective 2:43

That 2014 idea, the generative adversarial network, shows why spotting fakes is hard ⁵. Picture a forger and a detective. The detective learns to tell real pictures from the forger's fakes ⁵. The forger is trained with one goal: make the detective wrong ⁵. So every improvement in the detective teaches the forger what to fix ⁵. Take it as a picture of the pressure, not as how every fake is made.

A coin flip on faces 3:13

How good are people at this? In 2022, Sophie Nightingale and Hany Farid asked volunteers to sort faces into real and computer-made ⁷. The first group averaged forty-eight point two per cent, just under a coin flip ⁷. A new group, given training and feedback, reached fifty-nine per cent ⁷. And people rated the computer-made faces as slightly more trustworthy than real ones ⁷.

What ears can do 3:42

What about voices? In 2023, a University College London team found listeners spotted fake speech seventy-three per cent of the time, and warned its fakes came from older methods ¹¹. In 2025, Sarah Barrington, Emily Cooper and Hany Farid at Berkeley cloned two hundred and twenty real voices ¹². About eight times in ten, listeners judged a clone to be the same person ¹². They called a clone AI only about sixty per cent of the time, where pure guessing gets fifty ¹².

Fakes from the wild 4:17

So can software do better? In 2025, researchers at TrueMedia.org, the University of Washington and elsewhere collected suspect clips that people flagged online in 2024, real and fake ¹³. They gathered video, audio and images from eighty-eight websites in fifty-two languages ¹³. Then they tested free detectors that had scored superbly on their original test sets ¹³.

Close to guessing 4:45

Detectors are often scored by AUC: the chance a detector rates a random fake as more suspicious than a random real clip ¹⁴. One is perfect, a half is a coin flip ^14,13. On their own test sets, nine detectors scored nought point eight seven or better ¹³. On the real-world fakes, they scored at most nought point six three, close to guessing ¹³. The authors suggest they'd learned quirks of older fakes ¹³. Commercial detectors did better, but none reached ninety per cent accuracy ¹³. As of October 2026, Australia's eSafety Commissioner says detection tools are lagging behind ¹⁵.

The false-alarm trap 5:28

Even good detectors have a second problem. Imagine one that's right ninety per cent of the time, on fakes and real clips alike. Give it ten thousand clips, one in a hundred fake. It catches ninety fakes, but wrongly flags nine hundred and ninety real clips. So only one alarm in twelve is actually a fake. When fakes are rare, even an accurate detector mostly cries wolf.

Flip the question 5:56

So some companies flipped the question: instead of proving something fake, prove where it came from ¹⁶. In February 2021, Adobe, Arm, the BBC, Intel, Microsoft and Truepic founded the Coalition for Content Provenance and Authenticity, or C2PA ¹⁷. Its standard attaches Content Credentials to a file: a signed record of what made it and how it was edited, including any use of AI ¹⁶. Leica sold what it called the first camera with them in 2023, and Google says its Pixel 10 phones, from 2025, add them to every photo from their camera app ^18,19.

A fingerprint that breaks 6:40

How does a record show it hasn't been tampered with? It uses a cryptographic hash, a fingerprint calculated from every bit of the file ¹⁶. Here's a tiny picture, eight squares by eight, and its fingerprint. Brighten one square by one shade, and sixty-three of its sixty-four characters change. The camera or app signs that fingerprint with a digital certificate, a bit like a wax seal ¹⁶. Change the picture or the record, and the seal no longer matches ¹⁶.

What a seal can't tell you 7:13

But a seal tells you who sealed it, not whether the letter is true. The C2PA's own explainer says provenance alone can't tell you whether content is true, accurate or factual ¹⁶. Can they be removed? Its answer is yes ¹⁶. OpenAI's help page says a screenshot can lose them ²⁰. So a credential tells you where a file came from, and a missing one proves nothing ^16,20.

Ink inside the picture 7:41

So companies also hide a signal inside the content itself: a watermark ²¹. Google DeepMind's SynthID, launched in 2023, changes pixels in ways your eye can't see but software can, though Google says it isn't foolproof against extreme changes ²¹. In December 2024, Meta released Video Seal, an open-source video watermark, whose authors say its robustness and security need more work ²². As of October 2026, OpenAI says images from ChatGPT carry both Content Credentials and SynthID, and warns heavy cropping or compression can defeat the watermark ²⁰.

How a watermark hides 8:25

Here's the idea as a toy, not any company's real method. Nudge every pixel of a smooth picture up or down by at most two shades out of two hundred and fifty-five, following a secret pattern. You can't see it. But compare the picture's fine detail with the secret pattern, and the match scores one, while an unmarked picture scores zero. Blur it heavily and add grain, and the score drops to almost nothing. In 2024, University of Maryland researchers showed that several published watermarking methods could be stripped, and even faked ²³.

Labels, on paper and in practice 9:05

Since March 2024, YouTube has required creators to disclose realistic altered or synthetic content, a box this channel ticks ²⁴. In 2024, Meta began labelling AI images from other companies' tools, then renamed its Made with AI label after it landed on photos only touched up with AI tools ^25,26. TikTok said it was the first video platform to read Content Credentials ²⁷. But in October 2025, the Washington Post uploaded an AI-made video to eight social apps ²⁸. Only YouTube told viewers it wasn't real, and only in the description ²⁸.

What Australian law says 9:47

Some fakes are crimes as well as lies. Since September 2024, Australian federal law has explicitly made it an offence to share sexual material of an adult without their consent, whether it's real or created or altered by technology, such as deepfakes ²⁹. The maximum penalty is six years in prison, or seven in aggravated cases, such as when the sharer also made it ²⁹. Australia's eSafety Commissioner can help get digitally altered images removed ¹⁵.

The liar's dividend 10:20

There's a twist. In 2019, the law professors Robert Chesney and Danielle Citron named the liar's dividend ⁶. Deepfakes, they argued, make it easier for liars to escape accountability for things that are true ⁶. A sceptical public is primed to doubt real evidence, and the dividend grows the better people are taught about deepfakes ⁶. The goal isn't doubting everything; it's checking the right things.

What you can actually do 10:49

So what can you do? If a call asks for money urgently, hang up and call back on a number you already know ^3,4. The FBI suggests a secret word with your family ³⁰. Ask whether it's what that person would say or do, there, at that time ¹⁵. Glitches around the mouth or eyes only sometimes give away a cruder fake ¹⁵. A label or credential is information, not proof, and no label doesn't mean real ^16,28. Report scams to Scamwatch, and image abuse to eSafety ^4,15.

Made to fool you 11:24

So, can you still tell what's real? Not reliably by eye or ear, and not reliably with software ^7,13,12. But you can check where something came from, and how it reached you. These fakes were made on purpose, to fool people. Sometimes, though, AI does harm nobody intended, by treating some people worse than others. Why does AI sometimes treat people unfairly?

Sources

Every factual claim in the episode is tied to one of these. Spotted an error? Tell us.

1. Agence France-Presse, "Multinational loses HK$200 million to deepfake video conference… — hongkongfp.com
2. K. Magramo, "British engineering giant Arup revealed as $25 million deepfake scam… — edition.cnn.com
3. A. Puig, US Federal Trade Commission, "Scammers use AI to enhance their family… — consumer.ftc.gov
4. ACCC Scamwatch, "How scammers use technology and AI" (published 23 March 2026; read 16… — scamwatch.gov.au
5. I. J. Goodfellow et al., "Generative Adversarial Networks", arXiv:1406.2661 (10 June 2014) — arxiv.org
6. R. Chesney, D. K. Citron, "Deep Fakes: A Looming Challenge for Privacy, Democracy, and… — californialawreview.org
7. S. J. Nightingale, H. Farid, "AI-synthesized faces are indistinguishable from real… — pmc.ncbi.nlm.nih.gov
8. C. Wang et al. (Microsoft), "Neural Codec Language Models are Zero-Shot Text to Speech… — arxiv.org
9. OpenAI, "Navigating the Challenges and Opportunities of Synthetic Voices" (29 March 2024) — openai.com
10. Google, "Fuel your creativity with new generative media models and tools" (Veo 3,… — blog.google
11. K. T. Mai, S. Bray, T. Davies, L. D. Griffin, "Warning: Humans cannot reliably detect… — arxiv.org
12. S. Barrington, E. A. Cooper, H. Farid, "People are poorly equipped to detect… — nature.com
13. N. A. Chandra et al., "Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of… — arxiv.org
14. Google for Developers, Machine Learning Crash Course, "Classification: ROC and AUC"… — developers.google.com
15. eSafety Commissioner, "Deepfake trends and challenges: position statement" (last… — esafety.gov.au
16. C2PA, "C2PA and Content Credentials Explainer", specification 2.4 — spec.c2pa.org
17. C2PA, "Technology and media entities join forces to create standards group aimed at… — c2pa.org
18. Leica Camera, "Leica M11-P" press release (26 October 2023) — leica-camera.com
19. Google, "How Pixel and Android are bringing a new level of trust to your images with… — blog.google
20. OpenAI Help Center, "Provenance signals in OpenAI-generated content" (read 16 October… — help.openai.com
21. Google DeepMind, "Identifying AI-generated images with SynthID" (29 August 2023) — deepmind.google
22. P. Fernandez et al. (Meta), "Video Seal: Open and Efficient Video Watermarking",… — arxiv.org
23. M. Saberi et al. (University of Maryland), "Robustness of AI-Image Detectors:… — arxiv.org
24. YouTube, "How we're helping creators disclose altered or synthetic content" (18 March… — blog.youtube
25. Meta, "Labeling AI-Generated Images on Facebook, Instagram and Threads" (6 February 2024) — about.fb.com
26. Meta, "Our Approach to Labeling AI-Generated Content and Manipulated Media" (5 April… — about.fb.com
27. TikTok, "Partnering with our industry to advance AI transparency and literacy" (May 2024) — newsroom.tiktok.com
28. K. Schaul, "We uploaded a fake video to 8 social apps. Only one told users it wasn't… — washingtonpost.com
29. Criminal Code Amendment (Deepfake Sexual Material) Act 2024 (Cth), No. 78, 2024… — legislation.gov.au
30. FBI Internet Crime Complaint Center, "Criminals Use Generative Artificial Intelligence… — ic3.gov

Researched and scripted with AI assistance, fact-checked claim by claim, with synthetic narration and diagrams drawn in code. How we make episodes.