Latest / AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence / When Your AI Lie Detector Misses the Truth
Transcript
- Lucas: There's a company called Voyager Labs — they've been selling an AI system that claims to detect deception from video footage. Airports, border control, even some employers have started piloting it. Luna: Right, I've heard of them. They analyze facial microexpressions, body language, voice tone — basically a digital polygraph. Lucas: Exactly. But here's the problem: the underlying science is shaky at best. In 2025, the ACLU filed a complaint with the FCC after one of Voyager's trials at a US airport flagged 40 percent of honest travelers as deceptive. Luna: Wait — 40 percent false positive rate? That's not a lie detector, that's a harassment generator. Lucas: That's exactly the argument. The core technology is based on Paul Ekman's work on microexpressions from the 1970s — but later studies have failed to replicate his findings reliably. And even if microexpressions were a perfect cue, the AI is interpreting them out of context. Someone could be anxious about flying, not lying about their destination. Luna: So we have a system that's both inaccurate and deployed in high-stakes environments. What's the actual legal framework here? Can they just roll this out? Lucas: That's the unsettling part. In the US, there's no federal law specifically regulating ai based deception detection. The FCC complaint argued that Voyager's system constituted an unauthorized 'communications interference' — essentially, that analyzing facial expressions without consent violates privacy statutes. But the FCC hasn't ruled yet. Luna: Meanwhile, the EEOC has been looking at whether using these systems in hiring discriminates against people with certain disabilities or anxiety disorders. If you have a nervous tic, the AI might flag you as deceptive. Lucas: Right, and that's not hypothetical. There was a case in 2024 where a job applicant with Tourette's syndrome was rejected after an AI interview analysis gave her a low honesty score. She sued under the Americans with Disabilities Act, and the company settled. Luna: It feels like we're repeating mistakes from other AI domains — facial recognition, predictive policing — where the technology gets deployed first and the harm is documented later. Lucas: Absolutely. And the stakes here are particularly personal. A false positive from a facial recognition system might mean you get stopped by security. A false positive from a deception detection system could mean you're denied a job, denied boarding, or flagged in a database that follows you around. Luna: Speaking of following you around — there's also the insurance angle. Some auto insurers have started using voice analysis on claims calls to detect fraud. If the AI thinks you're lying, your claim gets flagged for investigation. Lucas: That's a great point. One startup, ClearQuote, advertises a 30 percent improvement in fraud detection. But they don't publicly release their false positive rates. And if you're an honest claimant who speaks with hesitation — maybe English isn't your first language — you could be punished for a communication style, not deception. Luna: And there's no independent validation body for these systems. The companies self-report accuracy metrics, often based on their own benchmarks. Lucas: That's the core governance gap. Compare it to medical devices: before a new MRI machine can be sold, it has to go through FDA trials. But AI software that makes a determination about your honesty? No equivalent requirement. Luna: It's a real blind spot. And I think listeners can feel how concerning this is. If today's conversation gave you something useful — a new angle on AI risks, or just a concrete example to think about — it's worth mentioning that the reason we can dig into this without commercials is listener support. People chipping in at buy me a coffee dot com slash fexingo. Lucas: Yeah, it genuinely makes a difference. Keeps the show independent, keeps us free to follow the story wherever it goes — no advertisers to please. Luna: Exactly. And on that note, let's get back to the technology itself. What do we actually know about how these algorithms work under the hood? Lucas: Most of them are deep learning models trained on labeled datasets of 'deceptive' and 'truthful' videos. The problem is, those datasets are typically created in lab settings where participants are told to lie about something trivial — like whether they stole a playing card. That's a far cry from real-world lying about smuggling contraband or faking an insurance claim. Luna: So the training data doesn't generalize. And the models end up learning spurious correlations — like if the person looks up and to the left, that's a sign of lying — which is a myth that's been debunked. Lucas: Right. There's a famous study from the University of Cambridge that found that when they tested commercial deception detection systems on real courtroom footage, the accuracy dropped to near chance. But the companies kept marketing the lab results. Luna: It's almost like the Turing test for deception — but the AI is the one being tested, and it's failing. Lucas: Exactly. And yet, the market is growing. Grand View Research estimated the global deception detection market at $2.8 billion in 2025, with AI systems making up a growing share. Luna: What's the regulatory outlook? I know the EU AI Act categorizes some of these as 'high-risk' — does that apply? Lucas: It does. The EU AI Act, which came into force in 2025, classifies AI systems used for 'emotion recognition' and 'deception detection' in law enforcement and hiring as high-risk. That means they need conformity assessments, human oversight, and transparency. But enforcement is still ramping up. Luna: And in the US, we've seen some state-level action. California's proposed AI Accountability Act would require impact assessments for systems that make 'significant decisions' about individuals — including honesty assessments. Lucas: But those bills are still working through committees. Meanwhile, the technology is being deployed. There's a gap between the pace of deployment and the pace of regulation — which is the recurring theme of this show. Luna: It really is. And I think the fundamental question is: should we ever trust a machine to judge a human's truthfulness? Especially when the cost of a mistake is so high. Lucas: My answer is no — not with current technology. The false positive rates are too high, the science is too contested, and the consequences are too personal. We should treat these systems like we treat lie detector tests in court: inadmissible. Luna: That's a strong standard. But it makes sense — if you wouldn't let a polygraph decide a case, why let an AI? At least the polygraph operator is a human you can cross-examine. Lucas: Right. And with AI, the 'operator' often doesn't understand the model's reasoning either. So you have a black box making a judgment about your honesty, and no one can explain why. Luna: That's a good place to leave it. I think the takeaway for today is: if you encounter an AI that claims to read your truthfulness, be skeptical. And demand transparency. Lucas: Absolutely. Thanks for the conversation, Luna. Luna: Thanks, Lucas.