Reading · Truth Decay by James Macleod

Over the last couple of months, I’ve made the same complaint and observation on Substack Notes more than once: the sheer, relentless volume of AI-generated video passing itself off as real. Someone reposts a clip captioned “cute,” “impressive,” “you won’t believe this,” and in my judgement, more often than not, it’s been generated or manipulated from the ground up. In fairness, it happens a little less on Substack itself and more on the platforms I’ve already left (X and Instagram / the Meta hellscape, which I still watch my friends scrolling through like digital crack addicts. Where it strikes me that a genuinely enormous share of what’s circulating is artificial in some way.

But here’s the thing: At this point, even I’m finding it harder to call it, particularly when I'm using my phone. Six months ago I’d have maybe backed myself. Now, especially on a phone, I’m not so sure. And that’s the point of this piece — because I’ve realised the main age group of people reading Truth Decay skews older than I ever pictured when I started this thing, and if the technology is only going to get better at fooling us, the only defence any of us have left is our own judgement. It’s worth sharpening it now.

One thing I’ve noticed that still works (for now): if you can throw a suspicious video from your phone onto a big screen — a TV, a monitor, anything bigger than your palm — do it. AI has become very good at rendering faces and even text that sits front and centre, like the writing on someone’s t-shirt. It is still routinely sloppy with the details it thinks don’t matter: Background stuff that you don’t notice on your phone. A book spine on a shelf, the writing on a mug, a reflection in a window, a shadow that’s pointing the wrong way. Security researchers testing 2026-generation detection methods flag exactly this — things that look like language, but don’t actually display into real words. Lighting and shadows that behave illogically. More so if you're looking at a fake of a real person speaking what a lot of AI videos will do is make the lips look blurry. They tend to do this by making it look like it's internet breakup, the kind of thing that you see when the connection is not strong on your phone. It's quite difficult to notice without a big screen, with one it really stands out. Blow the image up and those errors get harder to miss. I wouldn’t bank on this working for much longer. Every month the gap closes. But right now, it’s one of the few free tools you’ve got.

Only a year ago, the advice for spotting a scam was straightforward: if an email or phone call feels off, verify it on a video call. That advice no longer cuts it in 2026, and not in some hypothetical, future-tense way I was gesturing at two years ago. It’s already happened, at scale, to serious money.

In January 2024, a finance employee at Arup — the London engineering firm behind the Sydney Opera House and Beijing’s Bird’s Nest stadium — received an email claiming to be from the UK-based CFO, requesting a confidential transaction. He was suspicious, so he did the sensible thing and asked for a video call to confirm. On that call, every face and every voice he saw and heard — the CFO, several colleagues he recognised — was an AI deepfake, built from publicly available footage scraped from places like LinkedIn, YouTube, and recorded earnings calls. Convinced, he made fifteen wire transfers totalling roughly $25.6 million to five Hong Kong bank accounts, all in a single day. He only found out something was wrong when he called Arup’s actual headquarters afterwards. Hong Kong police called it the first known case of scammers using deepfakes to deceive an institution this way.

That’s the corporate case. The domestic version is worse, because it targets the exact instinct we’re least able to override: a parent’s fear for their child. I wrote about this case well over a year ago, but it’s worth bringing up again now. McAfee found that a free voice-cloning tool needs only three to four seconds of audio to produce an 85% voice match — a wrong number call, your voicemail message, a voice note, a TikTok, a birthday video posted publicly is more than enough raw material. The FBI has been warning Americans that criminals now use AI to mimic the voices of family members and pressure victims into sending money, and it isn’t rare: the FBI’s 2025 Internet Crime Report broke out AI-enabled fraud for the first time and logged 22,364 complaints totalling around $893 million in losses, with voice-cloned “distress” calls alone accounting for over $5 million of that. That was last year.

One Florida woman, believing her daughter had been in a car accident and was in legal trouble, sent $15,000 to a courier the same day before realising she’d never spoken to her daughter at all. Another victim sent $5,000 after hearing what sounded exactly like her daughter pleading for help. In both cases, the money was gone.

Here’s the bit to really think about: if a parent who has heard their child’s voice every day of their life can be fooled by a clone, and the caller ID on the phone has been spoofed to match too, what hope does an employee have of spotting a fake CEO they’ve met once, on a video call, mid-workday? And the more publicly visible a person is, the worse this gets — a Prime Minister, a president, a well known newsreader, a household-name CEO has thousands more hours of footage circulating than any of us, which means a more convincing deepfake. We are, right now, at the point where telling the difference is genuinely difficult. We are probably months away from certain kinds of fakes becoming impossible to tell, at which point intuition — trained intuition — is the only tool we have left.

Deepfakes are the newest costume on an old con. Plenty of entirely legitimate-looking companies have spent years claiming to be something they weren’t, and the bill for believing them has run into the hundreds of billions.

Let’s start with Enron — the company Fortune named “America’s Most Innovative” five years running while it hid its losses in off-the-books shell entities and inflated its earnings on paper. When the accounting collapsed, so did the stock, from $90 to under a dollar in weeks, and the total losses are estimated at around $74 billion.

Remember Elizabeth Holmes? Her company Theranos raised over $700 million on a $9 billion valuation for a blood-testing device that never actually existed, it was a black box that did… Nothing. Lying Lizzie and Sunny Balwani were eventually ordered to pay $452 million in restitution. Liz and Sunny are currently holidaying at the pleasure of the U.S. federal prison service.

WeWork, at its Adam Neumann peak, burned through roughly $4 billion, most of it SoftBank’s money, on the promise of “elevating the world’s consciousness” through shared office space. Think about that for a moment: SoftBank, a Japanese multinational, mammoth of an investment holding company, and one of the largest investors in blue chips around the world, got sucked in to an investment that seemed too good to be true, because it was.

And then there’s FTX — once valued at $32 billion, collapsed within days once it emerged that more than $8 billion in customer deposits had been quietly funnelled into Sam Bankman-Fried’s hedge fund and personal spending.

I never actually wrote about this one at the time, but FTX was showing every one of these signs a full year before it hit the rocks. I’d been hunting for a crypto exchange — exchanges are, to my mind, one of the weakest points in the entire crypto chain, above and beyond the risk of the assets themselves — and FTX was the name everyone online wouldn’t shut up about. I don’t pay much attention to YouTubers shilling the latest token, but the recommendations were everywhere, relentlessly. I gave the website less than five minutes, probably closer to two or three, and it looked shady as hell despite looking genuinely slick. No information anywhere about who was actually running the place. Next to nothing on how the platform was actually secured. And plenty of the kind of buzzwords I’m about to spend the next section picking apart — impressive-sounding, meaning nothing. I walked away. Partly gut feeling, mostly the total absence of transparency, which remains one of the most reliable tells there is. Sometimes the information is technically there. The thing you actually need to check is whether it checks out.

Every one of those companies had confident, charismatic frontmen or women. They all had lawyers, PR teams, and glossy keynote decks. None of that made a word of it true. Charisma is not evidence. Sam Altman is a genuinely compelling speaker. Do I believe everything he says? Nope. Would I invest in OpenAI on the strength of his delivery alone? Fuck no!

There’s a subtler con that doesn’t even bother lying to you outright — it just says things that sound impressive and mean nothing when you actually stop and pause. Every tech keynote since the first iPhone launch runs on this fuel: “unlike anything you’ve seen before,” “we’ve never done this,” delivered about a rectangular slab of glass, metal and silicon that is, structurally, extremely similar to the one you already own. None of that is technically a lie, because it’s technically not saying anything.

Watch for things like: “up to.” “Up to 20 hours of battery life,” “up to 50% faster.” “Up to” is filling a lot of the blanks here. It’s true, even if the real number is a fraction of it, because the whole range below the ceiling counts. We know this stuff. We know that it’s the same trick as a “sale prices up to 70% off” sign where one item, somewhere, is actually 70% off, and the rest are barely discounted. Tech companies run the same play with comparisons: this year’s chip is faster than last year’s, sure — except the benchmark quietly uses a two- or three-generation-old model to make the gap look bigger, and by the time anyone notices, the comparison has quietly shifted again. But when we are interested in something, we often drop our guard.

And when the claims are outright false rather than just empty, they cost real money once someone finally checks. Red Bull agreed to pay more than $13 million in 2014 to settle a lawsuit alleging its “gives you wings” marketing and claims of superior energy and performance weren’t backed by science — the plaintiff argued the product offered no more benefit than a cup of coffee, and years later, Red Bull settled a nearly identical Canadian case for around $850,000 CAD. Volkswagen’s “clean diesel” campaign is the heavyweight example: the company installed software specifically designed to cheat emissions tests, detecting when a car was being tested and switching on pollution controls it otherwise left off. The core US settlement alone came to $14.7 billion, and the FTC later confirmed Volkswagen and Porsche had repaid more than $9.5 billion to deceived car buyers — on top of billions more in environmental penalties and dealer settlements. All of it dressed in the word “clean.”

Here’s a trick worth stealing: take any of these claims — “experts approve,” “scientifically proven,” “clean,” “natural,” “up to” — and read the sentence back with a question mark on the end. “Scientifically proven?” Proven by whom, using what study, funded by whom, replicated where? That question mark will prompt you to do the work the marketing department is hoping you don’t.

None of this works without two very human weaknesses: the fear of being the only one who doesn’t buy in, and the desire to believe something because we want it to be true.

The first is why cult recruitment, pyramid schemes, and stage-managed investment seminars all follow the same script: a charismatic speaker, a packed room full of hyped up positivity, and an unmistakable social pressure that if you’re the one person not nodding along, you’re the problem. It’s how cults begin. It’s how megachurches sustain themselves. It’s how Multilevel Marketing events conventions fill arenas. The room itself is the sales pitch.

The second is quieter and more personal. If someone promises you a product will change your life and you already want it to, you’re primed to believe them. If someone promises fast returns on an investment and you need the money, you’re even more primed — and that’s precisely when the sensible questions stop getting asked. Hopium is expensive. It gets far more expensive the moment you stop checking.

There’s one question that cuts through most of this, whether you’re looking at a video, hearing a pitch, or a call from your “boss”: what has to be true for this to be believable?

List it out. If the honest answer is “an enormous amount, several of which they can’t verify,” that’s your signal. Ask the person directly, if you have the chance — a scripted scam rarely survives an unscripted question. Hang up and call the real number you already had saved, not the one the caller just gave you. On a video call, ask them to get something, show you something, turn their head to the side or wave a hand in front of their face — deepfakes still struggle with head turns and hand occlusion in real time, and a live call that suddenly drops or glitches the moment you ask is not a coincidence. Set up a family safe word if you haven’t already — it sounds paranoid until the day it isn’t. Don’t email it, don’t save it to a notes app. Agree on it face to face, only use it or ask for it when you feel you need to.

And yes — use the AI itself. Ask it to show you the evidence that supports a claim and the evidence that challenges it. Reasoning models are, unsurprisingly, decent at reasoning, and asking the right question does a fair chunk of the legwork for you. It won’t replace checking the primary sources it gives you, and it shouldn’t — check them, but it’s a weapon currently sitting in your pocket, aimed the right way for once. Yes, it drinks a worrying amount of water to answer you, but in these cases you may prefer to sacrifice a glass of water then the contents of a bank account to somebody (who possibly isn't anybody) pretending to be your CFO.

We used to worry about scam emails and dodgy phone calls. Those haven’t gone away, they’ve gotten better. The verification method we relied on, seeing and hearing someone in real time, has been quietly defeated. We’re watching companies get caught lying about emissions, energy drinks, and blood tests for billions of dollars at a time, and we’re watching tech keynotes say increasingly little while sounding increasingly confident about it.

None of this is solvable with just hope and trust. It’s solvable, imperfectly, by slowing down: bouncing the suspicious video to a bigger screen, asking what has to be true, calling the number you already had, reading the claims back with question marks on the end, and noticing when you want something to be true a little too much to ask if it actually is. The con is very much on, they are everywhere and they are moving faster than most of us are adjusting. That gap is exactly where they operate. Closing it isn’t about becoming paranoid. It’s about deciding your own judgement is worth defending, and then actually using it.

If this held your attention, the kindest thing you can do is pass it to someone who needed to read it — particularly if they’re the type to trust a video just because it played on their phone. Truth Decay stays free to read because people share it, not because of an algorithm doing me any favours.

If you’d like these landing in your inbox rather than relying on Substack’s mood, subscribing is free and always will be.

Truth Decay runs on a pay-what-you-want model. Paid support starts at $5/month if you’re able and willing, but nobody is ever locked out of anything for not paying. It genuinely helps keep this independent.

Seen a fake that fooled you, or one that nearly did? Tell me about it below — the more real examples in the comments, the more useful this becomes for the next reader.

The Big Screen Trick: How to Catch AI Before It Catches You.

Comments