Mato just raised pre-seed Read more

The $55 Billion Hallucination: AI's Fraud Genre

  • Jul 1, 2026
  • 16 min

Show notes

What the episode covers

Nate app founder Albert Saniger claimed his shopping bot was 93 to 97 percent automated by AI, when the real figure was reportedly closer to zero, with actual orders allegedly processed by humans typing behind the scenes. On this episode of Vaporware, Derek and Grant use that case as a launchpad to ask a bigger question: is Nate a one-off con, or the first documented instance of a repeatable AI fraud playbook?

The hosts build out a three-part taxonomy of AI fraud, unpack Cluely CEO Roy Lee's public admission that his viral $7 million ARR figure was inflated by roughly 35 percent ahead of a $15 million funding round, and break down the specific accounting tricks used to dress up thin numbers as explosive growth. They also dig into the SEC's new Cyber and Emerging Technologies Unit and why regulators tend to show up only after investors have already lost money.

Next episode, the hosts follow the money further, asking whether AI companies can even afford to charge what their products actually cost to run.

Timeline

In this episode

8 moments worth skipping to. The timecodes match the player above.

  1. 0:15Introduction
  2. 1:40Recall: The Zero-Percent Automation Rate
  3. 4:04A Taxonomy of AI Lies
  4. 6:12Cluely: The Confession Nobody Expected
  5. 8:23The Three Tricks: How ARR Gets Faked Without Technically Lying
  6. 10:29The SEC Finally Shows Up
  7. 12:52The Real Bubble Nobody's Prosecuting
  8. 14:58Outro

Quick answers

Straight from the episode

The questions this one settles, without the listen.

What was the real automation rate behind the Nate App, according to the episode?
Derek reveals that despite Saniger's claims of 93-97% automation, the actual automation rate was allegedly zero percent. Saniger reportedly restricted employee access to dashboards showing real numbers, treating them as a trade secret.
What are the three types of AI fraud identified in the episode?
Derek lays out a taxonomy of humans-behind-the-curtain schemes, benchmark cherry-picking, and AI-washed metrics, arguing that AI's definitional fuzziness lets vague claims survive scrutiny longer than in other industries.
Why did Cluely CEO Roy Lee admit his $7M ARR claim was inflated?
Roy Lee publicly admitted his viral $7M ARR figure was inflated by about 35%, which had anchored a $15M a16z round at a $120M valuation. Bloomberg reportedly found a similar pattern across nine other startups, and Lee also admitted to lying about how the original interview happened.
What accounting tricks did Roy Lee use to inflate Cluely's revenue?
Derek breaks down three individually-legal tricks: annualizing a single best month of revenue, counting unsigned pipeline as revenue, and front-loading prepaid contracts to inflate the ARR figure.
Has the SEC created a unit specifically for AI-related fraud?
Yes, the SEC has established a Cyber and Emerging Technologies Unit, with the Nate App case serving as its flagship prosecution. Grant notes enforcement is structurally reactive, only arriving after investor money is already gone.
How does the episode connect AI fraud to AI's broader pricing problems?
Derek cites a SemiAnalysis stress test showing steep gaps between AI subscription prices and actual compute costs, arguing that capability fraud and affordability/pricing fraud are essentially the same underlying bet on future AI pricing power.

Transcript

The full conversation

Every word of the episode, 2,396 of them, in the order they were said.

Read the transcriptHide the transcript

DerekThanks for watching!

GrantWelcome back to Vaporware. I'm Derek, that's Grant, and dude, we are back on the AI grift beat.

Speaker 3We never really left it, if we're honest.

GrantFair. Okay, so get this. Remember the Nate app, the shopping bot that was supposedly 93 to 97% automated?

Speaker 3The one where it was allegedly 0% automated and just guys in an office typing really fast?

GrantAllegedly, yes. We're reopening that file today because it turns out... out it might not be one weird case.

Speaker 3Wait, you think this is bigger than one company lying?

GrantThat's the question driving the whole episode: is Nate one-off, or is it Exhibit A in an entire category of fraud?

Speaker 3Okay, I need receipts for that claim.

GrantOh, you're going to get them. We're talking about a viral CEO admission over inflated revenue, a brand new SEC unit built specifically to chase this stuff.

Speaker 3Hold on, there's a whole task force now?

GrantThere is, and then, get this, we're following the money all the way down to whether AI companies can even charge with things actually cost.

Speaker 3Oh good, so it's not just fraud, it's fraud on top of a math problem.

GrantBasically, it's a fun day at the office.

Speaker 3I've got my spreadsheet brain ready. Let's see if any of this actually adds up.

GrantSwitching gears, back to April 2026, the month everything unraveled at once.

Speaker 3Alright, let's find out how deep this thing goes.

GrantOkay, quick refresher before we go anywhere new. Albert Saniger, Nate App-ring a bell?

Speaker 3The "AI Does Your Checkout For You" app.

GrantYeah.

Speaker 3I remember.

GrantRight, so he told investors his app had a ninety three to ninety seven per cent automation rate-basically robots buying your stuff for you, no humans involved.

Speaker 3And that was a lie.

GrantWait for it-the DOJ and SEC alleged the real number was effectively ZERO! Zero per cent.

Speaker 3Zero? Not lower than claimed-zero?

GrantZero-every single purchase done manually by contractors in the Philippines, hundreds of them.

Speaker 3So there's no AI in the AI shopping app?

GrantThere's a UI And behind the UI A call center. That's the product.

Speaker 3Okay, that's a hell of a gap-ninety three to ninety seven percent versus zero? That's not rounding error, that's a different company.

GrantAnd here's the part that gets me every time: Saniger allegedly restricted employee access to the internal dashboard that tracked the real automation numbers.

Speaker 3Wait back up. His own employees couldn't see it?

GrantCalled it a trade secret.

Speaker 3Wow.

GrantNot commercially sensitive-a trade secret, which conveniently means nobody inside the company could hold him accountable either.

Speaker 3So he wasn't just lying to investors, he was lying to the people building the thing.

GrantExactly-and they didn't know they were the punchline.

Speaker 3Okay, so we've covered Nate before. Why are we back here?

GrantBecause I keep asking myself one question: was this a one off, one guy, one bad app, one very specific lie?

Speaker 3Or-

GrantOr is Nate actually the first documented case of something bigger, a whole category of fraud that's about to industrialize?

Speaker 3You think there's a pattern here-like a repeatable playbook?

GrantI think humans pretending to be AI is just one flavor, and once you start looking for it...

Speaker 3You find more flavors.

GrantYeah, so what if Nate isn't the scandal? What if Nate is actually the instruction manual?

Speaker 3All right, now I'm nervous. What are the other flavors?

GrantBuilding on that, let's name the genre. There are three flavors of AI fraud, and Nate is just flavor one.

Speaker 3Okay, lay it out for me.

GrantFlavor one, humans behind the curtain. That's Nate, a person in a back room clicking buy while the pitch deck says neural network.

Speaker 3Sure,

GrantFlavor two, benchmark cherry picking. You rig the test so your model wins, like showing off a car that only drives downhill.

Speaker 3Technically it drove.

GrantExactly, and Flavor three, AI-washed Metrics—you don't lie about the tech, you lie about what the tech is doing for your revenue.

Speaker 3Wait, how is that different from just lying about revenue? Companies have inflated numbers forever.

GrantBecause AI-driven growth is unfalsifiable in a way that we sold more widgets isn't. Nobody agrees on what counts as AI, so the claim survives longer before anyone can even test it. Test it.

Speaker 3Ha! so it's fraud with plausible deniability built in.

GrantThat's the mechanism—and it's not just my theory; legal analysts tracking securities class actions say filings over alleged AI misrepresentations roughly doubled between twenty twenty-three and twenty twenty-four.

Speaker 3Doubled—in one year?

GrantRoughly. Yeah; same ambiguity scaled across an entire industry.

Speaker 3So the fuzziness isn't a bug people stumble into.

GrantIt's the exploit.

Speaker 3Wow.

GrantEvery case this episode gets sorted into one of these three buckets, sometimes all three at once.

Speaker 3Which is a very organized way of saying everybody's cheating differently.

GrantI prefer taxonomy. Sounds more respectable in a script.

Speaker 3Noted. So which bucket are we opening next?

GrantBucket three, the AI-washed metrics one, and we've got a case that's basically still warm.

Speaker 3Warm like recent?

GrantRecent like the CEO was posting corrections on X a few months ago.

Speaker 3Ugh! I already don't trust this.

GrantYou shouldn't. A viral ARR claim, a nine figure valuation, and a number that moved by millions depending on who was asking. Now flip that on its head, because Cluely isn't hiding behind humans or benchmarks. This one's a straight-up confession.

Speaker 3Wait, a confession, like willingly,

GrantSummer 2025, CEO Roy Lee tells a TechCrunch reporter that ARR doubled in a week to $7 million. That anchors a $15 million Series A from Andreessen Horowitz, company's valued at $120 million.

Speaker 3Doubled in a week? Nobody at A16Z pulled up the Stripe dashboard on that.

GrantApparently not before wiring the check. And then, get this, March 2026, Lee posts on X admitting the real number was around 5.2 million. TechCrunch reported he also misrepresented how that original interview

DerekWhat even happened?

GrantSo he lied about the lie—that's a special kind of chaos.

DerekRight? Roughly a thirty-five percent gap between the headline and reality, according to TechCrunch's reporting on his admission.

GrantOkay, but here's my actual question: why confess? Nobody was holding a gun to his head.

DerekUnclear, but it didn't stay isolated. Bloomberg's follow-up reporting reportedly named nine other startups with similar Similar gaps between claimed and verified revenue.

GrantNine; so Cluely's just the one who got caught talking.

DerekPretty much; and the wild part is, Cluely's actual business wasn't fake; they built a real growth machine; deliberately provocative viral content, an army of clippers pushing it everywhere.

GrantWait, wait; so the company works, the growth is real, and he still lied about the number?

DerekYeah, that's the part that gets me; he didn't need to invent a company. He invented a decimal point.

GrantThat's almost worse-that's a founder choosing the lie when the truth might have been fine.

DerekMight have been. We don't know if five point two gets you the same valuation.

GrantProbably not the same headline, though.

DerekNo, definitely not the same headline, and that's the thing-the mechanics of how he actually inflated that number, the specific tricks.

GrantOh, there's tricks-plural?

DerekThree of them, and none are illegal by themselves.

Speaker 3you

DerekSo, for the non finance folks in the audience, let's actually open the hood on how you inflate a ARR without forging a single document.

GrantPlease—I need this in plain English.

DerekTrick one: annualizing the best single month by multiplying it by twelve. You have one great month, you times it by twelve, and boom, that's your run rate.

GrantThat ignores every bad month you ever had.

DerekEvery one. Trick two: counting pipeline. Deals that are committed but unsigned as if the ink's already dry.

GrantWait, so a handshake counts as revenue now?

DerekApparently. In trick three, the sneaky one, you get a client to prepay for a full year and you book all 12 months as ARR on day one instead of spreading it out.

GrantSo three tricks, zero fake invoices.

DerekZero. ThePlanetToolsai reported that's the exact. Exact combination Roy Lee laid out in his own confession thread. None of it's illegal on its own.

GrantWhich is exactly why it spreads, right? Nobody's breaking a law; they're just optimistic.

DerekOptimistic accounting. Now here's the part that should worry you. This is one end of the spectrum.

GrantThere's a worse end?

DerekOh, way worse. TechCrunch's coverage of a separate AI startup case described investors paying roughly thirty-five times a revenue figure that turned out to be fabricated.

GrantThirty-five times a number that wasn't even real?

DerekNot massaged—invented. That's the criminal end. Cluely leads the gray end.

GrantSo one guy fudges the math and gets a bridge round. Another guy fakes the number outright and gets indicted.

DerekExactly the spectrum. And ARR itself isn't a bad metric. The problem is contract definitions leave enough wiggle room that a founder can sh

Speaker 4-

Derekcan shape the figure without ever touching a fake invoice.

GrantSo the tool's fine, it's the discretion built into it that's the loaded gun.

DerekAnd loaded guns tend to get regulators interested eventually.

GrantPlease tell me somebody's finally aiming at this.

DerekSo shifting to the regulators, turns out the SEC actually built a unit for this exact genre of fraud.

GrantWait, an actual dedicated unit, not just like a task force nobody funds?

DerekThe Cyber and Emerging Technologies Unit, launched February 2025, built specifically to go after securities fraud that uses blockchain, AI automation, basically the vehicle, not the crime itself.

GrantOkay, in their headline cases...

DerekNate, same month we opened the... Open the episode with April 2025, SEC and DOJ file parallel charges against Saniger, alleging he fraudulently raised over $42 million on false AI claims.

Grant42 million. So this is the flagship case. Like this is the one they point to when they say, see, the unit works.

DerekExactly the one. And their fiscal year 2025 results lump it together. AI, crypto, cybersecurity. All filed under emerging technology fraud categories.

GrantWhich honestly makes sense, same play book, different buzz word.

DerekRight. But here's the uncomfortable part.

GrantGo on.

DerekHere's the uncomfortable part. Every one of these-Nate, Presto Automation, the whole growing docket-became public only after the money was already specked or a whistleblower forced it out.

GrantSo the SEC's basically showing up to the house fire with a hose. after the house is a foundation.

DerekThat's not far off.

GrantI mean the math doesn't work and everyone kind of knows it. You can't enforce your way into catching something before the cash is going. . .

Speaker 4.

GrantEnforcement's a lagging indicator by design.

DerekIt's reactive by nature-the charges are the receipt not the alarm.

GrantGreat system. Really airtight.

DerekLook, it's something-it's more than we had two years ago.

GrantSure, but it means for every Nate that gets caught . . .

DerekThere's probably three more running the exact

Speaker 4same pattern.

Derekexact same play right now quietly.

GrantComforting.

DerekAnd that's just the broadside.

GrantHmm.

DerekThere's a whole other layer to this the SEC literally cannot touch.

GrantWait, what do you mean cannot touch? Like legally can't or just doesn't?

DerekBecause it's not illegal, it's just the entire economics of AI itself might not add up, even for the honest companies.

GrantHold on, you're saying this isn't just a fraud problem anymore?

DerekNope. Wait for it, because what's coming next makes Nate look almost quaint. It was quaint. Shifting from enforcement to the bigger number nobody's charging anyone for, there's a stress test making the rounds. SemiAnalysis bought every tier OpenAI and Anthropic sell, ran them into the ground with coding agents, and checked what it cost at straight API rates.

GrantAnd?

DerekThe maxed out $200 ChatGPT Pro plan can represent something like $14,000 in compute, Claude's top tier lands near $8,000.

GrantWait, fourteen thousand for two hundred dollars?

DerekThat's the ceiling for the heaviest users, sure, but the gap is the point. Every subscription is a bet that you'll get hooked before the company has to charge what it actually costs.

GrantSo it's the AI playbook just with real math instead of fake automation?

DerekKind of, yeah. DSHRs blog on AI's affordability crisis calls it the drug dealer's algorithm: subsidize now, addict later. Hope pricing power shows up before the venture money runs out.

GrantSo, and that number keeps growing, right? I remember early estimates of the industry's revenue gap being smaller.

DerekAnalysts have revised the gap upward pretty much every few months. For two years now nobody's landed on a final figure-that's the tell.

GrantBecause if you can't even agree on the size of a hole-!

DerekYou definitely can't agree on who's lying about filling it. That's the connective tissue for this whole episode: Klieg inflated a revenue line, Nate inflated an automation rate, the industry as a whole is inflating a business model.

GrantDifferent lines, same bet-look profitable, look smart before the bill lands.

DerekExactly-and it works because nobody's agreed on how to measure it. As your AI in the first place; no shared yard stick means no shared definition of fraud (which is a hell of a loophole to build a trillion dollar industry on). When it's new enough that nobody agrees on the ruler, that's exactly when the con gets its cover.

GrantSo watch the subscriptions, watch the ARR slides, and watch what happens the day the free tokens stop being free. That's the whole show, really; the math doesn't work, and everyone with a calculator already knows it.

DerekYeah, so that's Nate, and honestly, the trade secret line still gets me, hiding a zero behind a lawyer's phrase.

GrantRight, and the part that sticks with me isn't the fraud, it's how long nobody asked to see the dashboard.

DerekWhich is kind of the whole episode, vaporware thrives on vague claims because there's no shared way to check them.

GrantThat's the pattern across every deal we've covered: the check comes after the money's already gone.

DerekSo if you take one thing from today, ask for the receipts. Receipts before you're impressed by the pitch.

GrantOr at least ask if the dashboard's still turned on.

DerekFair. All right, if this one got under your skin, send it to a friend who still believes every deck they read.

GrantSubscribe, leave us a review-it actually helps.

DerekWe'll be back next week with another one of these.

GrantThanks for hanging out with us.

DerekSee you then.

More episodes

Keep listening

Other episodes of Vaporware, newest first.

All episodes of Vaporware

Sources

Where this came from

35 reports behind the episode. Every one of them opens where it was published.