Imagine walking into a grocery store, picking up a baguette, a sandwich, maybe a drink, and just walking out. No checkout, no scanning, no line. The bill finds you later. That was the promise, and for a while it looked like the future had arrived.
And the bill sometimes took hours to arrive, which we now know is a clue.
We'll get to the clue. Daniel wrote in with a whole thing about this, and it's one of those stories that sounds like a joke until you realize it's just what happened. He wants the actual detail of the Amazon Go scandal, for anyone who missed it or never set foot in one of those stores. What did Amazon claim? What was actually going on? Did we ever get first-hand accounts or whistleblowers from those back-office workers in India? Were the shoppers' faces pixelated or blurred? And underneath all of it, the question he keeps circling: is it fair to say that what shoppers thought was an automated billing system actually involved somebody on another continent watching you deliberate over which baguette to choose for lunch?
It's the whole episode, really. Let's start with what Amazon actually promised, because the gap between the promise and the reality is the story.
December 2016. Amazon drops a promotional video for Amazon Go, and the narration is almost poetic. "What if we could weave the most advanced machine learning, computer vision, and AI into the very fabric of a store so you never had to wait in line?" That's the pitch. The world's most advanced shopping technology.
And the store itself was tiny. The original Seattle beta was about eighteen hundred square feet.
A corner shop. But the ceiling was a different matter. Over a hundred cameras, rigid item placement so the computer vision had a fighting chance, and the whole thing sold as computer vision plus sensor fusion plus deep learning. Amazon's own blog would later describe it in exactly those terms, with GAN-generated synthetic training data, millions of AI-generated synthetic images.
So they trained the model on fake shoppers in fake stores before they let real ones in.
That's the claim. The Seattle store opened to the public in January 2018, after a year of employees testing it. And for six years, the story was that you walked in, the cameras tracked you, the shelf sensors confirmed what you picked up, and a virtual cart tied to your account tallied it all up.
Then April 2024 happens.
The Information reports that roughly a thousand workers in India were reviewing transactions. And the number that made everyone sit up: about seven hundred of every thousand Just Walk Out sales required human review in 2022. Amazon's internal target was twenty to fifty per thousand.
So they missed their own target by more than an order of magnitude.
By roughly a factor of fifteen to thirty-five, depending on which end of the target you use. And that gap is the crux of the whole story, because Amazon's response was that humans only validate a small minority of visits. The Information says it was seventy percent of sales. Those two sentences cannot both be true.
Unless they're measuring different things, which is the charitable read.
It's possible. But the charitable read requires Amazon to explain what it's measuring, and they haven't. They've said associates don't run the whole thing, they annotate data and validate a small minority of visits when the AI can't determine a purchase. Okay. What percentage is a small minority? They won't say. The Information says seventy percent. That's the shape of the dispute.
So what was actually happening inside those stores? Let's get into the mechanism, both the one Amazon described and the one The Information found.
The claimed mechanism is elegant, and I want to give it its due because the engineering is not stupid. You walk in, you scan a code, that code ties your account to a virtual cart. Ceiling cameras track you as a person moving through the space. Shelf sensors, which are essentially weight sensors, register when an item leaves or returns to a shelf. Computer vision identifies what the item was. The system reconciles all of that into a receipt. On exit, the temporary numeric code disappears. Amazon's blog is explicit that Just Walk Out does not use or collect biometric information, and that the code is temporary.
No facial recognition, in other words.
That's what they say. The code is the identity link, not your face. Which matters for the pixelation question, and we'll come back to it.
What did The Information say was actually happening?
That a thousand people in India were doing the work the cameras couldn't. Reviewing what customers picked up, set down, walked out with. And the receipts-latency detail is the one I find most damning, because it's a symptom you can observe from outside. The Information reported that the reliance on backup humans explains in part why it could take hours for customers to receive receipts.
If the system were fully automated, why the delay?
A computer vision system that identifies your items in real time produces a receipt in real time. A system that queues your transaction for a human reviewer in another time zone produces a receipt when that reviewer gets to it. Hours is not a mystery. Hours is a queue.
That's the tell. You don't need a whistleblower to notice a queue. You just need a stopwatch.
And people did notice. There was a lot of grumbling about receipts arriving long after people had left the store, and at the time it was attributed to processing. Now we have a different explanation.
Let's do Amazon's rebuttal properly, because they pushed back hard and it deserves to be stated accurately.
April seventeenth, 2024. Dilip Kumar, who's a VP of AWS Applications, publishes a blog post calling the reports untrue and erroneous. His line: associates don't watch live video of shoppers to generate receipts, that's taken care of automatically by the computer vision algorithms. Amazon tells USA Today that the framing of human reviewers watching shoppers live from India is misleading and inaccurate, and that associates validate a small portion of shopping visits by reviewing recorded video clips. And they tell Gizmodo that associates validate a small minority of visits when the AI can't determine a purchase.
So there are two distinct claims in there. One is about the volume of human review. The other is about whether it's live or recorded.
Right, and they get conflated constantly. On volume, Amazon says small minority, The Information says seventy percent. Unresolved, though the burden is arguably on Amazon to give a number if they want to dispute one.
On live versus recorded?
Amazon is unambiguous. Recorded clips, not live feeds. And this is where I have to be careful, because The Information's framing and a lot of the coverage implies near-real-time human involvement, but The Information's actual reporting, as far as I can tell, doesn't nail down whether reviewers were watching live. The receipts-latency point suggests asynchronous review, which actually supports Amazon's version. If a human were watching live, the receipt would be instant.
So the hours-long delay is evidence for the recorded-clips story.
It's evidence for queued review, which is consistent with recorded clips. It doesn't prove it, but it points that way. And I'll say plainly: I don't know for certain, and neither does anyone outside Amazon.
Which brings us to the pixelation question. Did the reviewers see faces?
This is the one where I have to be honest about what we don't know. Amazon says Just Walk Out doesn't collect biometric information and links shoppers only to a payment method via a temporary numeric code. That's their privacy claim. There is a relevant Amazon patent, US Patent 11600094, titled Obfuscating portions of video data. It describes obfuscating portions of video in real time using pixilation and other techniques so that a reviewing agent cannot identify people.
That sounds like exactly the answer.
It sounds like it, and it isn't. The patent concerns fulfillment center operators, not grocery shoppers. It's about protecting warehouse workers from being identified by reviewers, not about protecting shoppers in a store. So it's suggestive of how Amazon thinks about this problem, but it does not confirm that grocery shoppers' faces were blurred for the India reviewers.
So the honest answer to Daniel's question is: we don't know.
No source I've seen confirms or denies it. Amazon says no biometrics, which is a different claim from no faces visible. You can collect no biometric data and still have a human looking at an unblurred face. Those are separate things, and Amazon's statements address the first, not the second.
That's a real gap. And it's the kind of gap that a company could close with one sentence, and hasn't.
Correct.
Now the whistleblower question, which is the other half of what Daniel asked.
The Information's reporting relied on an unnamed person who has worked on Just Walk Out technology. One anonymous source. Not a named India-based back-office worker. Gizmodo cited a senior team member who was let go, who said nearly all the Just Walk Out engineers were laid off, leaving a skeleton crew. That's an engineer, not a reviewer.
So the thousand-workers figure traces back to a single anonymous source plus Amazon's own acknowledgment that human moderators exist.
That's right. And I want to be precise here, because this is where the story can get overclaimed. Amazon has never denied that humans are involved. They've denied the scale and the live framing. The thousand number is The Information's, from one source, and Amazon hasn't confirmed it. It's credible reporting from a publication with a good track record, but it is not a named whistleblower with a first-person account.
No one from that back office has ever spoken on the record.
Not that I can find. Which is itself notable. A thousand people doing that work, and not one named account. That tells you something about the employment structure, the NDAs, the geography, the power imbalance. It's a workforce that is structurally hard to hear from.
That's the mechanism. Now let's talk about why this story matters beyond Amazon, because it's not a one-off.
It's a genre. There's a joke in tech circles, the "Actually Indians" meme, which started as a riff on the "Actually Finnish" developer joke and got repurposed for this. The pattern is: company claims AI, actually there are people. Amazon Go is the famous case, but there's also Nate, the shopping app that raised a lot of money on the claim that its AI completed purchases for you, and which allegedly involved humans in the Philippines doing the ordering. And there was a fintech app found to be humans in the Philippines as well.
So the pattern has a name now. AI-washing.
Parmy Olson put it well in the Straits Times. She said there's a grey area in artificial intelligence filled with millions of humans who work in secret, often hired to train algorithms but ending up operating much of their work instead. That's the structural insight. The humans aren't a bug in the system. They're load-bearing.
Which is a different thing from fraud, and I want to poke at where that line actually is.
It's the right question. Almost every AI application has human fallback. Waymo has human operators who can teleoperate its vehicles. That's not fraud, that's a safety layer, and it's disclosed. The difference with Nate, allegedly, is that the CEO claimed the system worked without human intervention. That's a factual claim about the mechanism, and if it's false, that's fraud.
And Amazon?
Amazon never claimed zero humans. They said associates annotate data and validate a small minority of visits. So the fraud case is weaker. But the marketing case is strong, because the marketing never mentioned the India contractors at all. You watched that 2016 video, you read the blog posts about computer vision and sensor fusion and deep learning, and you would not come away knowing that a thousand people in India were reviewing transactions. The humans were disclosed in the fine print of a rebuttal, not in the pitch.
The pitch is the product.
The pitch is the product. James Bridle in The Guardian was blunter: this is how these bosses get rich, by hiding underpaid, unrecognised human work behind the trappings of technology. And the economics back that up. Cheap offshore labour is what made the unit economics work. The AI framing is what made the valuation and the press work. Both are doing labour in that sentence, and only one of them is doing it honestly.
So the question isn't whether humans were involved. It's whether the gap between the claim and the reality was material to how the thing was sold, valued, and covered.
And the answer is yes, obviously, because the entire press cycle was about the technology. Nobody wrote about the labour until The Information did in 2023 and then again in 2024. Six years of coverage about computer vision, and the humans show up in year seven.
What happened to the stores?
They're gone. April 2024, Amazon removes Just Walk Out from many of its Amazon Fresh stores and replaces it with Dash Cart smart carts, which are just fancy shopping carts with scanners. Gizmodo reports Amazon killed most internal Just Walk Out development, laying off nearly all the dedicated engineers. And then in January 2026, Amazon announces it's closing its Amazon Go and Amazon Fresh physical stores entirely and converting various locations to Whole Foods Market stores.
So the format is dead.
The format is dead, and here's the part that gets me. There was never a headline admission that the tech didn't work. No mea culpa, no "we overpromised." The stores just quietly became Whole Foods. The story ends with a conversion, not a confession.
But the technology lives on.
It does. Amazon still sells Just Walk Out to third parties. By 2024 there were a hundred and forty-plus third-party locations, sports stadiums, airports, a university. Lumen Field, where the Seahawks play, reported an eighty-five percent increase in transactions and a hundred and twelve percent increase in sales per game with Just Walk Out. That's the pitch to the next buyer. And the next buyer is buying the same thing, which means the same hidden labour is presumably part of the package.
So the question for the next operator is: did you buy the cameras, or did you buy the queue?
You bought the queue. You just don't know it yet.
There's something else in here, and it's the thing I keep circling. The story is usually told as a tech story. It's actually a labour story wearing a tech costume.
Say more.
The cameras are real. The computer vision is real. The sensor fusion is real. None of that is fake. What's fake is the implication that the system is autonomous. And that implication is what lets a company book the labour savings, claim the innovation premium, and skip the awkward conversation about who's actually doing the work at two in the morning.
The labour didn't disappear. It moved somewhere the org chart doesn't show and the press release doesn't mention.
And the customer never knew. That's the part that makes it a scandal rather than just a business decision. You walked out of that store feeling like you'd used the future. Somebody in another country was looking at your shopping.
Or a recording of it. We don't know which, and that's the point.
We don't know which.
Let's pull back and look at what we still don't know.
We don't have a named whistleblower from the India back office. We don't know for certain whether shoppers' faces were pixelated for the reviewers. The live-versus-recorded dispute is unresolved, though the receipts-latency evidence points toward recorded. And Amazon closed the stores without ever fully accounting for the gap between the claim and the reality.
The number that started all this, seven hundred per thousand, is still just The Information's number against Amazon's refusal to give one.
Which means the most important fact in the story is a number nobody has confirmed and nobody has denied with a counter-number.
That's where it sits. And as AI-washing becomes a recognized pattern, with Nate and the Philippines fintech app and Amazon Go all in the same bucket, the question for the next case is whether the marketing or the labour will change first.
The technology lives on in third-party stores. The workers are still there, somewhere, watching.
Somewhere.
That's a hell of a place to leave it, but there's one more thing.
Hilbert: Does it matter whether it was live?
Hilbert.
Hilbert: I ask because everyone keeps asking that. Live or recorded. And I spent most of a year doing that kind of work, so I have a view.
Go on.
Hilbert: It was a contract firm. Retail analytics, they called it. Night shift. I sat in a room with about forty other people and watched clips of people in stores. Flag the item, move to the next one. Queue never emptied. Never once did anyone tell me what the footage was for or who was on the other end of it.
And the live question?
Hilbert: It doesn't matter. That's my point. Live or recorded, the job is the same. You're watching a stranger shop. You're making a judgment call about what they picked up. You're doing it at a pace set by a queue that never stops. Whether the footage is thirty seconds old or six hours old changes nothing about the work. It changes what the company can say in a press release.
It's a PR distinction, not a labour distinction.
Hilbert: That's what I said. The training manual had a section on privacy compliance. Told us faces would be blurred. In practice the blurring was inconsistent. You learned to recognize the regulars anyway. Not by face. By gait, by jacket, by the way they shopped. Never knew their names. Knew their habits.
So the pixelation question has an answer, at least in your case.
Hilbert: In my case. Different company, different contract. But the manual said blurred and the reality was sometimes blurred. That's the honest version.
And the manual called it data annotation.
Hilbert: Data annotation to improve the AI. That's the phrase. But the queue never emptied, and the volume suggested the AI wasn't improving much. You don't need a thousand people on the night shift if the model is learning. You need a thousand people on the night shift if the model isn't.
That's a hell of a detail.
Hilbert: Someone's waiting for me in the car park. She's been there a while.
Let's pull back and look at what we still don't know.
We still don't have a named whistleblower from the India back office. We still don't know for certain whether shoppers' faces were pixelated for Amazon's reviewers. The live-versus-recorded dispute is unresolved. And Amazon closed the stores without ever fully accounting for the gap between the claim and the reality.
The most common wrong belief about this whole story is that Amazon Go was a fully automated system that just failed commercially.
It wasn't fully automated. Roughly a thousand people in India were reviewing transactions, and by The Information's count, seven hundred of every thousand sales needed a human in 2022. The automation was the pitch. The humans were the product.
The technology lives on in third-party stores. The workers are still there, somewhere, watching.
Somewhere.
One open question to leave you with. As AI-washing becomes a recognized pattern, the question for the next case is whether the marketing changes first or the labour does. My guess is the marketing gets better at hiding the labour, and the labour stays exactly where it is.
Which is the least comforting possible answer, and probably the right one.
Thanks as always to Hilbert Flumingtop, who produces this show and apparently has a life running in parallel to it.
This has been My Weird Prompts.
If you want to tell us what you think about this one, email us at show at my weird prompts dot com. We read them.
We'll be back soon.
See you then.
