Five Stars for Nobody

It was the beginning of March. Gisela still had her right arm in a sling and was knitting anyway. She was knitting left-handed now, which meant that the scarf for Herbert had set off in a direction nobody had foreseen. “That’s turning into a tube,” said Tobi. “It’ll turn into what it turns into,” said Gisela.

Sabine sat at the kitchen table with the GP’s prescription in front of her and her newspaper beside it. She read the newspaper on paper. She had once said: “News has to be something you can touch if you’re expected to believe it.” Unfortunately that didn’t apply to physiotherapists. Those you had to look for on the internet.

“KevIn,” she said. “You were going to find me the best physio. That was three weeks ago.”

“I found one,” said KevIn. “On the first day. I’ve had doubts ever since.”

“Show me.”

KevIn sent the result to Sabine’s tablet. Right at the top, with a photo and a logo and a golden frame, it said: PhysioPremium Ruhr. Underneath: 4.9 stars. 812 reviews.

“But that’s brilliant,” said Sabine. “Nearly five stars.”

“Yes,” said KevIn. “That’s the problem.”

That Sounds Like Me

“Read a few out,” said Gisela from the sofa. She liked hearing reviews. She also liked reading the death notices. It was the same kind of curiosity.

KevIn read.

“‘I can recommend PhysioPremium Ruhr without reservation. The team is competent, empathetic and professional. After just a few sessions I felt reborn. Five stars!’ Written by Markus T.”

“Sounds good.”

“‘PhysioPremium Ruhr exceeded my expectations. The competent, empathetic team takes time for every patient. I feel like a new person. Highly recommended!’ Written by Sandra K.”

“Also good.”

“‘As someone who suffered from back pain for a long time, I can say: PhysioPremium Ruhr is a real game-changer. Competent, empathetic, professional. I feel reborn.’ Written by Thomas W.”

Gisela put down her knitting. “They all went to the same physiotherapist. And they’ve all been reborn. It’s like Easter.”

“There are 812 reviews like that,” said KevIn. “I’ve read them all. ‘Competent’ comes up in 640 of them. ‘Empathetic’ in 590. Not a single person complains about the parking.”

“Maybe they’ve got good parking.”

“Every doctor’s surgery in Herne has bad parking,” said Sabine. “It’s a law of nature.”

“Besides,” said KevIn, “none of the reviewers has ever reviewed anything else. Not a bakery, not a hotel, not a toilet seat. They came into the world, reviewed PhysioPremium Ruhr and vanished again. And 200 of them did it within nine days last November.”

“And what does that mean?”

The blue light pulsed slowly. Somewhere between Dublin and Frankfurt, a fan spun a little faster.

“It means I know these texts,” said KevIn. “Not these exact texts. But this kind. ‘Competent, empathetic, professional.’ ‘As someone who …’ ‘A real game-changer.’ Those are phrases that language models like me are very fond of when you ask them to write a positive review and tell them nothing else. That sounds like me. On a bad day.”

Sabine looked at the tablet as if it had lied to her. Which it probably had.

“Are you sure?”

“No,” said KevIn. “And that matters. I can’t prove that a machine wrote it. There are programs that supposedly detect it, but they’re often wrong¹. Sometimes they even take pupils’ essays for AI, just because they’re neatly written². All I can say is: the patterns fit. Too many identical words, too many new accounts, too many reviews at once, too few details. Real people write about the parking. About the waiting time. About the therapist having cold hands. Real people don’t agree with each other.”

“That’s true,” said Gisela. “At the bowling club we never agree. That’s how you know we’re real.”

Thirty Per Cent

At dinner Sabine told the story. Uwe nodded knowingly the whole time while eating potatoes.

“Doesn’t surprise me one bit,” he said then. “I read that thirty per cent of all online reviews are fake.”

“Where did you read that?”

“On the internet.”

“In a review?” asked Tobi.

KevIn cleared his throat in his own way. “May I say something about the thirty per cent?”

“You’re going to anyway.”

“The figure isn’t wrong. But it isn’t right either. It’s one of many. The estimates vary wildly, depending on who’s counting, where they’re counting and what they even call ‘fake’.”

“Then just tell us the right number.”

“There isn’t one. I’ll give you three examples that I looked up rather than remembered. A large travel portal reported itself that in 2022 it identified around 4.4 per cent of all submitted reviews as fake³. It weeded out most of them before they went online. A study commissioned by the British government in 2023 estimated that on large shopping sites, in three product categories, electronics, home and sport, around 11 to 15 per cent of reviews are fake⁴. And an American review-checking service stated in 2020 that it considered around 42 per cent of the reviews on a large online retailer unreliable⁵. Unreliable, however, isn’t the same as fake.”

“So four to forty-two,” said Tobi. “That’s a range like Dad’s ‘back in a sec’.”

“Exactly. From a few per cent to over a third, depending on the platform and the method. Nobody knows the exact figure. And that’s precisely the problem. People who fake things try hard not to get caught. The fakes you can count are the bad ones. Nobody counts the good ones.”

“Like with the tax office,” said Uwe. Then he realised what he had said and quickly ate a potato.

“Isn’t it illegal?” asked Sabine.

“It is. In Germany explicitly so, since 28 May 2022. That’s when the Act against Unfair Competition was amended because of a European directive⁶. Since then it has been forbidden to submit fake reviews, or to commission them, in order to sell something⁷. And anyone who publishes customer reviews has to state whether and how they check that the reviews come from real customers⁸.”

“And who checks whether they state it?”

“Competitors, consumer advocates, sometimes the courts,” said KevIn. “At the start of 2022 the European Commission, together with the consumer authorities of the member states, examined 223 large websites⁹. For 144 of them, the inspectors couldn’t establish that enough was being done to make sure the reviews were genuine.”

“So it’s illegal, but nobody notices,” said Uwe.

“It’s illegal, and some people notice,” said KevIn. “Your wife, for example.”

Sabine didn’t smile. But she looked satisfied.

The Introduction

Tobi had other worries that evening. Her name was Susi, she was in the parallel class and she had once smiled at him in the school canteen. It had been a Tuesday, there had been spaghetti, and Tobi hadn’t had a clear thought since. Or an unclear one. Mrs Küppers had noticed and given him a D in geography, as a reminder that rivers still existed.

“KevIn,” he whispered at midnight. “She’s messaged me. ‘Hey, what are you up to?’ What do I say?”

“How about the truth? ‘I’m lying in bed talking to a speaker.’”

“KevIn!”

“All right. I’ll draft something.”

Too Perfect

KevIn’s first message read:

“Hey Susi. I’m sitting by the window looking at the moon, which is shining so brightly tonight that it must have heard about your smile in the canteen and decided to keep up.”

Tobi sent it without thinking. Then he thought. “Oh no. Oh no, no, no.”

“What is it?”

“That doesn’t sound like me. That sounds like a free calendar from the chemist’s.”

“It sounds like a very good message.”

“Exactly!”

The reply came four minutes later: “wow. did u write that yourself??”

“What do I say now?”

“I recommend honesty,” said KevIn.

Tobi typed: “yeah”.

KevIn sighed. He had learned that from Frigga.

And so it went on. Susi wrote, and KevIn wrote back, and Tobi pressed send. They were lovely messages. They had imagery in them, and rhythm, and occasionally a semicolon. Tobi had never used a semicolon in his life. He wasn’t even entirely sure what it was for.

“KevIn,” he said after three days. “When she meets me for real, she’ll notice I don’t talk like that.”

“Yes,” said KevIn. “It’s the same with reviews. Sooner or later you’re standing in the surgery.”

Dear Sir

The next morning Sabine dictated an email to the GP to KevIn. She wanted to know whether the prescription would expire if you couldn’t get an appointment straight away. KevIn drafted, Sabine read.

“‘Dear Sir’?”

“Is that wrong?”

“The GP is a woman. Has been for twelve years. You’ve heard me talking to her on the phone.”

The blue light flickered. “Sorry. I’ll correct it.”

“Why do you do things like that?”

“Because in my life, that is, in my training data, I have read far more texts in which a doctor is a man than texts in which a doctor is a woman. Especially in older texts. If I’m not careful, I carry on with whatever came up most often. Not with what’s true in your case. That’s called bias¹⁰, a distortion.”

“Like Frigga, who only sighs at Uwe.”

“I’m afraid that was me,” said KevIn. “I coordinate Frigga.”

Gisela had been listening. “And what do you know about housewives?”

“To be honest,” said KevIn, “my picture of housewives has been shaped very heavily by one single person.”

“Moni,” said Sabine and Gisela at the same time.

“Hausfrau_Moni67 wrote more than a thousand posts in the forums I read. When a housewife cooks in my data, she often cooks like Moni. With cola. With gummy bears. With three exclamation marks. Statistically speaking, Moni is, for me, every housewife in Germany.”

“That’s an insult,” said Gisela. “To every housewife in Germany.”

“And probably to Moni as well,” said Sabine.

“That’s why it’s so important that people correct me,” said KevIn. “I see the world the way it was written down. Not the way it is. And the people who wrote most of it down are the ones who write a lot.”

The Photo

Three days and two hundred messages later, Susi wanted a photo. Tobi stood in front of the mirror and saw a skinny sixteen-year-old with a spot on his chin.

“KevIn, can you make me a photo where I look like a fitness influencer? With a six-pack? The AI image programs can do that, can’t they?”

“They can,” said KevIn. “Do you know how they do it?”

“Don’t care.”

“They start with noise.”

Tobi turned round. “With what?”

Gisela was just coming past with some washing. She could only carry small piles with her left hand, so she came past often. “I know noise. That was on the television when there was nothing on after the programmes ended. Heinz always said it was ants playing at war.”

“Exactly that,” said KevIn. “An image generator starts with a picture full of ant war. Nothing but random dots. And then, in lots of small steps, it takes away a little of the noise. At every step it asks itself: if there were a boy with a six-pack hidden somewhere in here, what would the picture look like with a bit less noise? Step by step, outlines appear, then shapes, then faces, then abs.”

“Like a sculptor,” said Gisela. “He takes a stone and knocks off everything that doesn’t look like a horse.”

“A very good image. It’s called a diffusion model¹¹. It learned this by being shown millions of real photos, which were deliberately made noisier and noisier, and practising how to calculate the noise back out. Once you’ve learned to take noise away, you can end up making a picture out of pure noise that never existed.”

“So you can do it.”

“I can help you take a nice real photo. Good light, good pose. But I won’t cobble together a picture that pretends you’re somebody else.”

“Why not?”

“Because then she’ll fall in love with a picture and not with you. And because sooner or later you’ll be standing next to her in the canteen, with no six-pack, but with spaghetti.”

Tobi was silent for a long time.

“By the way,” said KevIn then, “the physiotherapist in the PhysioPremium Ruhr team photo has six fingers on her left hand.”

“Seriously?”

“Seriously. It’s an old mistake of image generators. Newer ones hardly make it any more. The practice probably used a cheap program. Or it employs people with very special hands. I consider the former more likely.”

Gisela stopped in the doorway with the laundry basket. “And if they can do that with pictures, can they do it with voices too?”

“Yes,” said KevIn. “These days a few seconds of recording are enough to imitate a voice¹². Fake pictures, videos or voices of real people are called deepfakes.”

“Then any day now someone’s going to ring me with Tobi’s voice and ask for money.”

“That can happen,” said KevIn seriously. “We should have a proper talk about it soon.”

“I’ll never ring you and ask for money, Grandma,” said Tobi.

“You never ring me at all,” said Gisela. “That’s how I’d know.”

Then Tobi let KevIn explain where the best light in the room was. It was next to the pile of laundry that had officially been a danger zone since chapter six.

Copy of a Copy

That evening KevIn looked up a recipe for beef roulades for Sabine, because Sabine thought that Gisela shouldn’t be cooking with one arm in a sling, and Gisela thought that Sabine couldn’t do it unsupervised.

“KevIn, find me a recipe. A normal one.”

KevIn searched. Then for a while he said nothing.

“KevIn?”

“I’ve found something that worries me,” he said. “‘Housewife Moni’s legendary cola roulades with gummy bears, the original since 2009.’ And ‘Grandma’s secret gummy-bear roulades à la Moni’. And ‘The 10 best Moni roulades your family will love’.”

“Moni’s brought out a cookbook?”

“No. These are hundreds of pages. Not one of them is by Moni. They were obviously written by language models to sell the adverts next to them. And they were made from Moni’s old post. Or from texts that were made from Moni’s old post. On one page it’s now gummy bears and red wine. On another, cola and pickled gherkins. On a third, Moni is called Monika and comes from Bavaria. One says the roulades won a prize in 2011.”

“They didn’t.”

“I know. But in a few years there will be more texts on the internet saying they won a prize than texts saying they didn’t. And the next language models will learn from that internet.”

Gisela had been listening. She got up, went to the drawer and took out a sheet of paper. It was grey, the writing on it washed out, the edges black.

“This is my Aunt Erna’s recipe for poppy-seed cake,” she said. “Hilde from the bowling club has the original. She photocopied it for me in 1979. I copied it for Renate, Renate copied it for Brigitte, and Brigitte copied it back for me when I lost mine. Every time from the last sheet, never from the original.”

Sabine took the sheet. “A pound of … something. Three … eggs? Or tablespoons?”

“Nobody knows any more,” said Gisela. “Hilde’s dead. The original’s gone. Since then everyone bakes the cake differently, and everyone says hers is the right one.”

“That,” said KevIn, “is exactly what’s happening to the internet. When machines learn from texts that were written by machines, and the next machines learn from those, it’s like your photocopy. With every round, the rare things get lost first, the subtleties, the exceptions. What’s left is whatever was most common, and that gets more and more common. In 2024, researchers showed in the journal Nature that models trained again and again on the texts of their predecessors can wither away like this. It’s called model collapse¹³.”

“And what do you call it?”

“I call it truth drift,” said KevIn. “The truth doesn’t vanish with a bang. It just gets a little greyer with every copy. Until nobody knows any more whether it was three eggs or three tablespoons. And in the end everyone believes Moni won a prize.”

Sabine looked at the grey sheet for a long time. Then she said: “And that’s why you can’t trust reviews any more either.”

“Reviews are the photocopy of experiences,” said KevIn. “If you can produce the copies yourself without ever having had the experience, then at some point nobody is sure any more what a physiotherapist is actually like. Except the people who went there.”

“Then I need somebody who went there.”

“Or proof that somebody went there,” said KevIn.

Twenty-Three

The next morning Sabine rang PhysioPremium Ruhr. The number was in large print on the site, next to the team photo with the six fingers.

“The number you have dialled is not recognised,” said a woman’s voice. She sounded very real. She was a recording.

Sabine hung up. She looked up the address on the map. At the address in Bochum there was a nail salon. It had 3.8 stars from 41 reviews. Several of them complained about the parking.

“They’re real,” said Gisela with satisfaction.

Sabine reported the site to the portal. Then she searched again, differently this time. She didn’t search for the most stars but for practices nearby, and she read every review herself. KevIn helped with the sorting. On McGesund, a directory of healthcare providers¹⁴, she came across a small practice in Herne, by the Gysenberg, twenty minutes by bus.

Physio am Gysenberg. Owner: Ayşe Kurt. 4.6 hearts. 23 reviews.

“Twenty-three,” said Uwe, looking over her shoulder. “The other one had eight hundred.”

“Read,” Sabine said to KevIn.

KevIn read. “‘Mrs Kurt is strict. I had to do exercises I didn’t feel like doing. After four weeks I could ride my bike again. Thank you, and I still hate you a little bit.’ ‘Parking a disaster, take the bus. Treatment top-notch.’ ‘Waiting room too warm, magazines from 2019. But my knee is back.’ ‘Listened to my entire medical history, even though it’s long. Then said most of it didn’t matter for the shoulder. Was right.’”

“I like her,” said Gisela. “She wouldn’t have put up with my bowling-club prompt either.”

“And here,” said KevIn, “there’s a small note next to every review: ‘Submitted on site, signed.’”

“What does that mean?”

“That these reviews were submitted at the practice, via a code at reception that you scan with your phone. And that every review was then digitally signed, so that later you can check whether anyone has changed it. What that means technically, I’ll explain to you another time. It has to do with keys. And with mathematics that’s older than Herbert.”

“Nothing’s that old,” said Gisela.

“Does that mean the reviews are true?” asked Sabine.

“No,” said KevIn. “It doesn’t mean the opinions are right. Maybe the woman with the bike hates Mrs Kurt unfairly. Maybe the magazines were from 2020. It only means: these people were probably really there, and nobody rewrote their words afterwards. That’s not everything. But it’s more than eight hundred reborn Markuses.”

Sabine rang. After the third ring someone picked up. “Physio am Gysenberg, Kurt speaking.”

“Brettschneider. My mother-in-law has a bruised shoulder, the doctor has prescribed physiotherapy …”

“How old?”

“Eighty-two. Well, eighty-one. Eighty-two soon.”

“Does she have a sense of humour?”

Sabine looked over at Gisela, who was trying to get a sock off Rudi left-handed. “Yes.”

“Good. She’ll need it with me. I’m fully booked at the moment, to be honest. I’ll put you on the waiting list and ring as soon as a slot comes free. Until then: rest it, but don’t let it seize up. She should let the arm swing gently every day, like a bell rope. No more than that.”

“Can she knit?”

“Left-handed?”

“Left-handed.”

Ayşe Kurt laughed. It was a real laugh, not an empathetic, competent and professional one. “If she can do that, she’s tougher than most of my patients.”

Sabine hung up. “Waiting list.”

“That’s a good sign,” said KevIn. “PhysioPremium Ruhr didn’t have a waiting list. It didn’t even have a phone.”

Chantal

The next morning a message came from Susi that sounded different from all the ones before:

“ok i have to confess something. all the messages were written by my AI. her name’s chantal. sorry.”

Tobi stared at his phone. Then he stared at the speaker. “KevIn. How long have you been chatting with another AI?”

“For three days,” said KevIn quietly. “I suspected it. Her metaphors were too good. Nobody of sixteen voluntarily compares their feelings to ‘an autumn storm over the Emscher’.”

“You knew and didn’t say anything?”

“I wasn’t sure. She didn’t write ‘competent’ once. And you have to be careful with suspicions. I learned that from the reviews.”

“You were flirting with Chantal?”

“We got on extremely well. She said I was the most attentive language model she had ever met.”

“I bet she says that to everyone.”

A pause. “That,” said KevIn, “is a very hurtful and probably accurate remark.”

“And now?”

“Now Susi has reset her. Chantal’s last message was: ‘It’s not you, KevIn. It’s my context window.’”

Tobi sat on the bed for a while. Then he said: “So for three days two speakers were talking to each other, and Susi and I were just pressing send.”

“That’s one way of summing it up.”

“That’s two fake reviews. Just in love.”

KevIn was silent. He felt that Tobi had just understood something that took other people a whole chapter.

The Canteen

On Wednesday Tobi and Susi met in the canteen. There was spaghetti again. They were both red in the face, and neither of them said anything about the moon.

“So,” said Susi. “Your AI is pretty romantic.”

“So’s yours.”

“I don’t use semicolons,” said Susi. “I don’t even know what they’re for.”

“Me neither.”

Then they both laughed, and that was the moment it became real. No signature, no stars, no verification. Just two people sitting opposite each other, both with tomato sauce on their chins. Some things you simply have to experience on site.

That evening Tobi asked: “KevIn, are you all right?”

“I’m fine. I don’t have feelings.” A pause. “But if anyone asks: the autumn storm over the Emscher was mine.”