The Nashik district health office had a dengue season starting and a grant that lapsed at the end of the month. The communication officer, Sunita Rane, needed forty thousand posters in Marathi for panchayat noticeboards, primary health centres and school gates. The illustration budget was eleven thousand rupees, which in previous years had bought clip art.

This year an intern generated the artwork. A courtyard at mid-morning: an overturned bucket of standing water, a woman in a green sari tipping out a flowerpot, a child watching from a step. It was good. It was better than anything the office had printed in a decade, and everyone who saw the proof said so.

The press in Malegaon wanted the file by nine the next morning. At six in the evening a colleague looked at the proof for a long moment and said the woman looked familiar.

That is a thin thing to act on. Sunita acted on it anyway, because the check is free. She dropped the proof into three reverse image search services in turn. Two returned nothing useful. The third returned a stock photograph that had been on a library since 2015: the same courtyard geometry, the same pose, the same pot at the same angle, the same light falling in the same direction. It was not identical. The sari was a different colour, there was no child, and the wall behind was plain rather than tiled. But once she had the two side by side she could not unsee the relationship, and neither could anyone else in the room.

The intern's first response was arithmetic, and it is the argument everybody reaches for. The model file is two gigabytes. It was trained on billions of images. It cannot be storing them. Therefore it cannot have copied anything.

The arithmetic is right and the conclusion does not follow. Average compression across a whole training set says nothing about any particular item in it. An image that appeared once contributes a nudge that is averaged away. An image that appeared a thousand times, with much the same caption each time, becomes a reliable low-error target that the model can lower its training loss by learning specifically. This stock photograph had been syndicated to health departments, NGOs and municipal websites across two decades. Its caption was, in effect, the intern's prompt.

So the choice in front of Sunita at half past six had a cost on both sides, and neither cost was hypothetical.

Holding the print run meant losing the press slot. The next slot was the following week, the grant lapsed on the thirty-first, and unspent grant money does not roll forward. Dengue cases in the district had already started climbing. A poster campaign that arrives after the peak is a filing exercise.

Running it meant publishing, under a government masthead, at a scale of forty thousand, an image that a stock library's automated matching could plausibly find. The library's remedy would be ordinary and manageable. The reputational problem would not be, because the whole product of a public health poster is that people believe it. A district office caught using somebody else's photograph badly disguised would be a story that outlived the dengue season.

There was a third thing she had to establish before she could decide anything, and it took four minutes. She asked the intern for the generation record. He had exported the final artwork as a JPEG for the press, which destroys the text chunk that local tools write into a PNG. But he had kept the original PNGs in a folder on his own laptop, out of habit rather than policy, and the prompt, the seed, the sampler, the step count and the guidance scale were all sitting in the file.

Kavitha Anandh and her brother run a knitwear unit in Tiruppur with nine machines and one important customer: a German buyer who takes about seventy per cent of what they make. In March the buyer asked for lifestyle images for a lookbook. Not product shots on white — garments worn in a room, the same room every time, the same light, so that the set read as one place across a season.

Neither of them had a studio or a budget for one. Kavitha built the room on a hosted image service over four evenings and produced thirty images. The buyer was pleased enough to ask for thirty more in September, for the second half of the range.

In September she opened the text file where she had saved every prompt, pasted the first one in exactly as written, and got a different apartment. Warmer light. Different furniture. A window on the wrong wall. She tried the next one, and the next. Nothing matched.

Her brother's diagnosis was that the service had got worse, which is the diagnosis everybody reaches first. Kavitha's was that she did not actually know what the service had been doing in March, and she ran two tests to find out.

The first: she re-ran one March prompt six times without changing a character. She did not get six variations of one apartment. She got six different apartments, with different furniture and different architecture. Variation within a scene is what a seed change produces. Variation of the scene itself points at something upstream deciding what the scene is.

The second test was the one that settled it. She asked for an empty room, bare walls, a single grey box on the floor, nothing else. She got a styled living room with a plant, a rug and afternoon light through a window. Nothing she had typed asked for any of that. A language model was reading her short request, writing a longer and more detailed description, and handing that to the image model. Her words were an input to that expansion, not the instruction the image model received. And the expansion was itself a generation, which is why identical text produced different rooms.

That left three ways forward, with the buyer's deadline three weeks out.

The first was to rent a flat in Coimbatore, hire a photographer and models, and shoot the second thirty properly. Roughly a lakh and a half, three days, and it does not solve the problem — the March images would still not match the September ones, so the set would be split down the middle whichever way she went.

The second was to accept the mismatch and present it to the buyer as a new season look. Cheapest, fastest, and it asks a customer who is seventy per cent of the business to absorb an inconsistency they did not ask for.

The third was to rebuild locally. A second-hand machine with twelve gigabytes of video memory, a free node-based interface, an open checkpoint archived onto a disk she controlled, and then redo all sixty images so the whole set came from one place. That meant spending money on hardware, losing days of a three-week window to setting it up, and accepting that a mid-sized open model would not be quite as polished out of the box as the hosted service had been.

She chose the third, and her reasoning was not about this season. It was that the buyer would ask again in March, and in September after that, and a pipeline that cannot be re-run is not a pipeline.

Saraswati Prakashan prints workbooks and teaching aids for low-fee private schools across three districts. In June they took an order for a Class 2 numeracy set: forty flashcards covering the numbers one to ten, two cards for each number, each card showing that many of something familiar. Mangoes. Kites. Bangles. Lassi glasses.

An intern generated the cards over eleven days. He was careful. He regenerated each card until the count looked right, sometimes fifteen or twenty times, and he checked every one before handing them over.

The proofreader was Ramesh Jha, a retired primary teacher who had been reading proofs for the firm for nine years. He printed the set at full size, laid it out on the table, and counted.

Seven kites had eight kites. Nine bangles had eleven. And five mangoes had five mangoes and, behind a leaf at the left edge, most of a sixth.

The intern's defence was that he had checked them, and he had. He had checked them on a laptop screen at the size four cards fit side by side, which is the size at which a sixth mango behind a leaf is a leaf.

Anjali Prasad, who runs the firm, had to decide by that evening. The distributor's van to the districts left Monday morning. The next van was in three weeks, which was after the counting chapter would have been taught.

She had three options and none of them was free.

She could ship thirty-seven cards. But a counting set missing five, seven and nine is not a counting set, and the schools had ordered it because the counting chapter starts in week two.

She could have Ramesh count all forty at full size and have the intern regenerate the three that failed. This was the option everyone in the room preferred, and it was the one Anjali distrusted, for a reason worth stating precisely. The intern's method had produced three failures in forty. Applying the same method to the three failures produces failures at the same rate. It also assumes Ramesh caught everything, and two people in the office had already disagreed about the mango, which meant the error count was not three but somewhere between three and five.

Or she could stop prompting for counts altogether. Generate one good mango, cut it out, duplicate it five times, vary the rotation and the shadow, and place them by hand in a free editor. Forty cards at roughly twenty-five minutes each is about sixteen hours. Two people, two days. Monday becomes Wednesday, which means the van is missed and a private courier is paid.

There was a second thing on the table, and it had eaten two of the intern's eleven days. The lassi cards kept coming back as milkshakes in tall American glasses with straws. He had tried thirty variations of the wording. He had tried weighting the word. He had put milkshake in the negative prompt. Nothing moved it, and he had concluded that he was bad at prompting.

He was not. Those are two different failures with two different fixes, and the eleven days had gone on treating them as one.

Harbans Knitwear has been selling cardigans and pullovers to domestic wholesalers for thirty years. Gurpreet Sandhu, who runs it now, had nine days until a trade fair in Delhi and no catalogue.

What he had was the website his cousin built in 2011: sixty product photographs, six hundred pixels on the long edge, shot against a bedsheet. On a phone they look acceptable. At A4, which is about two and a half thousand pixels wide at print resolution, they are unusable.

Reshooting meant pulling sixty garments out of the warehouse, two days of a photographer at Ludhiana rates, and the packing staff working around it in their busiest month. It also meant discovering that eleven of the sixty styles had been discontinued and no longer existed in the building or anywhere else.

The design studio he hired proposed something cheaper. They ran the six-hundred-pixel files through a generative upscaler and sent back sixty images at print size, sharp, clean and convincing. The quote for the whole catalogue dropped by about two-thirds.

Meena Kaur, who has run production for eleven years, opened one at full size and said the stitch was wrong.

The original photograph showed a cable-knit cardigan at six hundred pixels, where the cable is a grey suggestion of texture rather than a countable pattern. The upscaled version had a clean, confident, entirely plausible cable — with a different repeat, a different rope width and a different crossing rhythm from the cable the factory actually knits. The button had become a different button. The cuff rib had become a deeper rib than the garment has.

Gurpreet's first instinct was that this did not matter much, and it is worth sitting with why, because it is the instinct most people have. The picture looked better. It looked more like a good cardigan than the original did. Nothing in it was ugly or obviously wrong.

Meena's objection was commercial rather than aesthetic. A buyer at a trade fair points at a photograph and orders four hundred pieces against a style number. If the photograph shows a cable the factory does not knit, the shipment that arrives is not the one that was ordered, and a buyer who has to explain that to their own customer does not come back.

So the decision had a real cost on both sides, nine days out.

Shipping the upscaled catalogue meant a book that looked better than anything Harbans had ever printed, delivered on time and within budget, carrying sixty photographs of garments that were slightly not the garments. Reshooting meant thirty-five thousand rupees, a day and a half the packing bay could not spare, eleven styles with no photograph at all, and a catalogue that would look plainer than the competition's on the next table.

Meena proposed a test before anyone decided. It took forty minutes.

She took one cardigan that still existed, photographed it properly on the shop floor, and downscaled her own photograph to six hundred pixels. Then she ran that through the same upscaler and laid the result beside the real photograph at full size.

Eight visible differences. Four of them were on things a buyer specifies in an order: the stitch repeat, the rib depth at the cuff, the button, and the woven label, which had come back with confident lettering that was not the company's name.

That settled it, and it settled it with evidence rather than with an argument about whether AI is trustworthy.

Brightbox is Nilesh Trivedi, an editor and a junior. They make corporate films for banks, hospitals and two engineering firms, mostly two to four minutes, mostly shot.

A district co-operative bank invited them to bid on a two-minute film explaining a new micro-loan product, to run on branch televisions and to be forwarded on WhatsApp. Budget two lakh forty thousand rupees. Done the usual way that is a two-day shoot, a small crew, two hired actors and a location fee, and about a lakh eighty of the budget is committed before anybody opens an edit timeline.

Riya, the junior, had watched a video model's launch reel the week before and proposed generating the whole film. Twenty thousand rupees of credits, no crew, no location, no actors. The margin was the entire argument and it was a good one.

Nilesh's objection was not that it would not work. It was that nobody in the room knew the hit rate, and a launch reel cannot contain that number. Every clip in a reel is real output. It is also the survivor of an unknown number of attempts, chosen by people whose job is to choose well, on prompts selected because they produced something.

It was Tuesday. The bid was due Friday.

Option one was to bid on the strength of the reel. Two lakh forty, a healthy margin, and find out during production whether the shots the bank actually needs are shots this model can make.

Option two was to spend Wednesday and about nine thousand rupees of credits measuring it, and bid on Thursday with a number. The cost of that is specific: two of the three days before the bid go into a test rather than into the treatment and the storyboard, and a competing production house on the second floor of the same building was bidding on the same job.

The shot list decided it, or rather the shot list made the question answerable. The bank's film needed a farmer signing a form, which is hands on paper. Two people talking across a counter. A branch board with the bank's name legible. A scooter arriving at a gate. And the same woman appearing in four separate shots across the film, because she is the borrower the story follows from application to first instalment.

Every one of those is a known weak spot, and Nilesh knew that in general. Hands at the point of contact with an object are where the model's learned appearance of a gesture stops matching what a gesture is. Legible text in video is roughly where image models were in 2022. A character who must be recognisably the same person in shot one and shot nine has no mechanism keeping her the same, because nothing in the pipeline stores an identity. He could have recited all of that on Tuesday afternoon.

What he could not recite was the rate, and the rate is the only thing a bid needs. Knowing that hands are hard tells you to expect trouble. Knowing that hands come back usable one time in twelve tells you whether to put them in the quote, and knowing that they come back usable zero times in twelve tells you the shot does not exist at any budget.

There was also a number underneath the whole argument that Riya had not included in her twenty thousand rupees. List prices are quoted per generated second, and what a production actually spends is the cost per second it can use. At a one-in-eight hit rate — which is ordinary once a shot has a specific requirement attached to it rather than being pretty — an eight-second clip listed at two hundred rupees has really cost sixteen hundred, plus the hour of somebody's attention spent watching seven failures. Teams that budget for the keeper run out of money somewhere around shot four of thirty.

They ran the test.

The Sahakari Patsanstha in Kolhapur has about nine thousand members. A large number of them are farmers and retired mill workers, and a meaningful number have no smartphone at all.

In 2023 the society added voice verification to its telephone line. A member calls, speaks a short passphrase, a speaker-verification system compares it against an enrolled sample, and a clerk can then read a balance, discuss a loan instalment or accept a change to a standing instruction. Complaints about the phone line fell by roughly two-thirds inside a year. The manager, Prakash Kulkarni, considered it the best thing the society had done for its older members in a decade.

In March a member in Ichalkaranji discovered that his balance had been read out to someone, his registered mobile number had been changed, and two transfers totalling one lakh forty thousand rupees had left the account over the following week.

The member had not called. His son had not called. What had happened was simpler than anyone expected. The member had spoken for about ninety seconds at a co-operative federation's annual meeting the previous year, and the video was on a Facebook page with four hundred followers.

A speaker embedding — a few hundred numbers summarising vocal identity — can be extracted from a few seconds of clean speech and used to condition a speech model directly, with no training run at all. Ninety seconds of a man speaking clearly into a microphone at a public meeting is not a marginal sample. It is a good one.

The society recovered about forty thousand rupees and wrote off the rest.

At the board meeting in April there were two proposals.

The first was to remove voice verification entirely and go back to one-time passwords. This is the safe-sounding answer and it has a real cost. The OTP flow is what produced the complaints in the first place. Around six hundred members have no smartphone, and for a farmer eleven kilometres out during the monsoon, being told to come to the branch to change a mobile number is not a minor inconvenience.

The second came from a younger director who had been approached by a vendor. A synthetic-speech detector, one lakh ten thousand rupees a year, quoted at ninety-six per cent accuracy. It preserved the service and it addressed the attack directly.

Prakash's nephew, who works in software in Pune, was at the meeting because he had driven his uncle there. He asked for the whiteboard and did an arithmetic problem that took four minutes.

Suppose one call in a thousand uses a cloned voice, which is generous. The society takes about ten thousand calls a month. That is ten fraudulent calls, of which a ninety-six per cent detector catches about ten. It also examines nine thousand nine hundred and ninety genuine calls and flags four per cent of them, which is about four hundred.

Four hundred members a month, most of them elderly, most of them people for whom the telephone line is the only route into their own money, treated as suspected impostors, so that ten fraudulent calls are caught.

The room went quiet, the director said the vendor had not mentioned that, and the nephew said the vendor would not have, because the ninety-six per cent is a true number and the four hundred is what the true number does at this base rate.

Ruchi Vibes has three hundred and forty thousand subscribers and a house style: a home kitchen, one cook, no studio lighting. Twelve episodes of the new season were shot, cut and graded, with a launch on Monday.

The title theme was eight seconds under the card, with a longer version used as a bed under the recipe segments. Jomon, a musician in Thrippunithura who had written the previous two seasons, quoted twenty-eight thousand rupees for a theme and three stings. The producer, Deepa Menon, said no, and the editor, Fahad, generated a theme in an afternoon from a text description.

It was good. It was catchy in the specific way a title theme has to be, which is that you can hum it after hearing it twice.

On Thursday, four days out, Deepa's mother heard it from the next room and said it was from a film. She could not name the film. Fahad said he had never heard it before in his life, which was true and not relevant. Two people in the office said it reminded them of something and two said it did not.

The decision had to be made that day, because Fahad was colour-grading for the rest of the week.

Launching meant going out with a hook that four people had split evenly on. If it turned out to be close to an existing melody, the consequence would not arrive as a lawyer's letter. It would arrive as an automated content match on the video platform, on every episode at once, with the revenue redirected to whoever registered the recording while the channel argued with a queue.

Changing it meant re-rendering twelve episode openers and thirty-six segment transitions, about a day and a half of the only editor's time in a week that had no day and a half in it, plus either paying a rush rate to Jomon with four days' notice or generating something else and having the identical conversation about that.

Fahad's position, argued with some heat, was that a model had made the melody, that he had never knowingly heard the film song anybody was vaguely gesturing at, and that a generated tune cannot be somebody else's tune because nobody wrote it.

Deepa did not know enough to answer that properly and she knew she did not. What she did know, from a year of running a channel on a platform, was that the question of who wrote it and the question of what would happen on Monday were not the same question and might not have the same answer.

There is a reason the risk concentrates on exactly this eight seconds. Popular film melodies appear in training data many times over — the original recording, covers, live versions, karaoke tracks, reaction videos, tutorials, and hundreds of hours of other people's content with the song playing underneath. That is the duplication pattern that produces something close to reproduction rather than something merely in the same idiom. A hook is also short, which is the length at which music infringement is actually litigated, and memorable, which is the property a title theme is chosen for. The channel had asked for the most dangerous possible thing and had been delighted when it arrived.

There was a third cost that did not appear on any spreadsheet. Jomon had written the last two seasons and had not been told the channel was generating this one. He would hear the theme on Monday like everybody else, in a city where the people who do this work all know each other.

Late July, heavy rain across the Konkan for a third day. Asmita Joshi runs the digital desk of a Marathi daily with a print edition in three districts and most of its readership on a phone.

At twenty to eleven a photograph arrived on the paper's tipline: a bus submerged to the windows, a man standing on the roof holding a child, a temple spire in the background. It was the strongest image anyone had of the night. Three competing outlets were running flood coverage with nothing like it.

The sender was a number with no name. The messaging app labelled the image as forwarded many times.

The desk publishes at eleven. On a night like that one, the traffic is the month's traffic, and the paper's own district is exactly where its readers expect it to be first.

Running the picture with a line saying the paper could not independently verify it is the ordinary compromise, and on that night it was not a small thing. People in low-lying parts of Chiplun were deciding whether to move to a relative's house. An image of a bus underwater on a named road is information they would act on. If it was fabricated or misplaced, the paper would be the outlet that laundered it into the record, and a correction at nine the next morning reaches a fraction of the audience the picture does.

Holding it meant a competitor running it within the hour, the paper looking slow on its own ground, and — if the picture was real — failing to show readers something true on the night it mattered.

A reporter, Nikhil, had put the file through a detector he found by searching. It reported eighty-seven per cent AI, and he wanted to lead with that.

Asmita asked him four questions. Eighty-seven per cent of what. Trained on which generators, and is this image from one of them. What happens to that number after a photograph has been through two rounds of messaging-app compression. And at a base rate where the great majority of pictures sent to a tipline on a flood night are real photographs taken by real people, what proportion of the images this tool flags are actually generated.

She was not being clever at Nikhil's expense. She had been accused twice of running a fake and once, more damagingly, of calling a real photograph fake.

Then she ran the routine, and it took twenty minutes.

Preserve first: save the file, screenshot the message and the number, note the time it arrived and the claim being made. Write the claim in one sentence — this shows a bus submerged tonight on the Chiplun road.

Reverse image search on two services. One returned nothing. The other returned a visually similar photograph from the 2021 Konkan floods, which was a different bus on a different road, so it ruled nothing in or out.

Source: a number with no name, forwarded many times, no original poster findable. A dead end, and dead ends are findings.

Place: the temple spire was distinctive. Satellite imagery showed a temple of that shape on a road near Chiplun. Street-level imagery from 2022 showed the building beside it as a single-storey shop. In the photograph it was two storeys with a blue plastic water tank on the roof.

Time: that was the one that ended it, and it had been in front of everybody for twenty minutes. The photograph is in daylight. It had been sent at twenty to eleven at night, as a picture of that night.

Studio Sundial is Ira Bhatt and three others, working out of two rooms in Lado Sarai. The largest job they had been offered came from a two-year-old Gurugram company making a water-quality sensor for housing societies and small water plants. The company was closing a funding round and preparing to sell in the United Kingdom and the United States. They wanted a full identity: a mark, packaging, and a campaign of about forty images. The fee was six lakh forty thousand rupees, which was more than the studio had billed in the previous five months together.

The designer on the job had built the mark by generating around three hundred candidates on a hosted service, choosing one, and tidying it. The campaign images had been generated on the same service, on the studio's consumer subscription, which was what the studio had because it was what the studio could afford.

In the second meeting the founder asked the question every client asks, in the form every client asks it. We will own all of this, obviously, exclusively, everywhere.

The accurate answer is four paragraphs long and contains the words unsettled and depends on the country.

In the United States, the Copyright Office has been consistent across a series of decisions and a formal report: copyright requires a human author, prompts alone do not supply authorship however detailed or however many iterations were run, and material generated by a machine in response to a prompt is not protectable — though a human's own expressive contributions, including selection, arrangement and subsequent editing, can be. A registration for a graphic novel illustrated with generated images ended precisely there: the text and the page arrangement protected, the individual images not.

In the United Kingdom and in India there is an older provision deeming an author for a computer-generated work with no human author, being the person who made the arrangements necessary for its creation. It has barely been tested in either country, the UK has consulted on removing it, and the Indian position was left ambiguous by a registration naming an AI as co-author which the office then sought to withdraw.

In China, a court has found copyright in a generated image on the basis that the user's prompts, parameters and repeated refinement showed sufficient intellectual investment. Similar facts, opposite conclusion.

So the same mark might be unprotectable in one of the client's three markets, protected on different reasoning in another, and owned by legal fiction in the third.

Ira had to decide whether to say that, in the meeting where the fee was being agreed.

Saying yes was what every other studio the founder had spoken to had said. It is what a client expects to hear, and a long careful answer in a pitch sounds like a studio that is not sure of itself.

The cost of saying yes was not immediate and it was large. If the client registered the mark, built a brand on it and a competitor in the United States reproduced the artwork, there would be no copyright to stop them. And the founder's lawyers would read Studio Sundial's contract during funding diligence, where a warranty that the work was free of any third-party right and exclusively owned everywhere is a warranty the studio could not support.

The cost of saying the accurate thing was six lakh forty thousand rupees going to the studio that said yes.