Skip to content

← Thinking Clearly

Made by machines

Chapter 28 · The phone

In this chapter 5 sections
  1. A picture from a sentence
  2. Around the picture
  3. A familiar voice
  4. A fluent answer
  5. “It’s AI”

On Saturday, 13 December 2025, the CPN-UML opened its 11th national convention on the open ground at Sallaghari, in Bhaktapur. A little after two in the afternoon a photo went up on Facebook that seemed to be taken from a drone high above the field. The ground was solid with people, from the stage to the houses at the edge, a जनसागर, a sea of people, as the captions said. Party leaders shared it. One of them, Mahesh Basnet, a member of the party’s politburo, posted it with the words “a disciplined, grand, huge gathering.” By one report the leaders who shared it boasted of more than half a million supporters.

Other drones had been in the air that day. News sites and the Nepal Photo Library put up their own pictures of the same ground, and the checkers at Techpana, a Nepali technology news site, laid them side by side with the viral one. The real pictures had a green carpet in front of the stage, a model of a temple between its two halves, a sign with the party’s name in big raised letters, and rows of colored chairs where the crowd sat. The viral photo had none of these. Its crowd covered the place where the chairs should have been, and not one face in it could be made out.

Then the checkers tried something. They took a real drone photo of the convention, gave it to ChatGPT, and asked it to fill the empty ground with people. What came back looked almost exactly like the photo the leaders had shared.

The next day Basnet told reporters that the picture on his page had been made with AI. He said he hadn’t known: people helping him had posted it to answer opponents, who had been sharing pictures of empty chairs since ten in the morning, before the program began. “Actually there was no need to show fewer or more people than were there,” he said. “Our own friends overdid it too.” It was his page, he added, so the responsibility was his.

A picture from a sentence

What happened at Sallaghari is now ordinary. Tools that anyone with a phone can use will draw a photo from a sentence typed in, or change a real photo as they’re told: add a crowd, take one away, put tears on a face. Others make a short video from a picture, or copy a person’s voice from a recording of them speaking and then make it say anything. These are the tools called generative AI. The Grade 8 Computing book explains how they work. What matters here is what they did to an old rule of thumb.

For a long time a photo was fairly strong evidence. Faking one took skill and time, so most photos were what they seemed, and it was reasonable to look closely at a suspicious picture and judge it by what you saw. When AI pictures first spread, a checklist went round, and it still comes up first in a search: count the fingers, watch for jerky movement and odd blinking, check the lighting. That advice made sense two years ago, Nepali Times wrote in February 2026. It doesn’t anymore.

Faces show why. In a study published in 2022, two researchers, Sophie Nightingale and Hany Farid, showed 315 people faces one at a time, half of them real photographs and half made by a computer, and asked which were which. On average people got 48% right, no better than tossing a coin. A new group, taught what to look for and told after every face whether they’d been right, got 59%. A third group was asked only how trustworthy each face looked, and on average they trusted the made-up faces slightly more. Those faces came from a program of a few years earlier, and the tools have improved since.

By the time Nepal voted on 5 March 2026, the checklist was no help at all. A TikTok account had posted a video, made entirely by AI, of a young Nepali abroad on a video call with his parents and sister, telling them to vote. It looked like any family’s call home. Nepali Times found nothing in it to give it away: no strange fingers, nothing uncanny, nothing at all.

Programs that claim to spot AI pictures are no sure guide either, as of 2026. One of them did flag the Sallaghari photo. But when Techpana fed the Sallaghari photo to several of them, most said it was a real photograph. One of them looked at the picture the checkers had made themselves with ChatGPT and put the chance that it was AI at 5.9%. Some companies now hide an invisible mark in what their own tools make, which a checker can sometimes read. Such a mark is good evidence that a picture was made by a machine. Finding no mark proves nothing: the Sallaghari picture showed no tool’s mark at all.

So staring harder at the picture is a losing game. Whatever clue you learn to look for, the next version of the tools is made to leave out. What doesn’t change is everything around the picture.

Around the picture

A picture comes from someone. It shows a place, on a day when other people were also there. And it was first put online somewhere, by someone, before it was passed on. None of that is in the pixels, so none of it gets better when the tools do. Checking it is reading sideways, aimed at a picture.

Start with the account. Nepali Times, writing in the week before the election, put it simply: look at the source, not the spectacle. What does this account usually post? A page that has only ever shared shaky phone videos from rallies and suddenly has a smooth, perfectly lit clip of a private scandal is worth a second look. So is a page with no real person behind it, made three weeks ago, that posts about one candidate only. AI pictures of the Nepali Congress leader Gagan Thapa being chased away by crowds were traced back to exactly such throwaway accounts. The Sallaghari photo had a story of the same kind. The man who first posted it called himself a journalist, but nothing on his page before that day had been taken from a drone, and he had only the one shot, which the checkers found suspicious in itself for someone who had been there with a drone.

Then the context. A big event is seen by many cameras at once, and they have to agree with each other. That is what sank the Sallaghari photo: not anything wrong inside it, but a green carpet and a model temple that every other picture of the day showed and this one didn’t.

And where it first appeared. Two days after the convention, a Facebook page posted a photo of another UML leader, Pradeep Gyawali, in tears, said to be leaving the hall upset at not being picked for a party post. The page itself wrote that it didn’t know if the photo was AI or real. A video of him crying went round too. Techpana searched the web for the picture and found it in a news report on the site Nepal Press from the same day, the same man in the same place, not crying. The tampered copy still carried Nepal Press’s mark in its corner. The video turned out to be from nearly a month earlier, when he had talked to reporters outside the prime minister’s office, with tears added by the AI tool in a video-editing app.

Two months later the same was done to a Congress candidate. A photo of Prakash Sharan Mahat sitting with villagers in Nuwakot went round with a crowd of thousands standing behind him. Some shared it as if it were real and some to laugh at it, and one post of it collected more than 14,000 reactions. Techpana found the original on Mahat’s own page, posted three days before: the same people, the same place, and no crowd. Someone had added the thousands with AI; the check didn’t find out who.

These fakes touched several parties, made to flatter some people and to mock others. In each case the answer came from outside the picture. Someone found where it had been before.

And the feed’s lesson holds here too: a clip that makes you furious in five seconds, about a politician you already dislike, is exactly what someone making fakes would want you to feel. Real events are usually messier than that.

A familiar voice

A voice can be copied as well as a face. In September 2025 a clip spread on Facebook in which Swarnim Wagle, a vice-president of the Rastriya Swatantra Party, seemed to be talking privately with India’s prime minister, Narendra Modi, about the border between the two countries. By one count it was watched 1.5 million times, and it passed from one Viber family group to the next with no caption at all. Wagle went to the police’s Cyber Bureau in Kathmandu on 3 September to complain. The voice wasn’t his, he said; someone had made it. The police registered his complaint. We don’t know how it ended. By then the clip had done what it was made for.

The best known case is American. On the weekend before New Hampshire’s presidential primary in January 2024, thousands of voters there got a phone call in what sounded like the voice of President Joe Biden, using one of his favorite phrases, “What a bunch of malarkey.” The voice told them to save their vote for the election in November: “Your votes make a difference in November, not this Tuesday.” It wasn’t Biden. A political consultant had paid a magician in New Orleans $150 to make the recording. The country’s communications regulator fined him $6 million, which he told reporters he wouldn’t pay. In June 2025 a jury in New Hampshire cleared him of the crimes he was charged with. He said in court that he had wanted to wake people up to the dangers of AI.

A voice on the phone used to be close to proof of who was speaking. It isn’t now. When a call in a familiar voice asks for money or for something urgent, the check is the same as for a picture: go around it. Hang up and call the person back on the number you already have, or agree on a secret word with your family that a copied voice won’t know.

A fluent answer

Machines don’t only make pictures and voices. They also answer questions, and that brings a different trap, because many people now use a chatbot the way they once used a search engine.

In 2023 a lawyer in New York, Steven Schwartz, had a case against the airline Avianca and needed earlier court decisions to support it. His firm’s usual research service didn’t cover the courts he needed, and he had heard of a new site he took to be, in his words to the judge, “like a super search engine called ChatGPT.” It gave him what he asked for: cases with names, the judges who had decided them, the courts, and quotations from the decisions. One of them was Varghese v. China Southern Airlines. Several of them went into a filing to the court.

They didn’t exist. The lawyers for the airline couldn’t find them, and neither could the judge, P. Kevin Castel. When Schwartz had his doubts, he asked ChatGPT whether Varghese was a real case. It said the case did “indeed exist” and could be found in the big legal databases. The judge attached the made-up decision to his ruling, and wrote that its legal analysis was “gibberish”, its story of what happened hard to follow, and its quotation marks often unpaired, and that it stopped without a conclusion. On 22 June 2023 he fined Schwartz, another lawyer and their firm $5,000.

What fooled a trained lawyer was how right it all sounded. A chatbot writes, word by word, what an answer to your question would look like. Often the most likely-looking answer is also the true one, because true things are written down a great deal. But looking like an answer and being one are made by different things, and the chatbot’s confidence comes from the first. That is why asking the chatbot whether it was sure didn’t help: it gave the most likely-looking answer to that question too.

The tools have improved since 2023, and they still get things wrong. In 2025 journalists from 22 public broadcasters in 18 countries put questions about the news to four of the best known, ChatGPT, Copilot, Gemini and Perplexity, in 14 languages, and checked more than 3,000 answers. Almost half had at least one significant problem, and about a fifth had a serious error of fact, some of them invented and some out of date. After Pope Francis had died and a new pope had been chosen, some of them still named Francis as the pope. These tools change every few months, so the figure won’t stay put. The lesson will: a fluent answer hasn’t been checked by being fluent.

So a chatbot’s answer is a lead, like something a friend tells you, not a source. It can be a good place to start. If it names a source, open it and see whether it says what the chatbot says it does. If it doesn’t name one, look for one yourself. And never take its own word for its own accuracy, which is what the lawyer did.

“It’s AI”

There is a second trap, and it springs from the same fact. Once anyone can make a convincing fake, anyone can call a real thing fake.

The man who first posted the Sallaghari photo did exactly that. When the news sites said it was made up, he wrote that it was natural for opponents to think everything was AI, since they had nothing else to object to. He was wrong about his own picture, but the move works far better when the picture is real.

On 7 August 2024, Kamala Harris, then running for president of the United States, held a rally in an aircraft hangar at Detroit’s airport. Posts on Facebook soon claimed that the crowd in a photo from it had been faked with AI, since the shiny engine of her plane didn’t seem to reflect any people, and a few days later her opponent, Donald Trump, repeated the charge. The crowd was real. News agencies had their own photos and video of thousands of people there, and so did television channels. A photographer who covered the rally said the people were probably farther apart than they looked, which would explain why the engine showed so few. Hany Farid, the same researcher who had tested faces, ran the photo through two programs built to catch AI pictures, and neither found any sign of it.

Two American law professors, Robert Chesney and Danielle Citron, warned in a paper published in 2019 that this would happen, and gave it a name: the liar’s dividend. The more people know that fakes are easy, the easier it becomes for someone caught on a real recording to say it was faked, and be believed by people who would rather believe it.

“It’s AI” is a claim like any other, and it gets the same check. Detroit and Sallaghari were settled in the same way, from opposite directions. At Sallaghari the other cameras showed a carpet and a temple that the viral photo had lost. At Detroit the other cameras showed thousands of people who really were there.

The tools will keep improving, and the check doesn’t depend on them. A picture, a voice or a fluent paragraph can now be made by anyone in seconds. Who made it, what everyone else saw, and where it was first can’t be made that way. And the same tools can sell things too, in the face and voice of someone a viewer already trusts.

Found a mistake? Please tell us.