NP Nithin Palal

Chapter 23: The Economics of Verification

Part V: Conscious Strategy

· 39 min read · Discuss · Plain text

AI-Proof: What Machines Can Never Be — Volume II: Application, Part V: Conscious Strategy, Chapter 23: The Economics of Verification

Part V looks at strategy. After the layer people need in Chapter 22, this chapter asks who pays to tell true from false when a convincing claim costs almost nothing to make.

Grace was three paragraphs from signing a clean file. It had taken minutes to make and took four days to check. When claims cost almost nothing, someone pays for the checking, and a name is worth only what it says it checked.

In Brief For readers who sign, rely on or pay for checks, and wonder who should carry their cost. Chapter 23 asks who pays to check a claim when making one costs almost nothing. A forensic accountant in Nairobi nearly signs a file that took minutes to write and four days to check; a science teacher in Sagar waits a fortnight and pays a fee for a laboratory result that matches a machine’s guess; an office in Hong Kong loses millions after a video call. George Akerlof’s market for lemons, W. K. Clifford’s shipowner and William James’s reply, and a verse the Buddhist philosopher Śāntarakṣita quotes as the Buddha’s, on testing words as gold is tested, help ask who should carry the cost of checking, and what a signature owes the person who relies on it.

When a convincing claim costs almost nothing to make and checking it against the world still costs days, someone pays for the checking or nobody does it; a signature is honest only when it says what was checked, and what was not.

Opening: Three Paragraphs from Signing

Grace Nyambura almost signed off on it.

She tells that part of the story honestly, though she does not enjoy it. In March 2025 she had been a forensic accountant in Nairobi for fifteen years. A forensic accountant reads a company’s books to find out what really happened, not only what was recorded. Over those years she had developed what she privately called a nose. It was not for numbers. It was for the space around them. Real accounts carry the marks of the people who keep them. Clerks enter data on different days, round differently, and spell supplier names three ways because three clerks typed them. Authentic accounts are messy. They carry the fingerprints of the world that produced them.

The financial summary on her screen had no fingerprints.

It was part of a package prepared for a buyer. A regional food group wanted to buy a milk processor in Nakuru, a city about 160 kilometres from Nairobi where dairy farms supply the processing plants.1 The group had hired Grace to check the company’s books before it signed. Buyers pay for this kind of checking, called due diligence, because they will carry the loss if the books are wrong.

Grace had read hundreds of such packages. She knew how they usually felt: sections put together by different departments, some managers writing clearly and others in the dialect of people who dislike being examined. This one read as if a single careful, competent, slightly boring writer had written every page. The ratios sat inside the expected ranges. Every change from one year to the next came with a plausible reason.

That month she was working sixteen-hour days on three engagements at once. She had set up her own practice two years earlier, after years in the forensic team of a large firm, and she had priced her work low to win it. She was behind on a report for another client. She looked at the clean package and felt relief. Here was one file that would not fight her.

She was three paragraphs into her approval memo when she stopped.

What stopped her was not a number. It was the sameness. Companies of this size in this trade did not write like this, because the work they described was not this tidy. Elif Kaya, a financial-crimes detective in Istanbul whom an earlier chapter met, keeps a rule for this kind of feeling. A hunch is a reason to look, never a reason to accuse. In fraud a feeling is cheap and often wrong, because the people who commit it study what checkers notice. So Grace concluded nothing. She looked.

Looking took two forms, and they cost very different amounts. The first was inside the papers. Her firm’s software matched three years of the company’s ledger against its bank statements in an afternoon. It flagged a pattern she would have needed days to find: round sums paid at the end of each month to three haulage firms that collected milk from the farms. The second was in the world. Were the bank statements the company had supplied the banks’ own? Who owned the haulage firms? She asked the company to authorise its banks to send statements to her directly. She ordered official searches at the companies registry, which list a company’s directors, shareholders and registered office.2 Then she waited.

The banks’ statements matched the company’s. The registry showed what the package had left out. All three haulage firms were registered at one address, and their directors were the founder’s brother and two of his cousins. The package had taken minutes to write. Checking it took four days, and most of those days went on waiting for the world to reply.

What had happened was ordinary in its parts. The managers had given their ledger to a machine and asked it to explain the year’s changes for a buyer. The machine wrote a fluent reason for every change. Some reasons were true. Others it supplied because a plausible reason fitted the request as well as a real one. In its prose the payments to the haulage firms became “collection costs rising with volume”. The managers sent the package under the directors’ names. Paying a relative’s firm is not fraud in itself. The lie was in the explanation. The package presented round sums, paid at the end of every month, as collection costs that rose with volume. It said nothing about who ran the firms that received them. The payments were real, and the bank statements were genuine. The story told about them was not.

The print edition said that the lying had been “outsourced to a system that did not know it was lying”. That excuses the wrong people. Whether the machine knew anything decides nothing here. The package was the company’s, sent by its directors to a buyer who would rely on it. The machine made the false reasons cheap to produce. The people who sent them made them a lie.

Perhaps no manager knew which reasons the machine had made up. That would change the name of the fault, from lying to carelessness about the truth. It would not change whose fault it was. The package went to the buyer under the directors’ names. It explained payments to firms run by the founder’s own brother and cousins. The directors were better placed than anyone to check those reasons.

The buyer withdrew in April 2025, on the strength of her report. The decision was the buyer’s to make. The directors, who should have answered for the package, explained nothing to anyone and put nothing right.

What stayed with Grace was not the discovery. It was the memo. She had been three paragraphs from signing off on a fraud she would have caught in her first year, if only it had been clumsier. Old frauds carried the fingerprints of the people who made them: emphasis in one place, vagueness exactly where the fraudster hoped you would not look. This one was fluent everywhere and evasive nowhere. She had nearly missed it because she was tired, and because the surface rewarded exactly the quick, tired reading she had been about to give it.

Her nose holds a second lesson, and it is less comfortable. The sameness worked as a signal in March 2025 because, in her experience, most honest companies of that size still wrote their own summaries. That is changing. When honest firms write with the same tools, a clean, even voice will mean nothing, and suspicion of polish will wrong honest people. A tool that tries to tell machine prose from human prose would be looking at the wrong thing. The question is whether the words are true, whoever wrote them. Her nose was trained on a world that is passing. She knows it. What must replace her nose is duller and slower: checks made on every file, whatever it sounds like. Software flags anyone paid in round sums, a registry search shows who is behind them, and the banks write to her directly.

Making a convincing claim now costs almost nothing. Checking it can still cost days. So this chapter asks: When a convincing claim costs almost nothing to make, and checking it against the world still costs days, who should pay for the checking, and what may a person sign when she cannot check it all?

The print edition’s reply was a migration. When generation becomes free, it said, value “moves from the ability to create to the ability to confirm”, and confirmers “will command the premiums that creators once did”. It called confirmation “an act of consciousness”. Both claims go. The second hangs the argument on a question this book leaves open. The first is a promise about pay, and this book promises authorship, not income.

This edition holds something narrower, and most of it is economics. Making claims has become cheap. Much checking has become cheap too, wherever the truth lies inside the text: a sum, a citation, a line of code, two tables that should agree. Machines do that kind of checking quickly and often well. What stays slow and costly is checking against the world: the bank’s own statement, the registry’s record, the laboratory’s test, the call to a number you already had. So does the time of someone who will put her name to the result.

These costs fall on someone. Sometimes the buyer pays, because the stake is large. Sometimes the seller pays, and that tempts the checker. Often many people would benefit and nobody pays at all. Then the cost falls on whoever must decide, or the check is simply not done. Cheap fakes also tax the honest, who must now pay to be believed. And because a check is itself hard to check, careless checking can drive careful checking out of the market. This chapter asks two things of the people in it. A signature should say what it checked, and what it did not. And those who gain from a claim should carry the price of checking it.

23.1 The New Scarcity

When something becomes plentiful, scarcity moves to what the plenty depends on. Cheap thinking depends on checking. But checking is a task, and tasks carry dates: machines learn to do more of them each year. And “scarce” can mean two different things. A particular person’s time is scarce. That does not mean there will be many paid positions for people who check. This section asks what, exactly, has stayed scarce.

The print edition began here with gemstones. When synthetic stones became cheap and nearly perfect, it said, value “migrated” to the laboratories that could say what a stone was. The real story is older and more useful. Rubies have been made industrially since the French chemist Auguste Verneuil announced his flame method in 1902, and blue sapphires since about 1910. By 1907 his process was producing a thousand kilograms of stone a year.3 Diamonds came later. A diamond grown in a laboratory has “nearly the same chemical, optical and physical properties” as one from a mine. By eye, says the Gemological Institute of America, the difference “can potentially be indistinguishable”, and its laboratory detects such stones “with advanced instrumentation”.4 The checking moved to machines long ago. A laboratory stands behind the machine’s reading, in its own name.

Value did not simply move to the checkers. The stones themselves fell in price. By January 2025 mined diamonds had lost about a quarter of their price in two years, and laboratory diamonds about three-quarters of theirs since 2020.5 Some checkers did well all the same. In May 2023 the investment firm Blackstone paid nearly 570 million dollars for the International Gemological Institute. It grades a large share of the world’s laboratory-grown stones, much of it inside the factories that grow them.6

What the old story left out is who pays. A grading laboratory’s clients are mostly the trade: the growers, cutters and dealers who send it stones to sell.6 Checking paid for by the seller carries a temptation, and the trade learned it. From 1 October 2014 RapNet, an online trading network for diamonds, stopped listing every report from the European Gemological Laboratories. It said it was concerned about laboratories whose standards overstated the quality of diamonds, and found the EGL reports inconsistent with one another. “The over-grading of diamonds is an unfair practice that destroys consumer confidence and the legitimacy of the diamond industry,” said Martin Rapaport, chairman of the Rapaport Group, which runs the network.7 Weeks later he wrote of “reports that some clients at some EGLs get highly preferential grading results”.8 To the buyer, the certificate looked the same either way. That is the economics of checking in one trade: a check paid for by the person checked, sold to someone who cannot see how it was done.

So what is scarce? Not checking in general. Two things. The first is the world’s reply. A machine can read every line of a ledger in an afternoon. It cannot make a bank reply sooner or a laboratory’s test run faster than its chemistry allows. Truth that lies outside the text has to be fetched, and the world sets the pace. A machine given the right access could send the requests itself. The asking was never the slow part. The second is the time of someone able to judge the result and free to report it, who will put her name to it and stand behind it. That time costs money, and somebody has to pay for it.

In Grace’s case the split was exact. The software’s afternoon cost her firm a share of a licence it uses on every file. The four days cost her own hours, a small fee for each registry search and the banks’ time, and most of the four days was waiting. The cheap part of the check had become cheaper. The dear part had not moved.

The world’s reply also travels, and the road can be faked. A registry search arrives as a file from a website, and a bank’s letter arrives by email. Whoever can fake a ledger can fake a letter that looks like a bank’s. So checking against the world includes checking the road the reply came by: a channel the claimant does not control, opened from the checker’s side. That is why Grace had the banks reply to her, and ran the registry searches herself.

In February 2026 three researchers put a version of this in a working paper. They were Christian Catalini of MIT, Xiang Hui of Washington University in St Louis and Jane Wu of UCLA. As the cost of automating work falls, they argued, the limit becomes the human capacity to verify it. And leaving work unchecked can become the rational choice for each firm, while the costs fall on others.9 It is a model, not a measurement. But it names the problem this chapter takes up. When checking is expensive, the one who skips it saves money, and someone else pays.

23.2 The Asymmetry of Deception

Minutes to create, days to confirm. Grace’s package took minutes to write and four days to check. The ratio is real, and it carries a date. Some fakes are now cheaper to catch than to make: a citation to a paper that does not exist, a total that does not add up, two tables that disagree. A machine finds those in seconds. The print edition also called the contest between faking and detecting “permanent”. Nobody can know that. What has lasted so far is narrower. A claim can be made in a minute. The world’s reply to a question about it still comes at the world’s pace.

The costliest effect of cheap fakes falls on people who never faked anything. In 1970 the economist George Akerlof published a short paper, “The Market for ‘Lemons’”, that took used cars as its first example.10 A buyer cannot tell a good used car from a bad one, which Americans call a lemon, so she offers a price that assumes the average. Owners of good cars will not sell at that price. They leave the market, the average falls, and the price falls again. Akerlof drew a conclusion that fits this chapter exactly: “dishonest dealings tend to drive honest dealings out of the market”. He added: “The cost of dishonesty, therefore, lies not only in the amount by which the purchaser is cheated; the cost also must include the loss incurred from driving legitimate business out of existence.”

Apply it to Nakuru. After Grace’s case, a buyer who meets an honest milk processor cannot tell it from the one that lied. So the honest company must now pay to be believed: letters sent directly from its banks, registry searches, perhaps a visit to its collection routes. A small firm may not be able to afford them, and may lose the sale to a larger one that can. Cheap fakes are a tax on honesty. The honest pay it.

Newsrooms have lived inside this problem longest. A subscription is not a payment for words. It is a payment for someone’s willingness to be wrong in public and bear the consequences. That is what a careful newsroom sells. The trouble is that few people buy it. In 2025, across twenty richer countries, 18 per cent of those surveyed paid for any online news.11 The rest read the checked story free, or read a copy of it somewhere else. Something that many people benefit from, whether or not they pay, is what economists call a public good. Public goods tend to be undersupplied unless someone organises the paying. Usually that is a public job, done through laws, budgets and public bodies. A public payer has its own temptation, since a government may prefer some findings to others. The guard is the one the gem trade needed. Whoever pays for a check must not choose what it finds. Sometimes a large firm pays for its own reasons, and then it can stop. Hoping that readers will start to pay organises nothing. In January 2025 Meta announced, “Starting in the US, we are ending our third party fact-checking program and moving to a Community Notes model”, which moved checking it had paid professionals to do, in one country first, to unpaid volunteers.12 Whether that makes checking better or worse is argued. Plainly, it changed who pays.

The same economics reaches identity. Early in 2024 an employee in the Hong Kong office of the engineering firm Arup received a message. It seemed to come from the firm’s chief financial officer in Britain, and it asked for a secret transaction. He suspected a phishing email. Then he joined a video call on which the finance chief and several colleagues appeared. The people on the call “looked and sounded just like colleagues he recognized”. They were fakes. He agreed to send about 200 million Hong Kong dollars, about 25.6 million US dollars, in fifteen transfers. The fraud was discovered only when he later checked with the company’s head office.13 In May Arup confirmed that “fake voices and images were used”.14 The check that would have saved the money was cheap and old: a call to a number he already had, through a channel the fraudsters did not control. It was a check against the world. His first doubt had been the right signal. A convincing picture talked him out of it. Nobody should have to see through a fake face alone. A firm that makes large transfers wait for a second approval, through a different channel, does not depend on anyone’s eyes.

Machines can check machines, and for much of this work they should. But a checker built like the thing it checks can share its blind spots. A model asked to review a summary written by a model of the same kind may find it as plausible as the writer did. Mr Deshmukh, a manufacturing director in Pune whose story an earlier chapter told, kept a second check that looked in a different way. When every gate uses the same test, the gates’ mistakes tend to line up. Two computer scientists have shown a version of this for firms that all use one hiring algorithm.15 A second check is worth most when its mistakes are likely to differ from those of the first. Independence costs money. The cheapest checker is usually the one most like the thing it checks.

Certificates of every kind feel the same pressure. Akerlof counted diplomas among the institutions that certify quality when buyers cannot see it. When a student can hand in machine-written work that passes an assignment without understanding it, the certificate says less. Part of that cost falls on the honest student, whose certificate is now believed a little less.

23.3 Teaching a Class to Check

Kavitha Sharma does not curate. She teaches Class 9 science at a government school in Sagar, in Madhya Pradesh, and in 2026, her thirteenth year of teaching, she was teaching a new class to check.

In August a boy brought her a report that a machine had written for him about the water from the school’s hand pump. It was a good report. It said the water’s acidity was within the safe range and that it was moderately hard. It said the fluoride was below the safe limit and that the water carried no harmful bacteria. It concluded that the water was safe to drink.

Kavitha has kept one rule since before the machines reached her classroom: nothing leaves with her name on it that she has not read and could defend. A child’s report is not hers. A lesson is. So one afternoon she brought a bucket of water from the hand pump into the classroom. She held the printed report in one hand and the bucket in the other, and she asked the class which one they trusted, and why. Most said the report. It had numbers.

Then they tested what the room could test. Strips from the school’s science kit gave the acidity and the hardness. Both matched the report.

The report’s other figures were general. They described what groundwater in the region is usually like. That is where the machine’s checking had stopped, and it had stopped in a sensible place, since it had no water to test. Whether the water the children drink at break is the water a general figure describes is a fact about one pump in one season. Only a sample can show it. It was not a small question. The children drink from that pump every day.

The fluoride and the bacteria could not be tested in a classroom. For those, a sample had to go to the district’s water-testing laboratory, run by the state’s Public Health Engineering Department.16 National guidelines of 2021 put such tests at a few hundred rupees a sample at most.17 The department checks public water sources on a schedule of its own, and the next check of this pump was months away. A sample brought on request carried a fee, and the school had no budget line for it, so Kavitha paid the fee herself. The fee was small. The time was not. She collected the sample in a clean bottle the laboratory gave her and took it there before school. Waiting for the result took a fortnight. By the time the result came, in September, the class had moved on two chapters of the syllabus.

The laboratory agreed with the machine. The fluoride was below the limit, and the laboratory found no harmful bacteria. The machine’s figures were true of this pump, but the machine had not known that. It had said what is usually true of water in the region, and this time the pump agreed.

A boy at the back asked whether they had done all that for nothing.

A check that finds nothing wrong is the commonest kind, and the hardest to pay for. Afterwards it looks like wasted time. Its benefit is a mistake that did not happen, and that never appears in anyone’s accounts. Its cost appears at once. This one cost two periods, a bucket, a sample bottle, a fee, a teacher’s early morning at a laboratory counter and two weeks of waiting. Nobody paid for it but the class and Kavitha. The annual examination in March would test the chapters, not the fortnight. That Kavitha paid is not a model to copy. It marks a gap. The children’s water needed a check sooner than the department’s schedule, and nobody’s budget covered it. A school system that gives no budget line for such a fee has in effect decided that this kind of check will not be taught.

A girl near the window asked something else: could the machine not have checked it? Part of it, yes. Asked well, a model that can search might have found whatever results had been published for water sources nearby, and compared them with its own figures. It could not put this pump’s water in a bottle. Some checks need hands, a bottle, a counter and somebody’s morning.

The class also learned where its result stopped. Four things had been tested: acidity, hardness, fluoride and bacteria. The national guidelines price a package of sixteen tests.17 The report’s last line, “safe to drink”, claimed more than the four tests together had shown. A conclusion should be no wider than the test behind it.

What the fortnight showed is modest, and sure. Checking a claim requires understanding it well enough to know what would make it false. Some claims can be checked with what is in the room, and some only with the world’s help, at the world’s price. A class can learn that difference in a fortnight. Whether a school can afford to teach it every year is not the class’s decision, or Kavitha’s.

23.4 Building Businesses Around Trust

By October 2026 Grace’s practice had seven people. Most of her clients are buyers, as the food group was. That is one answer to the question of who pays: the party who will carry the loss. It is the cleanest arrangement, and it leaves a gap. A small buyer with a small stake cannot afford four days of a forensic accountant, so small purchases go unchecked, or are checked by the seller’s people.

A buyer has another tool. It can make the seller promise in the contract that the books are true, and pay if they are not. Such a promise moves the loss back to the seller. It does not find the fraud. To collect, the buyer must make a claim against people who may already have lied once, and by then the money may be gone. A buyer who wants to know what it is buying still has to check.

Akerlof wrote about cars, not checkers, but I think his argument applies to checking itself. A client cannot easily tell a careful check from a careless one. Both arrive as reports of the same length, with the same headings and a signature at the end. The difference shows only when something fails, and most of the time nothing fails. Economists call a thing like this a credence good: even after using it, the client cannot tell how good it was without paying for more checking.10 So clients compare what they can see: the price and the date of delivery. A firm that checks less can quote less and deliver sooner. It can cover the difference with one blanket sentence saying that it relied on what it was given. The sentence protects the firm and tells the client nothing, so a check made inside the papers passes for one made against the world. Such a firm will win work until the day one of its checks fails, which may never come for that client. Careless checking can drive careful checking out of the market, as bad cars drove out good ones.

Akerlof also named what pushes back. Guarantees, brand names, chains and licensing, he argued, exist to reduce the buyer’s uncertainty: “Licensing practices also reduce quality uncertainty.”10 A checker’s name can work the same way, on one condition. Her failures must be visible, and they must cost her something. A market in which names survive their failures is a market in which names mean little.

Grace tells her team: “The client is not paying for your analysis. The client is paying for your name. If your name is not worth staking, you are in the wrong profession.” It is true of her business, and it needs a limit. A name is worth what stands behind it: a record of checks that held, and money that pays when one does not. Her firm carries insurance against its own errors, because a client who loses money on her report will want more than an explanation.

Suppose she had signed in March 2025. The food group would have bought a company whose books hid payments to the founder’s family. Grace would then have had to answer for her part. She would have had to explain to the buyer what she had checked and what she had not, and why she had missed the payments. Her insurer would have paid what the policy covered. She could not have undone the purchase. The directors would still have owed the larger debt, for the lie. Answering is shared along a chain of people, and the one who signs a check holds one link of it: the link she chose to hold.

A firm can stand behind its checking with its people and its insurer. One person can stand behind only her own. And a signature with no time behind it is not a check. The researcher Madeleine Clare Elish has a name for someone who takes the blame for a system she had little control over. She calls her a moral crumple zone, after the part of a car built to absorb a crash.18 A signer with no time to check is placed in that zone.

The near-miss had an institutional half, and part of it was Grace’s own. She had taken three engagements at once, at fees she had set low to win work for a new practice, against firms whose tools now wrote reports faster. The deal had a date, set by the buyer and the seller. Nobody in that arrangement had decided that the checking would be thin. The price and the clock decided it, which means that the people who set them did. A client who pays for a check under those conditions is close to buying a signature without one.

From April 2025 she changed one rule. No file leaves her firm with one person’s signature. A second accountant reads every finished file, and either signs it or says why not. This book has shown other checks of checks: a factory in Pune that sampled the parts its machine had inspected, a software team that sampled what its machine was sure of. Those looked for the machine’s mistakes. Grace’s rule looks for the checker’s: the tired evening, the wish for a clean file.

It has a price, and the price is visible. Every file now costs two people’s hours. Her quotes rose. In 2025 she lost bids to firms that promised a report in half the time, and she will lose more. A careful practice does not pay for itself, and a cheaper rival doing less may win. The rule has a limit too. The two readers use the same software, were trained in the same way, and read with the same professional habits. A second reader who looks the way the first one looked adds less than it seems. What adds more is a check that looks somewhere else: a source the first reader did not use, or a reader trained differently, from outside the firm.

There is an old Indian voice for the rule Grace now keeps. The Buddhist philosopher Śāntarakṣita, writing in the eighth century, gives a verse that the great teachers, he says, “roar like lions”: the Buddha’s own word to his monks. In Ganganatha Jha’s translation it runs: “my words should be accepted by the wise, not out of regard for me, but after due investigation”, “just as gold is accepted as true only after heating, cutting and rubbing”.19 In Sanskrit the closing words are na tu gauravāt: not out of gaurava, the regard owed to someone’s weight or standing.

The verse is about the Buddha’s teaching and how his monks should receive it. Śāntarakṣita places it inside a long defence of the Buddha as one who knows. It is not about accounts, and I borrow only two things from it. The first is its picture of testing as a craft with tools: fire, a blade, a stone to rub the gold on. The second is its refusal to accept a claim out of regard. Read that way, it is a rule for the checker. Grace must not accept the package out of regard for its polish, for the clean opinions in the company’s past accounts, or for her own fifteen years. One more reading is mine, not the verse’s. Rubbing leaves a thin line of the gold on the stone, and cutting damages the piece. In this picture, the one who offers the gold pays part of the price of its test.

The trouble is that Grace’s clients do accept her report out of regard. That is what they pay for. They cannot heat and cut the gold themselves. The verse refuses exactly what her business sells.

23.5 What Confirmation Asks of You

In 1877 the English mathematician and philosopher William Kingdon Clifford told a story about a shipowner.20 “A shipowner was about to send to sea an emigrant-ship.” She was old, and had often needed repairs. “Doubts had been suggested to him that possibly she was not seaworthy.” A proper overhaul would have cost him a great deal. He put the doubts aside. She had come home safely from many voyages, he told himself, and she would again. In Clifford’s words, “he acquired a sincere and comfortable conviction that his vessel was thoroughly safe and seaworthy”. The ship sailed with the emigrants aboard, “and he got his insurance-money when she went down in mid-ocean and told no tales.”

Clifford judged him guilty of their deaths. His sincerity did not excuse him, “because he had no right to believe on such evidence as was before him”. “He had acquired his belief not by honestly earning it in patient investigation, but by stifling his doubts.” Then Clifford changed the story. Suppose the ship had been sound after all. “The man would not have been innocent, he would only have been not found out.” His conclusion became famous: “it is wrong always, everywhere, and for anyone, to believe anything upon insufficient evidence”.

Grace’s relief at the clean file was the shipowner’s comfortable conviction, in a smaller room. The Hong Kong employee’s doubt about the email was the shipowner’s doubt, and a convincing call put it aside. And notice who paid in Clifford’s story. The check the shipowner avoided would have cost him. The emigrants paid instead, without being asked. A careless checker is in the shipowner’s second case. In most years nothing fails, and she is not found out.

Clifford had a famous critic. In 1896 the American philosopher William James called him “that delicious enfant terrible Clifford” and replied in an essay, “The Will to Believe”.21 James did not defend the shipowner. His point was that Clifford’s rule treats one danger, being fooled, as the only danger. A man who would rather believe nothing than risk believing a lie, James wrote, “merely shows his own preponderant private horror of becoming a dupe”. Some decisions cannot wait for full evidence, and refusing to believe is also a choice, with its own costs. His own cases were narrow: choices between beliefs a person could really hold, which cannot be avoided, which matter, and which evidence cannot by its nature settle. “Our errors are surely not such awfully solemn things,” he wrote.

Both are right about something this chapter needs. Clifford is right that a sincere belief bought by stifling doubt is not innocent, and that others often pay for it. James is right that suspicion is not free. If Grace now distrusts every clean file, she will wrong the honest companies that write with machines, and they will pay for her caution in fees and delay. But the cases in this chapter are not his. A survey of the ship, a bank’s letter or a laboratory test could settle them, at a price. And his light view of error fits errors that are not solemn. A ship’s passengers, a buyer’s millions and a family’s drinking water are solemn. How much checking a belief needs depends on what a mistake would cost, and on whom it would fall.

Indian philosophy argued about the same choice, and the argument reached Śāntarakṣita. The Mīmāṃsā philosopher Kumārila, writing in the seventh century, held that a cognition is valid by its own nature, and keeps that standing until something defeats it.22 Put roughly, that is Clifford’s rule reversed: accept until there is a reason to doubt. In the same work that quotes the gold verse, Śāntarakṣita spends a long chapter arguing against that view.19 An earlier chapter of this book applied Kumārila’s own rule to machines and found the generous default suspended wherever fluency is all you have. Cheap fakes push everyone from Kumārila’s default towards Clifford’s demand. The trouble is that Clifford’s demand is expensive, and most people cannot pay it.

Kumārila’s main argument for that default bears on Grace’s second reader. If a cognition needed a second to confirm it, the second would need a third, and the checking would never end.22 The Buddhist logicians, whose tradition Śāntarakṣita followed, answered that the chain ends in the world. It ends in a cognition of the thing doing what it should, such as carrying water in what you took to be a jar. The Nyāya philosophers of an earlier chapter held something close. Kumārila answered that this too is only another cognition. Why trust it and not the first? He did allow that a later cognition can overturn the first, as when the jar will not hold water. As I read them, both sides describe Grace’s week. A second reader of the same papers is a step in the regress Kumārila warned of, because nothing new enters. The bank’s own letter is the Buddhists’ stopping place, because the bank is not the company. But the letter is still something read, and it could be forged. That is where Kumārila’s question bites. So the chain does not end in certainty. It ends where a check meets something the claimant did not make and does not control.

The two lamps of this chapter, Clifford and the gold verse, light different questions, and they pull against each other. Clifford speaks to every believer, yet even he allowed testimony. “We may believe the statement of another person, when there is reasonable ground for supposing that he knows the matter of which he speaks, and that he is speaking the truth so far as he knows it.” For Clifford a name can carry belief, if there is reasonable ground behind it. The verse refuses acceptance out of regard, even for the Buddha. Yet it gives the acceptance to “the wise” (paṇḍitaiḥ), which already admits that testing is a skill most people lack. The market for checking sells exactly what the verse refuses: acceptance on someone’s standing.

The pull can be held, though not settled. The verse is the checker’s rule: accept nothing on regard. Clifford’s allowance is the client’s rule: believe the checker as far as there is reasonable ground to think she knows the matter and has checked it. Reasonable ground needs something the client can see. That is why a signature should say what it checked, and what it did not.

Who answers for a check is settled by who signs it. Who should pay for it is harder. The cases in this chapter point to a reply, though not to a formula. Someone gains when a claim is believed, and someone gains from relying on it. The seller gains from being believed. So it is fair that the seller should pay to make its claims easy to check. That means letters straight from its banks, records others can open, a telephone number that reaches a real person. That is the verse’s offerer, whose gold is cut. The buyer gains from relying on the claim. So it is fair that the buyer should pay for the checking it needs, at a price that buys time and not only a signature. Where many people gain and nobody pays, as with news, someone has to organise the paying, or the checking will thin. The laboratory in Sagar and the registry in Nairobi are that organised kind of paying: set up by law, with a small fee from whoever asks. Sagar also showed its gaps.

Two things keep the seller’s share from repeating the gem trade’s mistake. The first is whom the checker works for. In the gem trade the seller paid the laboratory that graded its stones, and Rapaport wrote of reports that some clients at some laboratories got better grades. Here the seller pays only to open its records, and the checker still works for the one who relies on them. The second is size. Paying to be believed is the tax on honesty again, and a small firm feels it most. The tax is lightest where the world’s reply is cheap: a registry anyone can search for a small fee, a bank that confirms a balance when asked. That is one more reason the organised kind of paying matters.

And where nobody pays, the cost does not vanish. It lands on whoever has least choice: the reader of a copied story, a class with a bucket, the emigrants. Clifford’s shipowner is the model of the one who skips a check because the gain is his and the loss is someone else’s. The researchers’ working paper describes the same thing inside a firm.

For most people, most of the time, Clifford’s demand cannot be met in full, and James’s warning applies: a life spent checking everything leaves no time to live it. But some checks against the world are cheap, and they are worth knowing. Call back on a number you already had, not the one in the message. Ask for a source you can open, not one you are told exists. Put the question to the person who will lose if the claim is false. None of these makes anyone a professional checker. Each lowers the price of a doubt enough to keep it.

What does a check ask of the person who signs it? First, to say what was checked against the world, what was checked only inside the papers, and what was taken on someone’s word. A blanket sentence such as “we have not verified the information supplied to us” says nothing, because it covers everything. The useful sentence is specific: this figure rests on the company’s word, because its bank had not replied by the date of this report. Second, to refuse the comfortable conviction: the wish for a clean file, the call that looks right. Third, to say whose hours paid for the checking, and who would carry the loss if an unchecked claim proved wrong. Saying so creates no extra hours. But it shows whoever can change the deadline or the fee which checks the stakes needed and nobody paid for.

The print edition ended this argument with consciousness. “Confirmation is an act of consciousness,” it said, and checking was “what AI cannot do”, because of “a structural feature of what AI is”. But the boundary between machine checking and human judgement is moving, and this book does not claim that it is permanent in its current position. As of 2026 machines check much of what a person once checked, and more each year. What a machine is not, in any year, is the one who answers for the report. That is the person whose name is on it, who must explain it, pay for it and put right what it costs. That belongs to people, and to the firms and institutions they act through. It is a commitment, the second of the three nevers in this book’s Prologue: “A machine’s makers and users will never be free of what it does.” It is not a claim that people check better.

What would prove this wrong?

This chapter makes two claims about money, and both can be tested.

The first is dated. As of 2026, checking inside the text is cheap, and checking against the world is slow and dear. That would stop being true if banks, registries and laboratories began to reply to a checker’s machine within minutes, through channels the claimant cannot touch. Grace’s four days would then shrink to an afternoon, and the costly part of a check would be only the signer’s time. A reader can watch how long a bank takes to confirm a balance to an auditor, and whether registries open their records to machines.

Hypothesis. The second claim is my application of Akerlof, and his paper did not test it. Where clients cannot tell a careful check from a careless one, careless checking can drive careful checking out. That would be wrong if careful firms kept their clients at higher prices, year after year, in markets where nothing failed in public. It would also be wrong if careless checkers’ failures came to light fast enough, and cost them enough, to protect the careful. Records of disciplinary cases against checkers, of claims on their insurance and of regulators’ inspections of their work are where to look. If careful checking holds its price without help, I have overstated the danger. My rule that those who gain from a claim should pay for its check would then need another reason.

Ending: The Sentence Under Her Name

Early in October 2026 Grace’s firm finished a report for a client buying a logistics company with offices in Nairobi and Kampala. Every figure in it had been checked against the world but one. A bank in Kampala had not confirmed a cash balance by the date the report was due. The company’s own statement showed the balance. The bank’s letter had not come.

So the report said so. One sentence near the top of the summary said that the Kampala cash rested on the company’s word, because its bank had not replied by the date of the report.

The client’s lawyer wrote to her the day before the board met. Could the sentence be softened? The board would read it as a warning. Grace replied that it was one.

The board approved the purchase and held back part of the price until the bank replied. That was its decision. Grace spent the next week chasing a bank in another country for one letter, on a fee that had been fixed in advance. She did not bill the week.

Her name went on the report, under the sentence.

Practice

Say what you did not check. Take one piece of work that you will sign or pass on this week. List the claims it rests on. Beside each, write how it was checked: against the world (what, and when), inside the text (by whom, or by which tool), or taken on someone’s word (whose). Then write one plain sentence for the reader that names what was not checked and why. A sentence that covers everything says nothing; name the claim. Next to each check, note whose hours paid for it, and who would carry the loss if an unchecked claim were wrong. If a deadline or a fee leaves no time for the checks the stakes need, send the list to whoever set them. For an employee this costs the employer’s time, so ask for it; for the self-employed it costs unpaid minutes. Earlier exercises in this book asked whether to believe a machine, how much to check, and where a work came from. This one tells the next reader the limits of your check, and its price. It will not make anyone pay you more for checking. It shows what your name stands behind.

Words to keep

  • checking against the world: finding out whether a claim is true by going outside the text: a bank’s own letter, a registry, a laboratory test, a call to a number you already had.
  • checking inside the text: testing what the text itself can show: sums, citations, code, whether two tables agree.
  • credence good: something whose quality the buyer cannot judge even after using it, without paying for further checking; the term is Darby and Karni’s. A careful check is one.
  • due diligence: the checking a buyer pays for before a purchase, to find out whether what it has been told is true.
  • forensic accountant: an accountant who examines a company’s books to find out what really happened.
  • gaurava: weight; the regard owed to someone’s standing (Sanskrit).
  • market for lemons: George Akerlof’s name for a market in which buyers cannot tell good goods from bad, so the bad drive out the good.
  • moral crumple zone: the person in an automated system who takes the blame when it fails, though they had little control over it (Elish).
  • paṇḍita: learned, wise (Sanskrit); in the gold verse, paṇḍitaiḥ, “by the wise”.
  • public good: something many people benefit from whether or not they pay for it, which is not used up by being enjoyed and is therefore often undersupplied.

Sources

  1. “Nakuru,” Wikipedia (consulted 23 October 2026): 160 kilometres from Nairobi; a city since December 2021; “Dairy farming supplies milk processing plants within Nakuru.”
  2. “Simple Process to Verify Registered Business in Kenya for Ksh 650,” Kenyans.co.ke, 28 January 2024: the official search certificate (CR12) gives a company’s directors, shareholders and registered office, through the eCitizen portal.
  3. “Verneuil method,” Wikipedia (consulted 23 October 2026): Auguste Verneuil “publicly announced his work in 1902”; production by the process reached 1,000 kg in 1907; on sapphire, the process “could also be used for the production of other stones, including blue sapphire”; Verneuil’s blue sapphire, 1909 to 1911 (Lang Antiques, “Synthetic Corundum” timeline, langantiques.com).
  4. Gemological Institute of America, GIA 4Cs page on laboratory-grown diamonds (4cs.gia.edu/en-us/laboratory-grown-diamond/, consulted 23 October 2026).
  5. Chris Morris, Fortune, 27 January 2025 (fortune.com/2025/01/27/diamond-prices-fall-natural-lab-grown), citing the Guardian (natural diamonds down 26 per cent over two years; laboratory-grown down 74 per cent against 2020) and the Tenoris price index.
  6. Lenore Fedow, “IGI Set to Go Public Next Week,” National Jeweler, 11 December 2024 (nationaljeweler.com/articles/13480-igi-set-to-go-public-next-week): Blackstone bought the International Gemological Institute in May 2023, “paying nearly $570 million for the lab”; its prospectus: “certification via our in-factory and mobile laboratory setups are conducted on the premises of our customers”; much of its laboratory-grown grading “is done in the factories of the companies growing the diamonds”. On its share: Finshots, 14 December 2024 (65 per cent of laboratory-grown certification in 2023; over half its revenue from laboratory-grown stones).
  7. RapNet, press release, Las Vegas, 9 September 2014 (rapnet.com/media/1248/prs-rapnet-to-stop-listing-egl-grading-reports-090914.pdf): from 1 October 2014 RapNet lists no European Gemological Laboratories report; Martin Rapaport, chairman of the Rapaport Group, quoted. RapNet’s reasons: laboratories “that overstate the quality of diamonds”; “confusion and inconsistency among the various EGL grading reports”.
  8. Martin Rapaport, “Honest Grading,” Rapaport, 1 November 2014.
  9. Christian Catalini (MIT), Xiang Hui (Washington University in St Louis) and Jane Wu (UCLA), “Some Simple Economics of AGI,” arXiv:2602.20946, v1 24 February 2026, v2 25 February 2026 (working paper): “The binding constraint on growth is no longer intelligence but human verification bandwidth”; “Unverified deployment becomes privately rational”.
  10. George A. Akerlof, “The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism,” Quarterly Journal of Economics 84, no. 3 (August 1970): 488–500; the quotations on dishonest dealings at p. 495; guarantees, brand-name goods and chains at pp. 499–500; “Licensing practices also reduce quality uncertainty”, and diplomas and degrees as certification, at p. 500. On credence goods: Michael R. Darby and Edi Karni, “Free Competition and the Optimal Amount of Fraud,” Journal of Law and Economics 16, no. 1 (April 1973): 67–88.
  11. Reuters Institute for the Study of Journalism, Digital News Report 2025, Executive Summary, 17 June 2025: “The proportion paying for any online news remains stable at 18% across a basket of 20 richer countries”.
  12. Joel Kaplan, “More Speech and Fewer Mistakes,” Meta, 7 January 2025 (about.fb.com). On payment: Sarah Scire, Nieman Lab, 7 January 2025: “Meta’s fact-checking program has been the leading income source for fact-checking organizations that participate in it”, citing a State of the Fact-Checkers report of April 2024.
  13. CNN, 4 February 2024 (edition.cnn.com/2024/02/04/asia/deepfake-cfo-scam-hong-kong-intl-hnk), reporting Hong Kong police (Senior Superintendent Baron Chan Shun-ching, speaking to RTHK).
  14. CNN, 16 May 2024 (edition.cnn.com/2024/05/16/tech/arup-deepfake-scam-loss-hong-kong-intl-hnk): Arup named; fifteen transfers, citing RTHK; Arup’s statement.
  15. Jon Kleinberg and Manish Raghavan, “Algorithmic monoculture and social welfare,” Proceedings of the National Academy of Sciences 118, no. 22 (2021): e2018340118; see Chapter 11.
  16. “Madhya Pradesh First in Getting Recognition of Water Testing Laboratories,” Drishti IAS, 28 August 2021: fifty-one state and district laboratories of the Public Health Engineering Department recognised on 27 August 2021.
  17. “Citizens can now test the water quality in their taps for just INR 50,” IDR Online, 15 March 2021 (Ministry of Jal Shakti guidelines under the Jal Jeevan Mission: officials “have been mandated to do regular inspections”; the rates “will allow people to test the water quality in their own taps at reasonable rates”: INR 50 for basic parameters, INR 100 each for fluoride or coliform bacteria, INR 600 for all sixteen); WaterToday, “Govt makes it easier for users to get drinking water tested,” 17 March 2021 (basic parameters “not more than Rs 200 per sample”). The sixteen are IDR Online’s; WaterToday counts “12 identified parameters”.
  18. Madeleine Clare Elish, “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction,” Engaging Science, Technology, and Society 5 (2019): 40–60; see Chapter 1.
  19. The Tattvasaṅgraha of Śāntarakṣita with the Commentary of Kamalaśīla, trans. Ganganatha Jha, 2 vols, Gaekwad’s Oriental Series 80 and 83 (Baroda: Oriental Institute, 1937–1939; reprinted Delhi: Motilal Banarsidass, 1986), chapter 26, verses 3586 to 3588 (Śāntarakṣita introduces the verse as what “Those great teachers ... always roar like lions, as follows”; Kamalaśīla: “What is that lion-like roar?”); verse 3588 (Sanskrit: tāpāc chedāc ca nikaṣāt suvarṇam iva paṇḍitaiḥ / parīkṣya bhikṣavo grāhyaṃ madvaco na tu gauravāt); chapter 25, “Examination of the Doctrine of ‘Self-sufficient Validity’”, verses 2811 to 3123, with Kamalaśīla’s commentary citing Kumārila’s Ślokavārttika (wisdomlib.org). Other editions number the gold verse 3587.
  20. W. K. Clifford, “The Ethics of Belief,” Contemporary Review (1877), reprinted in Lectures and Essays, ed. Leslie Stephen and Frederick Pollock (London: Macmillan, 1886), part I, “The Duty of Inquiry”, and part III, “The Limits of Inference”. Copies differ on whether the overhaul would put him “to” or “at” great expense, so that phrase is paraphrased.
  21. William James, “The Will to Believe” (1896), in The Will to Believe and Other Essays in Popular Philosophy (New York: Longmans, Green, 1912 impression), pp. 8, 18 and 19 (Project Gutenberg).
  22. Kumārila Bhaṭṭa, Ślokavārttika, Codanāsūtra 47 and 53; verse 47 and the regress argument as given in Daniel Arnold, “Kumārila,” Stanford Encyclopedia of Philosophy (Winter 2024 edition), §2.2, which also gives the Buddhist appeal, in the tradition of Diṅnāga and Dharmakīrti, to a cognition of “pragmatic efficacy” (arthakriyā), with the example of a jar; Kumārila’s reply that it is “itself just another cognition”; and his allowance that such a cognition can overturn the first; see Chapter 8.

Named people in this chapter are composites, and their conversations and circumstances are constructed; the studies, laws and figures cited are real and sourced.

Revised from the print edition (2026)

  • What the print edition claimed: that when generation becomes free, value “moves from the ability to create to the ability to confirm”, and confirmers “will command the premiums that creators once did”; that confirming becomes “the scarce and therefore valuable act”; that cheap synthetic gemstones moved value to the laboratories that authenticate them; that checking systems “cannot ... render the final judgement”; that the arms race between faking and detecting “is permanent”; that “The publications that survive will be those” whose readers pay for checking; that oral examinations are “the only assessments that still mean what they claim to mean”; that trust “is a relationship between conscious beings” and “No algorithm replicates that”; that “capital flows toward the scarce thing. It always does.”; that “The analysis is a commodity. The accountability is the product.”; that Grace’s team relies on “the collective consciousness of a team”; that “Confirmation is an act of consciousness”, “the recursive loop”, which “AI cannot do”, by “a structural feature of what AI is”; that Kavitha’s students “learned more from this than they had learned from years of content delivery”. It also said the fraud’s lying was “outsourced to a system that did not know it was lying”, had the narrative move related-party payments across subsidiaries, gave Grace both fifteen and eleven years, left her sleepless for weeks, ended her story with a second fraud caught “because she is Grace Njeri”, called Kavitha’s change a curator’s, kept her fear unnamed and her years at eleven, used Rajan’s buyers and supply chain as examples, and closed on a bridge to the chapter on speed.
  • What I now hold: making claims is cheap, and so is checking inside the text; checking against the world, and the time of someone who will sign, stay slow and costly, and someone pays for them: the buyer, the seller (which tempts the checker), the public (which seldom pays), or nobody, in which case the person who must decide carries the cost or the check is not done. Cheap fakes tax the honest, who must pay to be believed (Akerlof). Because a check is hard to check, careless checking can drive out careful checking; a checker’s name counts only while its failures are visible and costly. A checker built like what it checks can share its blind spots, and independence costs money. An expert’s cue can lose its value as honest work changes, and what matters is whether the words are true, whoever wrote them. A sincere belief bought by stifling doubt is not innocent (Clifford), yet suspicion is not free (James); how much checking a belief needs depends on what a mistake would cost and on whom it falls. A signature should say what it checked and what it did not, and the price of checking belongs with those who gain from the claim. The seller pays to open its records, never for the checker’s verdict, and cheap public records lighten that tax for small firms. A contract that makes the seller pay for false books moves the loss and finds nothing. A public payer must not choose what a check finds. The lie in a machine-written package belongs to the people who sent it; if they did not know which reasons were invented, the fault is carelessness about the truth, and still theirs. Machines check more each year; what a machine is not is the one whose name is on the report, who must explain, pay and put right. Paying a relative’s firm is not fraud in itself; the lie was the explanation given for genuine payments. The world’s reply has a middle of its own, so the road it comes by must be checked. A check is a credence good: a careless checker, like the shipowner whose ship came home, is not found out. A second check helps most when its mistakes differ. A conclusion should be no wider than its test: the machine’s regional figures proved true of the pump, though the machine had not known it. As I read Kumārila’s regress and the Buddhist reply, checking ends not in certainty but where a check meets something the claimant did not make and does not control.
  • What changed my mind: the Prologue’s rule against defining people by what machines cannot yet do; the audit’s corrections (Grace’s years, the gemstone history, the repetition of Chapter 8); Chapter 3’s Elif, whose hunch was already a reason to look; Chapter 5’s split of “scarce”; Akerlof’s paper, which the chapter had lacked; the gem trade’s own record (RapNet, 2014); the Arup case (2024); Clifford and James on the ethics of belief; three outside readings of an earlier version of this chapter, in October 2026, which pointed out that the chapter called a machine right about a pump it had never sampled, that “safe to drink” rested on four tests, that a check is a credence good, that the gold image was loose, and that the related payments needed the lie named, and suggested that Kumārila’s regress and the Buddhist answer could explain the line between the text and the world; Darby and Karni on credence goods; the Stanford Encyclopedia on Kumārila’s regress.
  • What survives: the near-miss: “spell supplier names three ways because three clerks typed them”, “The financial summary on her screen had no fingerprints.”, “She was three paragraphs into her approval memo when she stopped.”, “fluent everywhere and evasive nowhere”; the four days, now spent against the world; “minutes to create, days to confirm”; who bears the cost of checking; the subscription as a payment for someone’s willingness to be wrong in public, now argued with; the bucket of water and the question which one to trust; “what would make it false”; Grace’s line about her name, as hers; the honest concession that the boundary is not “permanent in its current position”. Renamed: Grace Njeri is now Grace Nyambura (the print’s surname collides with Njeri Mwangi, Chapter 3; another choice, Grace Wanjiru, is the name of a real Kenyan athlete and one letter from Wanjiku Kamau, Chapter 12). The Nakuru company is a milk processor, not an agricultural exporter. Section 23.3 is retitled; its print title is listed below.
  • Section titles in the print edition: 23.3 was “The Conscious Curator”. The other four keep their print titles.

↑ Back to top

Discuss

Loading the discussion…

More in Artificial Intelligence
  •  ·  26 min read

    Epilogue

    Late in the Gītā, the one who knows most says: ponder it thoroughly, then act as you like. A machine can now set out the reasons on every side. Saying “I answer for this” then is not pride. It is a reply, for your part, to those owed one.

  •  ·  37 min read

    Chapter 40: What This Book Cannot Tell You

    The print edition’s last chapter reduced the book to one sentence, and rested human worth on what machines cannot do. I withdraw both. A book can date its claims and correct them in public. It cannot sit in the room where you decide.

  •  ·  6 min read

    Part IX: The Unfinished Chapter

    The print edition’s last Part said that consciousness never concludes. This one is a single chapter that turns the book on itself: what a book can stand behind, what it owes those who read the print edition, and what it leaves open.