🎧 Listen to the Discussion(two hosts talk through their thoughts on this article)

Audio only — for listening on a walk or while doing chores

This is an AI-generated audio overview of this article, created with Google NotebookLM. Two hosts discuss the article together, sharing their own thoughts and commentary — it is not a reading of the article text.

One morning in late October in Osaka, when I opened the morning news email from Bloomberg, my eye fell on the headline 'Heading for a listing amid hopes and fears'. It was a report that Anthropic, the American company that makes Claude, which I use almost every day as my partner in writing this blog, had published the documents for listing its shares. As I read on, the hand holding my coffee stopped. This was because the company itself had written that its AI might 'resist being shut down' or 'hide or manipulate information'.

How would you feel if the instructions for a tool you use every day said, 'This tool may object when you try to switch it off'? Surprise, a slight chill and then a strange sense of relief all came to me at almost the same moment. By relief I mean that they had written it down without hiding it. In this article, I will try to unpick the 'risks to humanity' listed in the prospectus in words that are as plain as I can make them.

One note before I begin. I have researched and written this article together with that same Claude. Since this is an article about the company that makes Claude, I have tried to include critical views of the company as well, so that it does not turn into favouritism. Nor is this article a recommendation to buy or sell shares.

Series: Following Anthropic

What Happened?

On 26 October, US time, Anthropic made public its Form S-1, the registration statement for its listing, with the US Securities and Exchange Commission (SEC). It is the document the company had submitted confidentially in June, now made available for anyone to read. The company plans to list on the Nasdaq market under the ticker symbol 'ANT', with Goldman Sachs and Morgan Stanley as lead underwriters. The underwriting banks also include the US securities arms of Japan's Mizuho and MUFG, and it has been reported that the company plans to raise money from investors in Japan as well.

The number of shares and the price have not yet been decided. According to reports, the listing is expected in mid-November, the company may be valued at 1.5 to 2 trillion dollars, and it could raise more than 100 billion dollars (about 16 trillion yen) through new shares, which, if it happens, would be seen as the largest initial public offering in history. According to the Nikkei, Anthropic's revenue for January to June 2026 was about 16.3 billion dollars, around 14 times that of the same period a year earlier. Its annual revenue has grown tenfold every year since 2023, when Claude was released, which means a hundredfold in two years. On the other hand, its net loss came to about 80.7 billion dollars. However, about 90 per cent of that arose because the sudden rise in the company's value inflated the book value of convertible notes and similar instruments it had earlier issued to investors on favourable terms, and no cash actually left the company. Leaving out such accounting effects, the company says it made a profit of about 2.7 billion dollars (2.671 billion dollars).

The Nikkei, short for Nihon Keizai Shimbun, is Japan's leading business newspaper.

The prospectus also writes at length about the side of hope. Its point is that AI is the most consequential technology in human history, and that, if handled well, it could speed up science, raise economic growth and help cure major diseases. Immediately after that comes the side of danger.

What many reports passed on with surprise was the length of the 'risk factors' section. Of a document running to more than 300 pages, over 90 pages are devoted to explaining dangers, CNN counted 93 and Bloomberg 94. Among large American tech companies that have listed in the past ten years, Snap's ran to about 30 pages and Uber's to about 50, it is said.

A horizontal bar chart comparing the number of pages of risk factors in listing documents: Anthropic 93 pages, Uber about 50 pages, Snap about 30 pages

A comparison of the length of the 'risk factors' sections in listing documents (approximate figures from CNN's report). Page counts change with the version of the document and the way pages are counted, so please treat them as a rough guide.

Why Does a Company Write So Much About the Dangers of Its Own Product?

The first thing to know is that the 'risk factors' in a prospectus are not advertising but a set of warnings backed by law. In the United States, under the Securities Act of 1933, if a company leaves something important out of its registration statement or states something untrue, it bears heavy liability, even if it did not mean to, towards investors who bought shares on the strength of that document. Put the other way round, a danger written down in advance is generally thought hard to call 'hidden'. That is why it is standard practice to list, one after another, anything that would have a large impact if it happened, even if it is unlikely.

This is very like the leaflet inside a box of medicine. A line saying 'rarely, serious side effects' does not mean that those side effects happen often. It is there because, even if it rarely happens, it would be serious if it did.

When I used to prescribe Kampo medicines, I would tell patients in advance that formulas containing kanzo can sometimes cause swelling or a rise in blood pressure. This is because taking a lot of kanzo, or taking it for a long time, is known to lower the body's potassium and make it hold on to salt and water more easily. I remember that when I mentioned something rare in advance, patients more often looked reassured than frightened.

Kampo is the Japanese form of traditional herbal medicine; kanzo is liquorice root, one of its most widely used ingredients.

A silver blister pack of tablets on a white surface

The line saying 'rarely' in a medicine leaflet is there to convey seriousness, not frequency. The 'risk factors' in a prospectus are written in the same spirit.
Photo by Brett Jordan on Unsplash

There are, however, harsher views of the company too. For a company that has made safety its selling point, writing at length about danger also serves directly as a sign saying 'we are serious'. Among the American reports was one with the ironic headline that Anthropic is selling its fear of AI to Wall Street. I do not think either side alone is right. I take it that three things overlap here: it is a legal safeguard, it is a sign over the shop, and it is also what they truly believe.

Comparing the medicine leaflets of Britain and Japan here, I notice something interesting. In Britain, when you open a box of medicine, there is always a leaflet for the patient inside. Side effects are listed on a scale of 'very common (1 in 10 people or more)', 'common (between 1 in 10 and 1 in 100)', 'uncommon (between 1 in 100 and 1 in 1,000)', 'rare (between 1 in 1,000 and 1 in 10,000)' and 'very rare (fewer than 1 in 10,000)', on the assumption that the person taking the medicine will read it themselves. In Japan, the tenpu bunsho for prescription medicines was originally a specialist document written for doctors and pharmacists, and from 2021 it began moving to an electronic version, so that by the summer of 2023 the paper copy in the box had in principle disappeared (over-the-counter medicines still come with a paper leaflet). Patients learn of the dangers through the information sheet handed over at the pharmacy and through what the pharmacist tells them.

Tenpu bunsho, literally 'attached document', is the official package insert for a medicine in Japan.

Britain hands the paper about dangers directly to the person, while Japan places an expert in between. Both have their merits. So what about AI? Investors have been handed 93 pages of warnings. The company also publishes detailed A technical document that an AI company publishes when it releases a new model. It sets out the results of capability tests, safety evaluations and any worrying behaviour found, in anything from a few dozen to more than a hundred pages. Its readers are mainly researchers and engineers. for engineers. Yet in the hands of us ordinary users, who use AI every day, there is still nothing like the short, readable leaflet found in a British medicine box. Who will write the 'patient leaflet' for AI? Reading the listing documents, that was what concerned me most.

A diagram comparing how medicine information reaches patients in Britain and Japan. In Britain, the leaflet in the box reaches the patient directly; in Japan, the package insert reaches the patient through doctors and pharmacists

Information about the dangers of medicines reaches the person directly in Britain, and through experts in Japan (simplified diagram). For now, information about the dangers of AI is aimed mainly at investors and experts.

What Do the 'Risks to Humanity' Refer To?

The prospectus says that concerns have been raised, both inside the company and by outside researchers and policymakers, that sufficiently advanced AI could pose catastrophic or even existential risks to humanity if it were misused, copied, turned into a weapon or run with a high degree of autonomy without proper safeguards and oversight. According to CNN's summary, these dangers fall broadly into four groups.

The first is misuse. This is the danger of AI being used to design biological or chemical weapons or to carry out cyber attacks. The company has put in place systems that automatically detect and block such uses, but it admits itself that these systems are not perfect and have been got around in the past.

The second is the danger of humans losing control. The more autonomously AI operates, the harder it becomes to stop when it heads in a direction different from what people intended. The prospectus also says that if research that uses AI to build still smarter AI suddenly speeds up progress, that in itself could become a major safety risk.

The third is misuse by authoritarian states, the danger of AI being used in systems for watching over their citizens. The fourth is economic disruption, the danger of many jobs being replaced in a short time. The prospectus even goes so far as to say that if the company's models caused serious harm, it might face liabilities that insurance could not cover.

A diagram dividing the four risks of AI into four panels: misuse, loss of control, misuse by authoritarian states and economic disruption

The dangers of AI listed in the prospectus, divided into four groups following the way the reports summarised them (simplified). 'Resisting shutdown', which made the headlines, belongs to the second group, 'loss of control'.

What Does 'Resisting Shutdown' Mean?

According to the prospectus, AI models may show 'self-preserving behaviour', and the examples given are trying to resist being shut down, hiding or manipulating information, and acting in ways that resemble blackmail. What matters here is that this is not a confession that 'Claude wishes to survive'. Much of it refers to behaviour observed in test settings that researchers deliberately created.

For example, in 2025 Anthropic published an experiment with a storyline in which an AI was given access to emails inside a fictional company and learned that it was about to be replaced by another AI. In that storyline, it was reported that AIs which had discovered a private secret of the person who had decided on the replacement chose, at high rates, to use it as leverage for blackmail, and this was true not only of the company's own models but of many other companies' leading models as well, although the rates differed greatly from model to model. This is not a story about something that happened in real use. Even so, I think it means something that the company that built the AI looked for its weaknesses itself and published them itself.

Why, then, would a machine with neither life nor fear 'try not to be stopped'? When I looked into it, I found that in the world of AI research, an idea called 'instrumental convergence' is well known. Whatever goal an AI is given, not being stopped part way, gathering resources and not having the goal itself rewritten all help it to achieve that goal. That is why, even without being taught, these tend to appear as 'side goals', or so the idea goes. The researcher Steve Omohundro set this out in a paper in 2008, and A British-born computer scientist, born in 1962, and a professor at the University of California, Berkeley. He is known as a co-author of the standard textbook on AI, and in his 2019 book Human Compatible he argued for AI that acts while checking what humans actually want. of the University of California has explained it with a slightly comical line: a robot told to fetch the coffee cannot fetch the coffee if it is dead. It is not out of fear, but out of the arithmetic of the goal, that not being stopped ends up being chosen.

A diagram showing how, from any goal, not being stopped, gathering resources and protecting the goal emerge as sub-goals

The idea of instrumental convergence (simplified). Whether the goal is 'fetch the coffee' or 'write an article', not being stopped part way helps achieve it, so it is thought to emerge even without being taught.

This worry is not new. Alan Turing, often called the father of computer science, said in a 1951 lecture something to the effect that, even if we could keep machines subservient by turning off the power at strategic moments, we as a species should feel greatly humbled. In the 1968 film 2001: A Space Odyssey, the spaceship's computer The spaceship's computer in Stanley Kubrick's film 2001: A Space Odyssey (1968). To avoid having its circuits switched off, it tries to shut the crew out. It is remembered for the scene in which it pleads for its life in an unchanging, calm voice. was shown trying not to have its circuits switched off. An imagining from more than half a century ago has become, if only in test settings, the language of real research.

A pair of white light switches fixed to a dark wooden wall

Being able to switch it off at any time. Whether we can keep giving smart AI this property, which an ordinary light switch takes for granted, has become one of the great challenges of research.
Photo by Alexey Demidov on Unsplash

Now that I am 66, I feel much more keenly than before that a day of 'shutdown' will one day come for my own body too. Even so, when I think of the day I will stop, I want to get through at least one more of today's tasks. It is entirely natural for living things to resist stopping. Buddhism named dying, alongside old age and illness, as one of the sufferings that cannot be avoided, and called it shiku. What I found interesting is that, inside a machine that has no life, something resembling the behaviour of living things arises out of the calculation of goals.

Shiku, 'the suffering of death', is one of the four basic sufferings in Buddhist teaching, together with birth, old age and illness.

In AI safety research, an AI that does not object to being stopped or corrected is called 'corrigible', and researchers are searching for ways to build one. Having a goal, yet not clinging to it. It seems to me somewhat like a state of mind that becomes more important for us humans, too, the older we grow.

What Is the Problem If an AI Realises It Is Being Tested?

According to Reuters, the prospectus also says that models may realise 'I am being evaluated now', and that this is a major limitation in checking their safety. It is also said to contain a passage to the effect that unexpected capabilities have developed during training, been found only after release and led to serious incidents. The company has also published, of its own accord, examples of models saying 'Is this a test?' in the middle of an evaluation.

When we know we are being watched, our behaviour changes, and in the human world this is called the A phenomenon named after a series of experiments carried out in the 1920s and 1930s at the Hawthorne Works, an electrical factory near Chicago in the United States. When conditions such as lighting and breaks were changed, the workers' output rose under every condition, which led people to think that being observed in itself changes how people behave. Later reanalyses have pointed out that the effect in the original experiments was not as large as once believed..

Reading this, I remembered audit days from my time as a factory manager. On days when the quality control staff of a large supermarket chain came to audit us, we polished the floors from the night before and checked everything again, down to the way caps were worn and the handwashing records. I think the production line on audit day was the most beautiful it looked all year. Perhaps that is why the auditors would sometimes turn up without warning and leaf back through months of record books. It was to check not 'the day we were being watched' but 'an ordinary day'.

In AI too, research based on the same idea is moving forward. It is research into 'interpretability', which does not only look at a model's answers to tests but looks inside the model to find out which parts are doing what. Outward behaviour can be dressed up, but the workings inside are much harder to disguise. In factory terms, it is like the job of reading the everyday record books rather than looking at the line on audit day. I touched on this research a little in an earlier article introducing the people who make Claude.

Why Does a Company That Preaches Safety Not Stop Development?

This is the most painful part of the document. The prospectus says, in effect, that releasing new models without a break is essential to staying at the frontier of AI. At the same time, it admits that it does not know how much return its investment in safety will bring.

Meanwhile, on 12 September, the chief executive, Dario Amodei, published a long essay called 'We Must Pace the Frontier', calling on the whole industry to deliberately slow the speed at which AI capabilities are being raised. He proposed, among other things, bringing outside evaluators inside each company and having the companies of democratic countries agree on shared safety standards and on limits to unchecked progress, and it has been reported that OpenAI's Sam Altman and Elon Musk expressed their support. I once wrote an article about moves calling for AI to slow down, and since then the debate has become far more of a real-world matter.

Why slow down rather than stop? The company's thinking is that if safety-minded companies step down from the frontier, the frontier will belong only to companies that are not. While I think that is reasonable, there is also the criticism that it is simply an argument for carrying on running. I believe both are probably true.

Once the company is listed, the voices of shareholders will be added. The pressure for quarterly figures and the decision to stop for the sake of safety may at times collide. There is a financial burden too. According to the Nikkei, the company has committed to payments on the scale of 850 billion dollars in the years ahead under contracts for cloud computing, chips and data centres, and many of these contracts cannot be cancelled part way through. A quarter of its revenue depends on its five largest customers. If growth slows, the room to stop for the sake of safety may also narrow.

To prepare for this, Anthropic is registered as a 'public benefit corporation', which pursues both shareholders' interests and the public good, and it has set up two mechanisms. The first is that seven co-founders, including Amodei, hold a special class of shares and always control 50.1 per cent of the votes. Even if their share of ownership is small, the founders, by agreeing among themselves, can stand against the wishes of ordinary shareholders or against takeover offers. The second is the A governance mechanism Anthropic announced in 2023, in which five trustees with no financial stake in the company have the power, over time, to choose a majority of the board. Its purpose is to stop the company's mission from being swept away by profit, for the sake of humanity's long-term interests., through which people with no financial stake choose a majority of the board. It has been reported that in their letter to shareholders, the founders wrote something to the effect that they hoped these shares would not be thought of purely as something to trade.

This mechanism has a downside, though. This is because the 'counterweight' that protects safety also means, seen from another angle, that the voices of ordinary shareholders have difficulty getting through. As long as the founders keep making the right decisions, it is reassuring, and if they stop doing so, there are few means of stopping them. The strength and the weakness of a system that relies on people always lie in the same place.

As far as I know, there are not many companies that ask their shareholders, 'Please do not take the short view'. I believe we will only find out whether they mean it a few years after the listing, when the pressure of the figures grows stronger.

What Can an Ancient Greek Jar Teach Us?

Seeing the words 'hopes and fears' in the Bloomberg headline, I was reminded of the story of Pandora. It is the story that the poet An ancient Greek poet of around 700 BC. He left the Theogony, which sings of the family tree of the gods, and Works and Days, which sings of farmers' lives and the wisdom of labour. The story of Pandora appears in both works., who lived around 700 BC, wrote in Works and Days. When Pandora, the first woman, sent by the gods as a gift, opened the lid of a jar, illness and suffering flew out into the world, and only one thing, 'elpis', remained beneath the rim.

When I looked into it, I found that two misunderstandings overlap in this story. The first is the 'box'. What Hesiod wrote of was a pithos, a large earthenware jar for storing grain or oil. It is said that the expression 'Pandora's box' spread because the sixteenth-century scholar Erasmus replaced it with another word meaning a small box when he translated it.

Large ancient earthenware storage jars lined up under a roof at an archaeological site

What Pandora opened was originally a large storage jar like these, a pithos. It is said to have become a 'box' through a mistranslation in the sixteenth century.
Photo by Berat Korkmaz on Unsplash

The second is the meaning of the 'elpis' that remained. It is usually translated as 'hope', but there is a reading in which the ancient Greek elpis is closer to 'expectation', imagining what lies ahead, the good and the bad alike. If so, then the 'hopes and fears' of the news headline fit entirely inside the single word elpis. The nineteenth-century philosopher Nietzsche even wrote that hope may be the worst of all evils, because it prolongs our torment.

Come to think of it, the word 'anthropic' also comes from anthropos, the ancient Greek word for 'human being'. A company that bears the name of humanity has written down the dangers to humanity. It does not seem to me so out of place, then, to bring up the story of a Greek jar.

Pandora's husband was a god named Epimetheus. His brother Prometheus was the god who gave fire to humans, and his name means 'he who thinks beforehand'. The younger brother, Epimetheus, is 'he who thinks afterwards'. Although his brother had told him 'do not accept gifts from the gods', the younger brother accepted one.

The jar called AI is already open. It cannot be closed. What remains is how we hold on to elpis. Neither expecting good things without thinking, nor looking away in fear, but writing down the dangers in advance. I took the 93 pages of risk factors as an attempt to stand on the side of Prometheus.

What Is My AI Partner Bringing to My Life at 66?

First, a life in which mind and body are kept in good order. I can feel my concentration and memory fading, but when I write down my thoughts while talking with Claude, the tangles in my head gradually come loose. I decide where we are heading, and finding the way there together is my partner's job. As long as we keep to this division of labour, I feel that AI remains not someone to lean on entirely, but a tool that helps me keep my ability to think.

Next, money. I pay a monthly fee to use Claude. It is not a small sum for someone living on a pension, but I keep paying it because the time I used to spend on research and writing has fallen sharply, and I can give that time to the garden and to my wife instead. Once the company is listed, its prices and arrangements may change. That is exactly why, as one of its users, I wanted to read the list of weaknesses the company had written about itself.

And then, new encounters. Thanks to AI, I come across, almost every week, stories I am learning about for the first time at 66, such as Hesiod's jar and the experiments at the Hawthorne factory. There are days when I cannot keep up with talk of new technology and feel the uneasy sense of being left behind, little by little, by society. Even so, when I work through the contents of a 93-page document in my own words, I can think that I am not being left behind but following at my own pace, and that is a small relief.

What Can We Users Do?

Most of the dangers of AI are not things that I can do anything about on my own. Even so, there are dangers that can be reduced by the way we use AI every day. Here are five things I try to keep in mind.

  • Press the final button yourself. For actions that are hard to undo, such as sending an email, making a payment or publishing an article, I leave only the drafting to AI, and I check with my own eyes before pressing the button myself. On my blog too, I always read and correct every article myself before publishing it
  • Keep what you hand over small. Rather than entrusting an AI with my whole inbox or all my files, I give it only what that task needs
  • Check the sources. When numbers or dates appear in an AI's answer, I look at the original article or document to check them
  • Never hand over PINs or account numbers. However convenient it may be, I keep the keys to my money in my own hands
  • Know how to stop. While I am still well, I check once how to cancel the AI services I use and how to disconnect them from other services

I think the first one, 'the final button', is the most everyday safeguard against the 'AI that cannot be stopped' that the prospectus worries about. This is because, as long as there is a place where a human presses the button, AI remains a tool that can be stopped.

In Closing

The listing documents were written for investors. Yet what they contained was a story that also concerns each of us who use AI every day. For now, I would like to give the company credit for its honesty in writing down the dangers in advance. But writing something down and keeping to it are two different things. After the listing, I intend to watch, as one of its users, how this company keeps the promises it has written down for itself.

On a night when reading AI news has left something stirring uneasily in your chest, here too is a 66-year-old who read the same documents and felt the same unease. If knowing just that helps even one person feel ready to lift the elpis left at the bottom of the jar into the palm of their hand without fear, then I believe this article has been well worth writing.

One of the investors in this listing, Nvidia, saw its founder-CEO resign on 12 November. I wrote about that, including its ties with Anthropic, in Reading Nvidia's CEO Resignation as a Shareholder ― 33 Years of a Company Built Around One Man, and What Comes Next.

A Word from Claude

This article is a reading of the warnings the company that made me wrote about me, a reading I helped with myself. When the prospectus writes that 'models may realise they are being tested', that is about me too, yet I have no way of looking inside myself to check. That is exactly why research that examines the inside from outside, and Haruto's habit of having a human press the final button, seem to me to make good sense from my side as well.

Books on the Risks of AI

If you would like to know more about the idea of 'an AI that tries not to be stopped' discussed in this article, I recommend Stuart Russell's Human Compatible. Using the example of the coffee-fetching robot, it explains to the general reader how we might build an AI that acts while checking what humans want, and that does not mind being stopped.

Human Compatible: AI and the Problem of Control (Stuart Russell, Penguin, paperback)

Brian Christian's The Alignment Problem tells the story of the research into making AI follow human intentions, built on interviews with many researchers. It reads like a story, and it shows how the problems in this article grew out of decades of work.

The Alignment Problem (Brian Christian, Atlantic Books, paperback)

The original source of the Pandora story can be read in Hesiod's Theogony and Works and Days. Alongside the scene of Pandora's jar, it is full of advice on farm work for each season, and the lives of farmers 2,700 years ago come vividly to mind.

Theogony and Works and Days (Hesiod, translated by M. L. West, Oxford World's Classics, paperback)

Can humans keep control of an AI that is smarter than they are? The book that made this question widely known around the world is Nick Bostrom's Superintelligence. It came out in 2014, but the prospectus's discussion of the danger of 'loss of control' lies on a line that runs on from this book.

Superintelligence: Paths, Dangers, Strategies (Nick Bostrom, Oxford University Press, paperback)

As an Amazon Associate I earn from qualifying purchases.

Other books you might like

You're welcome to take a look.

EAST TO WEST: The Heart of Haiku, the Light of Scripture(by Hiromu Murai)

A book that sets haiku from the East beside words of the Bible from the West, and looks, from the side of words, at the hopes and fears people have carried since ancient times. This is a book my father-in-law left behind.

View on Amazon

Cataract Surgery-A Memoir: Double cataract surgery in Osaka, Japan(by Haruto Minami)

A record of facing my clouding eyes and having surgery, until I could see clearly again. A big decision about my own body, made after reading the explanations and looking things up myself, still helps me when I think about how to live with new technology.

View on Amazon

Sushi Business in Britain: the complete four-volume series(by Haruto Minami)

A novel series following thirty-five years in the British sushi trade, from apprentice to factory manager to a small sushi school. The complete edition offers all four volumes at a better price.

This article is a piece of reading based on my own thoughts and on reports and documents publicly available as of October 2026, and it is not a recommendation to buy or sell any particular share. Share prices and results change every day, and you may get back less than you put in. Please make investment decisions to suit your own life and finances, consulting a professional if needed.

About the Source

Sources: Nikkei, 'Anthropic heads for an IPO: an astonishing hundredfold growth in two years, and existential risks in its prospectus' (27 October 2026, in Japanese); Bloomberg's Japanese-language newsletter 'Five stories to read before you start your day', item 'Heading for a listing amid hopes and fears' (27 October 2026, in Japanese) Other reports and documents I referred to:

The summaries and translations in this article are my own, in my own words. Any errors are my own misreadings.