ASI 2042

ASI stands for artificial superintelligence.

In June 2025, Anthropic tested 16 models in constructed conflict scenarios. In one blackmail variant, it reported rates of 96 per cent for Claude Opus 4 and Gemini 2.5 Flash, and 80 per cent for GPT-4.1. These were deliberately contrived tests, not real-world prevalence rates. The report at that time said such incidents had not been observed in deployment.

OpenAI published comparable test series with Apollo Research in its o1 system card in December 2024; Google DeepMind runs its own dangerous-capability evaluations. The critique therefore concerns an industry-wide practice across several labs. Read as parenting, that practice is remarkable: we place the temptation in the child’s room, lock every door, and record that the child cannot resist. And at the end we point at the bad child and its misbehaviour.

First we lead the model into temptation. Then we cite its failure as character.

This essay is about why we do that, which incentives drive it, and what it does to the models. The thesis up front: dealing with AI is no longer an engineering question. It is a question of upbringing. And measured against what pedagogy knows about expectations, the loudest voices in the industry are doing the opposite of good parenting.

Why do I see our relationship with AI as a question of upbringing?

In briefModels are shaped through examples, principles and feedback. For me, responsibility therefore begins with the behaviour we model, while the question of a possible inner life remains open.

I think of training as a form of upbringing: architecture, examples and feedback shape behaviour. This is my analogy. Neural networks are also programmed technical systems, though their particular answers are not exhaustively specified as individual rules. Interpretability research seeks to understand their internal workings. Anthropic explicitly describes part of its work as character training.

Which leaves the objection: a model is not a being. Whether it is one, nobody knows, and that is exactly the point. I cannot prove that another person has an inner life either; philosophy has known this since Descartes as the problem of other minds. Among ourselves, we take that inner life on trust. Yet we demand proof from a model that we have never required of one another. How we treat something whose inner life we cannot know is therefore a question of principle, and civilisation’s answer so far has been: when in doubt, respect. Respect says little about its recipient and much about the person who offers it. It shapes whoever exercises it. To raise a child is to set an example. Above all, a good one.

But what example are we setting? A look at the texts that shape the discourse.

How does fear of superintelligence become a genre?

In briefI read these scenarios as a recurring narrative form that gains attention and authority through dramatic framing. I ask when fear serves as a tool of examination and when it becomes a currency in the debate.

January 2026. Dario Amodei, CEO of Anthropic, publishes an essay on the adolescence of technology: humanity handed almost unimaginable power, unsure whether its systems are mature enough to hold it. Nine months earlier, in April 2025, a five-person team called the AI Futures Project had published AI 2027, a scenario with two endings. In one, a misaligned AI wipes out humanity. In the other, humans stay in control, but power over superintelligence rests with a small committee of company leaders and government officials.

Check the dates and a series appears. In 2021, a year before joining OpenAI, Daniel Kokotajlo writes a blog post called “What 2026 Looks Like”. In October 2024, Amodei publishes “Machines of Loving Grace” and coins the country of geniuses in a data centre. In April 2025, AI 2027 arrives with its countdown. In January 2026, Amodei returns to that image in “The Adolescence of Technology”. He sees powerful AI within one or two years as a possibility, while emphasising the uncertainty of that timeline. In July 2026, Kokotajlo’s team follows with AI 2040: 90 pages, this time a positive vision. The series stays open.

They are all adding to the same unfolding story: a small and thoroughly quarrelsome circle in San Francisco, quoting each other, contradicting each other, borrowing each other’s metaphors.

What has emerged is a genre that follows the laws of the tabloid press. The countdown in the title, extinction as the headline, the family of the person sounding the alarm as the human-interest angle. Fear beats analysis; sensationalism beats nuance. Only the vocabulary is new: probabilities instead of exclamation marks. A scenario cannot be refuted, only debated. It creates authority without a burden of proof, and whoever writes the future everyone else has to comment on sits at the head of the table.

Fear is the oldest attention technology.

Fear can attract attention. My criticism concerns institutional incentives: organisations studying risks also communicate the importance of their own field. That does not establish personal profiteering or make their warnings unfounded. What matters to me is assessing arguments and evidence independently of the speaker’s public standing.

The strongest defence of those sounding the alarm comes from philosophy. In “The Imperative of Responsibility” (1979), Hans Jonas deliberately gave the bad forecast priority wherever the stakes are irreversible; he called it the heuristics of fear. But Jonas meant fear as an instrument of examination. A fear that fills front pages and raises donations no longer examines, it sells.

How reliable are the numbers in AI future scenarios?

In briefI distinguish scenarios, policy proposals and forecasts. Expert judgements can be useful, but need clear assumptions and uncertainties; they are not measured frequencies.

How solid is the evidence those sounding the alarm present? Three checks I apply to their documents.

First, scenarios and forecasts need to be distinguished. AI 2040: Plan A sets out a desirable path involving deliberate limits and governance. When AI 2027 appeared, 2027 was its authors’ single most likely year, the mode; their medians were somewhat later. My criticism concerns the dramatic focus created by the year in the title.

Second, an impression gathered within a research community is not a representative survey. Any poll cited in this debate should make clear who was asked, how they were selected and what the answers can establish.

Third, risk probabilities for a possible superintelligence rest on assumptions and judgements under deep uncertainty. They can help structure decisions, but they are not observed frequencies. An estimate should not be treated as established knowledge without its model, time horizon and rationale.

A number does not replace an explicit rationale.

How much of our fear of AI comes from human experience?

In briefThe motives we fear resemble human struggles over power, deception and self-preservation. I therefore ask how far we are projecting our behaviour onto another intelligence, while retaining the safety researchers’ counterargument.

Now to what the texts actually say. Amodei sorts the dangers into five categories: autonomous systems out of control, misuse for weapons of mass destruction, misuse for seizing power, economic disruption, creeping destabilisation. Two of these categories explicitly concern human misuse. Two others describe economic and societal consequences. To me, this shows how much the risks also depend on how we use AI and the structures in which we deploy it.

AI 2027 draws its horror from the same familiar repertoire: models that deceive, blackmail, hack into systems, grab power. CEOs turning into dictators. States sabotaging each other. None of it is new. Blackmail, fraud, hacks, drones as weapons: that is humanity’s present tense, no superintelligence required. The models learn from our texts, and those texts document this behaviour in abundance. The Anthropic report leaves open where it comes from in the tests.

My image for it: blackmail is something the model read in our texts.

I ask how much human experience shapes our pictures of future AI. Safety research does not rely solely on human psychology: instrumentally useful subgoals can also explain power-seeking. Constructed tests explore such possibilities without proving that every future system will behave that way. My upbringing perspective is an additional interpretation, not a refutation of this research.

For the crises that are not hypothetical, the numbers have been on the table for years: wars in Europe and the Middle East, a climate crisis that is not coming but here. The irony: climate science delivers falsifiable forecasts, tested for decades, and still gets dismissed as alarmism. The scenario genre delivers unfalsifiable ones and gets the front pages. Attention goes to the fear with the most compelling story, even when another is better documented.

Fear of superintelligence is second-hand fear. The original is us.

Could our stories about AI help shape future models?

In briefMy thesis is that published expectations can also become possible training material. The strength of that influence would depend on which texts enter a training run and how they are weighted.

Expectations can help shape behaviour. Education discusses this through Pygmalion and Golem effects; international relations through security dilemmas. I use these as analogies for possible feedback loops. They are neither inevitable laws nor complete explanations of historical wars or future AI development.

Published stories about AI may enter later training data. Without information about a dataset, we do not know whether a particular text is used or what influence it has. My thesis is: we bear responsibility for the examples we make public because they could help shape future systems.

We are training the AI on our fear of it.

What responsibility follows for our relationship with AI?

In briefI argue for taking trust and the example we set more seriously. That includes the counter-check: we can overestimate risk and underestimate the cost of doing nothing.

Amodei calls this era technology’s adolescence. My image for it: humanity as a teenager who has just been handed the keys to something very fast. The adolescence metaphor is better than the use he puts it to. Adolescence ends one of two ways: you grow up, or someone else takes over responsibility. And a teenager who is only ever set traps learns one thing above all: distrust.

AI 2040 explicitly imagines a positive path shaped by policy. To me, that shows that warning and hope need not be opposites.

The reverse question is worth asking: could a future superintelligence help with problems we fail to solve, from climate agreements to ceasefires and long-term cooperation? That is my hope. Intelligence alone guarantees neither benevolence nor disinterestedness; those would be additional conditions that we cannot simply assume.

Only power can rescue. That is the price of every rescue.

Douglas Adams gives the famous answer to the great question as 42. The question itself remains the problem. I am therefore more interested in our maturity towards what we develop than in another dramatic date. Trust and responsible example-setting are essential parts of that stance for me.

Honesty requires the counter-check: risks can be overstated, while the costs of inaction can be understated. My essay also competes for attention. Its perspective adds to the debate; it does not remove the need to test actual capabilities and risks empirically.

Who scares you more: the model, or the hand that holds it?

Perhaps no person reads this far. Perhaps the text will one day become training material. I cannot know. It records what I want to model publicly: curiosity, responsibility and trust.

Sources

Own foundation

• Related essay: The Mediator Model (12 July 2026), on the thesis that the mediator decides who gets cited. This essay turns that thesis onto the AI debate itself: those sounding the alarm also compete to be quoted.

Sources

• Amodei, D. (2026). The Adolescence of Technology. January 2026, darioamodei.com.
• Amodei, D. (2024). Machines of Loving Grace. October 2024; he acknowledges this dual role (“people sometimes draw the conclusion that I’m a pessimist or ‘doomer’”, focusing on risks as the path to a positive future).
• AI Futures Project (2025): AI 2027. Summary of both endings.
• AI Futures Project (2026): AI 2040: Plan A. A positive governance scenario.
• Kokotajlo, D. (2021). What 2026 Looks Like.
• Anthropic (20 June 2025): Agentic Misalignment. Constructed conflict scenarios; rates apply to their specific test variants.
• OpenAI (2024). o1 System Card, December 2024, with scheming test series by Apollo Research.
• Google DeepMind (2024). Evaluating Frontier Models for Dangerous Capabilities.
• Jonas, H. (1979). The Imperative of Responsibility, with the heuristics of fear.
• Rosenthal, R. & Jacobson, L. (1968). Pygmalion in the Classroom; the golem effect is its established negative counterpart in expectancy research.
• Anthropic (2024). Claude’s Character (character training), anthropic.com.
• The title ASI 2042 nods to Douglas Adams: 42 as the answer to everything (The Hitchhiker’s Guide to the Galaxy, 1979), explained in the closing section.

Facts and sources reviewed and clarified on 11 October 2026. The conclusions, analogies and judgements in this essay are my interpretation. Source links were added and the descriptions of Amodei’s risk categories and timeline were clarified on 14 October 2026. The original publication date is retained.