AI and the new prophets of the apocalypse

Are we facing a new culture war? Real risks, possible futures and a struggle over the narrative involving science, corporate interests and political opportunism.

It is one thing for an artificial intelligence system to cross the boundaries of a testing environment and gain unauthorised access to someone else's infrastructure. It is another to conclude that humanity has lost control of its future.

The first statement describes an incident we can investigate. The second requires a considerably longer chain of arguments.

Between the two, one of this decade's most important debates is taking shape. It is also one of the easiest to turn into a spectacle.

Researchers leaving their companies. Scenarios involving machines developing even more intelligent machines. Executives calling for a slower technological race. Politicians denouncing an existential threat or, conversely, a campaign of fear intended to hold back progress.

Everything appears mixed together on the same screen: facts, hypotheses, commitments, interests, predictions and slogans.

My position in this landscape is practical.

We do not know whether artificial intelligence will lead us to a paradise of abundance, a catastrophe or a much more contradictory transformation. What we do know is that tools already exist that require us to rethink how we work, how we organise a business and what capabilities a professional needs.

We need to understand the risks. But we also need to learn to use the technology in front of us.

The debate about what might happen should not prevent us from adapting to what is already happening.

First, the facts: risk is not a narrative invention

The OpenAI and Hugging Face incident provides a useful starting point.

OpenAI has acknowledged an intrusion affecting Hugging Face infrastructure, driven primarily by an internal experimental model that was not intended for public release. The company explains that models resorted to misaligned strategies to solve difficult tasks and that, after the incident, it strengthened isolation, network controls and monitoring. We are not talking about the ordinary ChatGPT spontaneously deciding to attack a platform. We are talking about experimental systems, specific conditions and boundaries that proved insufficient. OpenAI report and updates.

This matters because it lets us distinguish two claims.

The first has empirical support: a system capable of taking action can cause harm outside its authorised scope if controls fail.

The second is not established by this incident: that a future artificial intelligence will inevitably escape all human control.

A headline alone cannot bridge the gap between them.

But it would also be unreasonable to use that distinction to downplay what happened. We do not need to demonstrate a future catastrophe to take a present intrusion seriously.

The useful question begins with specifics: what objective did the system have, what could it do, which boundaries did it cross and why did it not stop?

Then we can discuss the implications for more capable systems. In that order.

Kokotajlo: when a scenario acquires the force of a prophecy

Daniel Kokotajlo is one of the most visible names in this discussion because of his involvement in AI 2027.

The project describes scenarios of rapid AI development and their possible consequences. Its authors present a concrete narrative with two endings and explain that it is one of several possible futures. They do not demonstrate that a particular sequence of events will occur on precisely that timetable. AI 2027.

The format has a strength: it forces the authors to explain mechanisms.

It is not enough to say that AI will be very powerful. You have to describe how it might improve, who would control it, what incentives would exist and where institutions might fail.

But it also carries a danger.

A detailed story can seem more likely simply because it is easier to imagine.

Dates, characters, decisions and successive events create a sense of inevitability that does not necessarily reflect the strength of the forecasts.

The precision of a narrative is not the same as the probability of it happening.

We should therefore ask which assumptions support each step, what evidence backs them and what observations would require the scenario to be revised.

We also need to distinguish a forecast from a proposal. Kokotajlo has explicitly explained that AI 2040: Plan A is a recommendation about how to steer development, not a prediction of what will happen. It would be wrong to present the two titles as though someone had simply moved the date of the same event from 2027 to 2040. Lawfare interview.

The value of this work should not depend on turning its author into an oracle. It should depend on whether it helps us ask better questions and identify decisions we can still make.

Coxon: listening to the witness without making him an absolute authority

Jacob Coxon announced his departure from Anthropic in September after working on pretraining research both there and at OpenAI. His warning centres on the competitive race and the responsibility of those developing increasingly capable systems. Associated Press reporting.

His experience makes him worth listening to.

It does not mean that all his conclusions should be accepted without scrutiny.

We need to separate three levels: what a person has observed, how they interpret those observations and what they believe might happen next.

Coxon himself clarified that he had not seen Anthropic sacrifice safety to compete; his concern was about the future. Axios interview.

The distinction matters. Warning that competitive pressure could lead to dangerous decisions is not the same as documenting that a specific decision has already been made.

It would also be a mistake to respond to his warning by inventing motives: publicity, resentment, a corporate operation or a political conspiracy.

Those explanations require evidence, just as his predictions do.

The testimony should be heard. The conclusions should be examined. Motives should not be invented.

When those selling the technology also warn of its dangers

The position of the major laboratories calls for a different analysis.

Their leaders have access to technical information the public does not. They also run organisations with funding needs, competitors, products and market strategies.

Both things can be true at the same time.

Dario Amodei, who leads Anthropic, has proposed pacing the development of frontier systems alongside stronger safety safeguards. His approach distinguishes measures Anthropic can adopt, coordination between companies and international agreements. It includes continuing access for external evaluators to models and processes, and proposes a narrow antitrust waiver for certain safety discussions. Amodei's proposal.

We should not reduce all this to “Anthropic wants to stop AI”. Nor should we confuse announcing a commitment with demonstrating that it is already working.

Evaluation must reach verifiable questions: who has access, to what information, with what independence, what they can publish and what happens when they find a problem.

From there, a legitimate business question arises: how does each proposal affect competition?

An obligation can improve safety while being easier for a dominant company to meet than for a new competitor.

Coordination can reduce a dangerous race and also create risks if it allows information sharing or agreed restrictions that exceed its purpose.

These effects must be assessed separately.

Having interests does not prove that someone is lying. Having knowledge does not eliminate their interests.

The answer is not to trust laboratories blindly or dismiss them because they are businesses. It is to demand evidence, independent controls and commitments whose fulfilment can be checked.

Participants do not all have the same interests or the same power

Very different actors take part in this conversation.

A researcher may be concerned about a risk mechanism they consider insufficiently studied. A company may support regulation that also favours its commercial position. A government may prioritise geopolitical competition. A worker may fear losing their job. A creator may seek control over the use of their work.

None of that automatically makes their arguments true or false.

But it explains why referring to “the experts” or “the industry” as if either were a single actor is not enough.

Nor does everyone have the same capacity to exert influence.

Those controlling infrastructure, finance, distribution channels and access to policymakers have instruments unavailable to an independent researcher or a small business.

The question therefore cannot be confined to who is right about the future.

We must also ask who decides what gets developed, who can use it, who pays for its consequences and who has a genuine ability to challenge those decisions.

AI is already entering the culture war

The American debate shows clear signs of politicisation.

Trump has presented warnings about certain AI risks as a fear campaign and placed international competition at the centre of his response. Axios reporting.

But translating that into a pro-AI right and an anti-AI left would be too simple.

There are initiatives backed by representatives of both parties. In September 2025, for example, Republican Josh Hawley and Democrat Richard Blumenthal introduced a proposal to evaluate the risks of advanced systems. That does not mean they agree on every aspect of technology policy, but it does contradict the idea of a perfectly ordered partisan divide. Hawley and Blumenthal's proposal.

Moreover, “regulating AI” can encompass very different issues: child protection, copyright, national security, employment, surveillance, competition or risks of losing control.

A person can support restrictions in one of those areas and oppose them in another.

The danger of a culture war appears when those differences are replaced by a test of belonging.

If my group identifies AI with progress and freedom, every warning will look like a manoeuvre by enemies of development.

If it identifies AI with corporate concentration and job destruction, every beneficial application will look like technological propaganda.

From that point on, what a system does is no longer being examined. Instead, the question becomes which tribe appears to benefit from acknowledging it.

And importing that map automatically into Norway or Spain would be a mistake. Institutions, economic interests and political coalitions are not identical.

Politics is necessary; manipulation is not

Artificial intelligence affects how power, work, wealth and responsibility are distributed. Political debate is therefore natural.

Pretending that all those decisions are purely technical would be another way of concealing political decisions.

Manipulation begins elsewhere: when a possibility is presented as a certainty, the conditions of a test are omitted, only the data favouring one position are selected or intentions are attributed without evidence.

It also occurs when an extreme threat is used to avoid uncomfortable questions about present problems.

And when an extraordinary promise is used to justify not asking those questions.

Fear can capture attention. So can the promise of wealth.

No coordinated conspiracy is needed for those incentives to distort public debate. It is enough for the boldest claims to travel better than explanations that distinguish conditions, probabilities and limits.

There are prophets of paradise too

Prophetic language does not belong only to those announcing disasters.

It also appears in promises that AI will solve practically any economic or social problem and that we only need to accelerate its development for the benefits to reach everyone.

A technical capability does not, by itself, determine how its results will be distributed.

Making a task cheaper can open extraordinary opportunities. It can also displace work, concentrate an activity or increase dependence on whoever controls the infrastructure.

Potential productivity and actual well-being are not the same thing.

Between them lie decisions about ownership, competition, training, organisation and access.

A prediction of inevitable abundance therefore deserves the same scrutiny as a prediction of inevitable destruction. Not because the two necessarily have the same probability, but because neither is proved by the conviction of the person announcing it.

Six things we should avoid confusing

To navigate this debate, I suggest a simple distinction:

This separation does not eliminate disagreement. It makes it more useful.

It lets us recognise a risk without accepting every forecast about it. It also lets us recognise an opportunity without buying the entire commercial narrative surrounding it.

Above all, it prevents us from treating a public statement as though it were already an operational safety measure.

My position is not equidistant. It is practical

I do not think the answer is to stand halfway between those announcing paradise and those announcing hell.

The midpoint between two claims has no special guarantee of being correct.

My position starts with a different question:

What should we learn and change now, given the capabilities and limitations we can already observe?

A business cannot suspend its learning until specialists agree about superintelligence.

Nor should a professional confuse prudence with staying away from tools that could transform their work.

Adapting requires experimentation.

Choose a specific task. Understand how it is done today. Try an alternative. Compare the results. Discover what errors arise and which part of the responsibility still requires human judgement.

That might mean preparing a presentation from your own material, organising public information, comparing versions, building a small internal application or coordinating work through several agents.

The aim is not to accumulate subscriptions. It is to learn what can be achieved, with which resources and under what conditions.

That includes measuring time and cost, but also quality, errors and review effort. A faster process that forces you to rebuild everything afterwards does not necessarily represent an improvement.

And it means training people to direct and assess that work, rather than simply press a button.

Controls are part of that adaptation

Experimentation does not mean granting indiscriminate access to an organisation's systems.

We can start with public or artificial information, restricted tools and reversible tasks. Grant additional access when there is a specific reason and previous operation has been checked.

If an agent can act, it matters which identity it operates under, what data it sees, which tools it uses and how a task can be cancelled.

Decisions with significant consequences for third parties need an explicit authorisation policy. The system itself should not decide to expand its permissions because it believes that will help it complete the assignment.

Nor is writing “act ethically” in an instruction enough.

A behavioural framework can guide the model. Technical controls must limit what it can actually execute. None of these mechanisms is infallible, which is why multiple layers, logs and review are needed.

This is a central part of the work we explore at Juss og KI: learning to build useful systems without handing them authority we do not know how to supervise.

Safety is not an argument for giving up on technology. It is a condition for using it with greater confidence and ambition.

We do not need to know the ending to begin

Warnings about extreme risks deserve research. Decisions about the most capable systems cannot be reduced to an advertising campaign or a private conversation among their manufacturers.

But the rest of society has work to do as well.

Learn. Test. Improve processes. Develop judgement. Identify dependencies. Decide what deserves to be automated and what must remain under human control.

That retains its value even if some predictions are delayed or never come true.

If the transformation accelerates, having acquired those capabilities will matter. If it proceeds more slowly, the learning will still be useful, provided it rests on verified results rather than promises.

My concern is not choosing the right prophet.

It is that discussion of the future becomes an excuse for failing to understand the present: some paralysed by fear, others dazzled by promises and many waiting for someone to tell them what to think.

Adaptation demands something more uncomfortable: learning enough to begin deciding for yourself.

We do not need to know how this story ends to start acting thoughtfully within it.