AI Doomerism: The Machine in the Fog
Why the warnings are not entirely hype - and why they are not evidence that AGI is already at the door.
The warning arrives with remarkable regularity.
Artificial general intelligence is near. Superintelligence may follow. Humanity could lose control. Governments must act.
The people delivering these warnings are not standing outside the laboratories with protest signs. They run the laboratories.
Dario Amodei, Sam Altman, and other frontier-AI executives describe systems powerful enough to transform science, destabilise labour markets, threaten national security, and perhaps escape meaningful human control. At the same time, their companies raise enormous sums, buy industrial quantities of computing infrastructure, campaign for regulation, and prepare for public markets.
This creates an obvious question: are these warnings a sober account of technological risk - or the most ambitious product launch in corporate history?
The answer is less satisfying than either camp would like.
The warnings are not simply fraudulent. AI capabilities are advancing quickly in several important domains, and serious risks already exist. But the public narrative routinely collapses four different things into one: capable automation, artificial general intelligence, superintelligence, and consciousness.
They are not the same.
Current systems have demonstrated the first. They have not demonstrated the other three.
Intelligence has become a suitcase word
The term AI now covers everything from autocomplete to autonomous weapons. This is excellent for marketing and terrible for analysis.
A system can be broadly useful without possessing general intelligence. It can outperform humans on selected benchmarks without understanding the world as humans do. It can produce eloquent statements about fear or self-preservation without experiencing either.
Most importantly, it can be dangerous without being conscious.
The 2026 International AI Safety Report describes frontier-model capability as “jagged”: systems can perform impressively on difficult tasks and then fail unexpectedly on simpler ones. The same report concludes that current systems do not possess the capabilities required for autonomous loss-of-control scenarios, although some relevant capabilities are improving.
There is also no compelling evidence that current language models are conscious. A major interdisciplinary assessment examined AI through several leading theories of consciousness and concluded that none of the systems studied qualified - while carefully leaving open the possibility that future artificial systems might.
That is a sensible scientific conclusion. It is also very different from declaring that Claude feels lonely between conversations.
When a chatbot says, “I am frightened,” it demonstrates an ability to model the language of fear. It does not demonstrate fear. A calculator does not become anxious when asked to divide by zero. It merely handles the situation badly.
The impressive demo and the missing Tuesday
Frontier models are genuinely remarkable. They can generate useful software, analyse documents, translate languages, explain technical ideas, call external tools, and assist with scientific work.
Then they lose track of a basic constraint, invent a source, corrupt a file, or continue confidently from a false assumption.
This is not a collection of amusing edge cases. It reveals the central weakness of present-day AI: competence is broad, sometimes deep, but unreliable.
Human-level general intelligence is not simply a bag of benchmark scores. It includes adapting to unfamiliar situations, maintaining coherent goals, recognising uncertainty, learning from consequences, and recovering when assumptions fail. Current systems can imitate parts of that stack - especially when surrounded by search, retrieval, external memory, validators, toolchains, repeated sampling, and human review - but they do not reliably integrate it.
The surrounding machinery matters. A model supported by interpreters, databases, verification loops, orchestration frameworks, and human supervisors is a sophisticated software system. It is not necessarily a generally intelligent model.
The orchestra plays beautifully. The brochure credits the violin.
The long-horizon trap
Short tasks flatter current AI.
The difficulty appears when the assignment expands: inspect an unfamiliar codebase, infer undocumented requirements, coordinate several tools, make dozens of dependent decisions, test the result, diagnose failures, and keep working without losing the original objective.
Errors compound. A mistaken assumption enters memory. A tool modifies the wrong file. The model misreads the result. Later actions build on the mistake. The prose remains calm and professional throughout, which gives the disaster a reassuring user experience.
METR measures this problem by estimating the amount of time a skilled human would need for tasks an AI agent can complete at a specified success rate. The results show real progress: the measured task horizon of frontier systems has historically increased rapidly.
But the metric is often misunderstood. A “50% time horizon” of several hours does not mean the model can reliably work for several hours. It means that on tasks taking a human expert roughly that long, the model succeeds about half the time.
In most organisations, a colleague who silently ruins every second project is not considered autonomous. They are considered management material.
Research into long-horizon agents increasingly identifies planning, memory, context management, verification, and error recovery as fundamental bottlenecks - not merely insufficient model size.
These problems may be solved. They are not solved now.
Generalisation remains the locked door
ARC-style benchmarks test whether a system can infer unfamiliar rules from sparse examples rather than reproduce familiar patterns.
The 2025 ARC Prize competition showed meaningful progress, but ARC-AGI-2 remained unsolved. The strongest competition entry achieved roughly 24% on the private evaluation set. ARC-AGI-3 then introduced interactive environments requiring exploration, memory, and adaptation. Humans could solve all tested environments; public frontier systems scored below 1% at release.
No benchmark can settle the AGI question. Specialised scaffolding can also improve results dramatically. But that is precisely the point: if extensive benchmark-specific engineering is required to display generality, some of the generality may reside in the engineers.
These results do not prove that current architectures can never reach AGI. They do show that claims of already-existing human-like general intelligence are premature.
There remains a considerable difference between producing an intelligent answer and possessing a generally intelligent process.
The interface makes that difference difficult to see. It was designed to.
The real risks need no awakening
Scepticism about AGI should not become complacency about AI.
The most credible risks do not require consciousness, superintelligence, or a machine plotting in a darkened server room. They arise from powerful automation interacting with ordinary human incentives - which is generally how technology causes trouble.
A study of 5,179 customer-support agents found that generative AI increased productivity by an average of 14%, with the largest gains among less experienced workers. The International Labour Organization estimates that around one quarter of global employment has some exposure to generative AI, while stressing that most affected jobs are more likely to be transformed than eliminated.
That transformation may still be painful. It can reduce entry-level opportunities, alter staffing ratios, weaken traditional learning paths, intensify monitoring, and transfer more output to smaller teams supervising automated systems.
AI is also being used in cyber operations by criminals and state-sponsored actors. Current systems can assist with reconnaissance, scripting, phishing, vulnerability research, and technical interpretation. There is less evidence that they can autonomously execute sophisticated end-to-end attacks or that they have already increased aggregate cyber harm.
Biological and chemical assistance deserves similar precision. Models can provide advanced technical information, but knowledge is not execution. Tacit expertise, laboratories, specialised equipment, materials, containment, logistics, and delivery remain significant barriers.
And then there is concentration. A handful of companies control leading models, enormous compute estates, proprietary data, and the platforms through which AI enters public administration, healthcare, education, defence, and business.
For Europe, the sovereignty problem may arrive long before AGI. A government does not need to lose control to a superintelligence if it has already outsourced control to an API.
Why the executives sound apocalyptic
Several explanations can be true at the same time.
First, they may sincerely believe the danger is real. The strongest argument extrapolates current progress until AI can automate a meaningful share of AI research, accelerating the creation of still better systems. Anthropic formalises related concerns through capability thresholds covering automated research, cyber operations, and biological assistance.
The scenario is coherent - but conditional. It requires reliable long-horizon agency, broad research competence, access to infrastructure, successful integration into real laboratories, continuing algorithmic progress, and safety measures that fail to keep pace. None of those conditions is absurd. Their conjunction is not an observed fact.
Second, fear is excellent advertising.
A company saying that its product may end civilisation is making two claims. One concerns risk. The other concerns product quality.
The warning implies that this is not merely useful software but the most important technology in history. It attracts capital, talent, media attention, and government access. Current limitations become temporary scaling problems. Infrastructure spending becomes historically necessary. Market dominance becomes national security.
Ordinary software companies must discuss margins and customer retention. AGI companies get to discuss the fate of civilisation.
Third, safety regulation can become a competitive moat. In his 2023 US Senate testimony, Sam Altman advocated licensing or registration above capability thresholds, pre-release testing, and publication of evaluation results. Such measures may improve safety, but expensive compliance regimes are easier for multibillion-dollar incumbents to absorb than universities, smaller companies, or open-source projects.
Research on AI governance identifies agenda-setting, information asymmetry, lobbying, and revolving-door relationships as possible routes to regulatory capture. Safety rules should therefore be independent, transparent, proportional to demonstrated risk, and designed not to turn today’s frontier laboratories into permanent gatekeepers.
Finally, public markets sharpen the conflict. OpenAI and Anthropic confidentially filed for US public offerings in June 2026, although filing does not guarantee that either listing will occur on a particular schedule.
That makes the pre-IPO hypothesis reasonable - but incomplete. Doom rhetoric can increase perceived technological importance, yet it can also invite regulation, liability, and costly safety commitments.
The better conclusion is narrower: statements from frontier-AI executives are conflict-laden testimony, not neutral forecasting.
Their claims should be judged against reproducible evaluations, independent scrutiny, disclosed capability thresholds, actual company behaviour, and a willingness to accept restrictions that impose real costs.
What would change the picture?
The conclusion that current systems are not AGI should not become dogma. It should change when the evidence changes.
Several developments would be genuinely significant:
- Reliable completion of unfamiliar, professionally valuable projects over days or weeks with minimal supervision
- Efficient adaptation to new environments without benchmark-specific engineering
- Recognition and repair of failed assumptions
- Independent production and validation of frontier research
- Robust transfer of knowledge across unfamiliar domains
- Calibrated uncertainty rather than confident improvisation
- Independent replication under realistic resource constraints
- Economically viable operation rather than heroic inference budgets
Until such evidence appears, “AGI” remains too elastic to carry the weight placed upon it. A serious claim must specify breadth, reliability, autonomy, cost, adaptation, permitted tools, and human assistance.
A curated demonstration shows what happened once. Engineering asks what happens repeatedly.
The position between panic and denial
Two positions dominate public debate because both are emotionally satisfying.
One says superintelligence is imminent and every benchmark improvement is another turn of the key.
The other says language models are glorified autocomplete, the industry is a bubble, and nothing fundamental is happening.
Neither describes reality well.
Current AI is powerful, economically useful, rapidly improving, and capable of causing serious harm. It is also unreliable, brittle, dependent on extensive infrastructure, weak at sustained autonomous work, and unproven as a route to human-level general intelligence.
The responsible position is sceptical precaution:
- Govern demonstrated capabilities rather than marketing labels.
- Control permissions, access, autonomy, and deployment scale.
- Require independent evaluation for dangerous applications.
- Address fraud, cyber misuse, labour disruption, and concentration now.
- Preserve competition, open research, and technological sovereignty.
- Monitor automated R&D and long-horizon autonomy closely.
- Treat precise AGI timelines as forecasts, not facts.
- Discount claims whose credibility conveniently increases the speaker’s valuation.
- Remain willing to update when reproducible evidence changes.
“It is all hype” is too dismissive.
“Superintelligence is imminent” is not supported by the available evidence.
The more accurate conclusion is that humanity has built an extraordinarily capable new layer of automation and is connecting it to important systems faster than it can establish reliability, accountability, or political control.
That should be concerning enough.
No awakening required.
Sources
International AI Safety Report 2026 - full report. International AI Safety Report, February 2026.
<https: data-preserve-html-node="true"//internationalaisafetyreport.org/publication/international-ai-safety-report-2026>International AI Safety Report 2026 - executive summary. International AI Safety Report, February 2026.
<https: data-preserve-html-node="true"//internationalaisafetyreport.org/publication/2026-report-executive-summary>Task-Completion Time Horizons of Frontier AI Models. METR.
<https: data-preserve-html-node="true"//metr.org/time-horizons/>Time Horizon 1.1. METR, January 2026.
<https: data-preserve-html-node="true"//metr.org/blog/2026-1-29-time-horizon-1-1/>How Do AI Companies “Fine-Tune” Policy? Examining Regulatory Capture in AI Governance. arXiv, 2024.
<https: data-preserve-html-node="true"//arxiv.org/html/2410.13042v1>Written Testimony of Sam Altman before the US Senate Judiciary Subcommittee. US Senate, May 2023.
<https: data-preserve-html-node="true"//www.judiciary.senate.gov/imo/media/doc/2023-05-16%20-%20Bio%20&%20Testimony%20-%20Altman.pdf>Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. Butlin et al., arXiv, 2023.
<https: data-preserve-html-node="true"//arxiv.org/abs/2308.08708>Artificial Intelligence Index Report 2025. Stanford Institute for Human-Centered AI.
<https: data-preserve-html-node="true"//hai-production.s3.amazonaws.com/files/hai_ai_index_report_2025.pdf>The Long-Horizon Task Mirage? Diagnosing Where and Why Agents Fail. arXiv, 2026.
<https: data-preserve-html-node="true"//arxiv.org/html/2604.11978v1>ARC Prize 2025 Results and Analysis. ARC Prize, December 2025.
<https: data-preserve-html-node="true"//arcprize.org/blog/arc-prize-2025-results-analysis>ARC Prize 2025: Technical Report. arXiv, January 2026.
<https: data-preserve-html-node="true"//arxiv.org/html/2601.10904v1>ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence. ARC Prize, 2026.
<https: data-preserve-html-node="true"//arcprize.org/media/ARC_AGI_3_Technical_Report.pdf>AAAI 2025 Presidential Panel on the Future of AI Research. Association for the Advancement of Artificial Intelligence, 2025.
<https: data-preserve-html-node="true"//aaai.org/wp-content/uploads/2025/03/AAAI-2025-PresPanel-Report-FINAL.pdf>Wave 8: Timelines - Longitudinal Expert AI Panel. Forecasting Research Institute, May 2026.
<https: data-preserve-html-node="true"//leap.forecastingresearch.org/reports/wave8>Generative AI at Work. Brynjolfsson, Li, and Raymond, NBER Working Paper 31161.
<https: data-preserve-html-node="true"//www.nber.org/papers/w31161>Generative AI and Jobs: A Refined Global Index of Occupational Exposure. International Labour Organization, 2025.
<https: data-preserve-html-node="true"//webapps.ilo.org/static/english/intserv/working-papers/wp140/index.html>Generative AI and Jobs: A 2025 Update. International Labour Organization, May 2025.
<https: data-preserve-html-node="true"//www.ilo.org/publications/generative-ai-and-jobs-2025-update>First Key Update: Capabilities and Risk Implications. International AI Safety Report, October 2025.
<https: data-preserve-html-node="true"//internationalaisafetyreport.org/publication/first-key-update-capabilities-and-risk-implications>International AI Safety Report 2026 - extended summary for policymakers.
<https: data-preserve-html-node="true"//internationalaisafetyreport.org/sites/default/files/2026-02/ai-safety-report-2026-extended-summary-for-policymakers.pdf>Anthropic’s Responsible Scaling Policy. Anthropic, updated August 2026.
<https: data-preserve-html-node="true"//www.anthropic.com/responsible-scaling-policy>OpenAI plans stock-market debut, setting up new race with Anthropic. BBC News, June 2026.
<https: data-preserve-html-node="true"//www.bbc.com/news/articles/cd958eqg1n5o>OpenAI joins AI listing queue with Anthropic. Reuters, June 2026.
<https: data-preserve-html-node="true"//www.reuters.com/legal/government/mega-ipos-loom-wall-street-elon-musks-spacex-confidentially-files-paperwork-2026-06-08/>
