A warning that AI could soon cause a temporary power or banking shutdown can be neither proved nor dismissed, but connected systems already fail in ways that justify preparing for such an incident now.
Key takeaways
- A warning reported by the Guardian says AI could trigger a large-scale incident, such as a temporary financial or power blackout, this year or next.
- The timing of any AI-linked infrastructure failure cannot be verified, but the exposure of critical systems to software faults and cyber attacks can be examined today.
- Backups, manual fallbacks and regular outage drills are useful against AI-linked failures and against ordinary software failures alike.
- Whether humanity would have been better off without AI is a separate judgement from whether AI could disrupt infrastructure, and the available evidence cannot yet settle it.
The case for preparing rests on exposure rather than prediction
The Guardian reports that a Nobel prize-winning computer scientist, one of the researchers often described as a godfather of AI because of foundational work on artificial neural networks, has warned that the technology could cause a large-scale incident this year or next. The examples given are temporary blackouts of financial or power systems. According to the same report, the researcher also suggested that humanity might eventually conclude it would have been better off never developing AI, while rejecting the label of doomer and maintaining that the technology could still lead to a very good future.
This article argues for a narrower position than the warning itself. The forecast cannot be checked in advance, and the summary available here does not set out a probability or a precise mechanism. Critical systems nevertheless depend heavily on software, and that dependence can be examined without guessing a date.
The argument runs as follows. Essential services rely on networks of connected digital systems. Those systems already suffer failures that spread beyond their point of origin. AI tools are being added to both sides of the picture, as an aid to people attacking systems and as a component inside the systems themselves. Each of these developments gives operators and regulators reason to prepare, whether or not the specific forecast proves correct.
Preparation also has an unusual advantage as a policy response. Manual fallback procedures, isolated backups and rehearsed recovery plans pay off if a failure comes from an AI system, a human error, a storm or a conventional cyber attack. A government or company does not need to accept a contested claim about AI in order to justify them. That is why the conclusion here is about planning rather than alarm: the risk is credible enough to act on, even though its size remains unknown.
Software faults already spread through connected systems
The first piece of evidence does not involve AI at all. Modern infrastructure has repeatedly shown that a fault in one place can cause trouble far from its source. Faulty software updates, configuration errors and failures at shared service providers have interrupted organisations that had no direct connection with one another, other than relying on the same supplier or platform.
Power and finance are particularly exposed to this pattern. Electricity grids must balance supply and demand continuously, and they depend on control systems and communications to do so. Payment systems depend on chains of banks, processors and networks, so an interruption in one link can stop transactions elsewhere. Both sectors build in redundancy, yet redundancy works best against failures that engineers have anticipated. Shared dependencies are harder to see, because several apparently separate systems may rely on the same underlying component.
This matters for the AI debate because a large-scale incident would not require a superhuman machine. An ordinary fault in a widely used system could be enough if it reached a sufficiently central point. The relevant question is therefore how AI changes the frequency, speed or reach of failures in systems that are already tightly linked. The history of non-AI outages does not show that an AI-linked incident is imminent. It does show that the pathway from a local error to a widespread disruption already exists.
AI tools can lower the effort required for cyber attacks
The second piece of evidence concerns deliberate misuse. Generative AI systems can write and explain computer code, summarise technical documents and produce convincing text in many languages. Those capabilities are valuable to legitimate programmers, and they can also assist people who are trying to find software weaknesses, write malicious code or craft persuasive phishing messages.
National cyber security agencies and AI developers have discussed this dual use openly, and it is a recognised concern in the field. The extent of the effect is less clear. Attacks on critical infrastructure usually require more than working code. An attacker may need knowledge of specialised equipment, access to protected networks and patience to move through them undetected. AI assistance does not remove those barriers, and public reporting on incidents often cannot establish how much any tool contributed.
Even a modest effect could still matter, because of how defence works. Organisations running critical services often operate old equipment that is difficult to update without interrupting supply. A defender has to protect many points continuously, while an attacker needs to succeed at only one. If AI shortens the time needed to find and exploit a weakness, ageing systems may have less time to be patched. Defenders can use the same tools, however, so the net balance is uncertain and may differ from one sector to another. That uncertainty is itself a reason to test resilience rather than assume that defences will keep pace.
Automated decisions can outrun human oversight
The third piece of evidence concerns AI as a component inside systems, rather than as a tool used against them. Organisations are increasingly using machine learning for tasks such as forecasting, fraud detection, trading and network management. More recent AI agents can carry out sequences of actions on a user’s behalf, including running software and changing settings.
These systems can behave in ways that their developers did not fully predict. Neural networks are trained rather than written line by line, which makes it harder to trace why a particular output occurred or to list in advance every situation that might produce an error. A model can perform well in testing and then behave poorly when conditions differ from those it has seen before. Unusual conditions are also the moments when infrastructure is under most strain.
Speed adds a further complication. Human oversight is a common safeguard, but it works only if people have the time and information to intervene. Where automated systems interact with one another at high speed, an error can spread before an operator notices it. Finance has long experience of this problem from automated trading, which is why markets use mechanisms that pause activity during extreme movements. Similar safeguards, including clear limits on what automated agents may change and the ability to shut them off quickly, are a practical response to the same underlying risk.
How far AI has been integrated into the most sensitive control functions of power and banking is not fully known from public information. Operators may have good reasons not to disclose such details, but the gap makes outside assessment harder.
The strongest case against treats the warning as unproven and possibly distracting
The best objection is that no public evidence set out in the available reporting demonstrates that an AI-caused blackout is close. Predictions about AI timelines have often been wrong in both directions, and expertise in building neural networks does not automatically confer expertise in grid operations or bank settlement. People who run those systems may reasonably point out that they already plan for severe failures, operate under strict regulation and keep many sensitive controls separated from public networks.
Critics can also argue that attributing a future outage to AI would be difficult and perhaps misleading. Complex failures usually have several causes, and an incident involving an AI tool may be better explained by poor maintenance, weak access controls or rushed deployment. Treating AI as the headline risk could draw money and attention away from those mundane problems, which may cause far more harm in practice.
A further objection concerns costs on the other side of the ledger. AI may help to detect fraud, forecast electricity demand, find software flaws and speed up recovery after an attack. Fear-driven restrictions could slow those benefits, leaving infrastructure less safe than it might otherwise have been. The suggestion that humanity might have been better off without AI also compares the present with a counterfactual that nobody can observe, so it cannot be tested in any straightforward way.
These objections are serious, and they limit what can honestly be claimed. They do not, however, count against most of the preparation proposed here. Manual fallbacks, better maintenance and strict limits on automated control address the mundane causes that sceptics emphasise as well as the novel ones that the warning describes.
Better incident data could strengthen or weaken this conclusion
Several kinds of evidence would change the assessment. The most direct would be detailed public investigations of serious outages that establish what part, if any, AI systems played. A run of incidents in which AI was a clear and necessary cause would strengthen the argument and justify firmer rules. If the forecast period passes without such an incident, that would count against the specific prediction, although it would not by itself show that the underlying risk is small.
Controlled testing could add useful information. If independent exercises showed that current AI systems give attackers little practical help against realistic infrastructure defences, the misuse argument would weaken considerably. Results showing that they substantially speed up attacks, or that defenders gain a lasting advantage from the same tools, would shift the balance in one direction or the other.
Disclosure by operators and regulators would also help. Sector-wide information on where AI is used, which functions remain under human control and how many systems share the same providers would turn a broad concern into a measurable one. Stress tests that simulate AI-related failures could reveal whether existing fallback plans work. If they did so reliably, the case for additional measures would be much weaker.
The larger claim about whether humanity would have been better off without AI would require something different: a long-run accounting of benefits and harms that does not yet exist.
Sources and further reading
- The Guardian: the report on the warning about AI and large-scale incidents, which is the source of the specific claims attributed here.
- National Cyber Security Centre (UK): further reading on how AI may affect the cyber threat landscape.
- Bank of England: further reading on operational resilience in the financial system.
- National Institute of Standards and Technology (US): further reading on frameworks for managing AI risk.
Surfaced from the rss:guardian_tech signal “AI infrastructure risk warning”. AI-assisted draft, editorially reviewed.
