If AI Went Down: Which Outage Should We Really Fear?

Darkness in Iberia

On 28 April 2025 the power went out across Spain and Portugal. The report of the expert group of the European network of transmission system operators (ENTSO-E), published in March 2026, concluded that the cause was systemic failures in voltage control, reactive-power management and the regulatory framework, not a single fault. Key voltage-control equipment was connected and disconnected manually, which slowed the response, and some conventional generators delivered less than 75% of the reactive power requested of them at critical moments. The lesson is that great systems fail through an accumulation of small faults, not one blow.

Is AI like electricity?

Electricity is a physical utility, and when it stops every device stops. AI is a software layer that sits on top of electricity, communication networks and cloud computers. So the first outage that threatens it is a power outage. The International Energy Agency estimates that data centres used roughly 1.5% of the world's electricity in 2024, and projections of their growth depend on assumptions; the figures are set out in the article "The AI Race: Who Owns the Chips, the Energy and the Rules?".

Nor is there a single switch that turns "AI" off. Each application depends on its own provider, model and cloud service. That is why the closer picture is cascading partial failures.

Failures that really happened

At 11:48 pm Pacific time on 19 October 2025 a fault began in Amazon's DynamoDB database service in the Northern Virginia region. Its cause was a latent defect in the system that manages DNS records, and the full incident ended on 20 October after more than fourteen hours. Many dependent services were affected, including compute, load balancing and sign-in. On 19 July 2024 a faulty update from CrowdStrike disabled about 8.5 million Windows devices according to Microsoft, less than 1% of Windows machines worldwide, yet the effects reached banks, airports and railways.

The lesson is not specific to AI: when services are concentrated in a few providers, a small fault becomes a big event. Every AI service used through the cloud inherits that risk.

The hidden danger: skill erosion

The worst harm may not be what happens during an outage but what builds up before it. In an observational study published in The Lancet Gastroenterology & Hepatology in 2025, researchers compared adenoma detection rates in 1,443 colonoscopies performed without AI assistance by doctors at four centres in Poland: 28.4% before the tool was introduced and 22.4% after three months of using it, a difference of 6.0 percentage points. The study is observational, so it does not prove that the tool is the cause, and its authors said only that continuous exposure "might reduce" detection. Even so, it is a reasonable warning for every profession: someone who uses the tool every time may lose the ability to work without it.

The answer, and what to do

We do not say "yes, and worse", nor "no cause for concern". An AI outage is more likely to resemble not the darkness of cities but slowdown and confusion in everything we have built on the tool, and its severity depends on the alternatives we have prepared. We take the means and place our trust in God.

• Map your dependence: which task in your work stops if the tool is unavailable?
• Prepare a written fallback for critical tasks, with a manual alternative or a second tool.
• Practise working unassisted at regular intervals, as evacuation drills are rehearsed.
• Do not hand a decisive decision to the tool alone, and review its output with human eyes (see the article "Three Myths About Artificial Intelligence: Fluency, General Intelligence and Neutrality").

It is enough to remember that the tool is a useful servant, and a prudent person does not burn their boats.