How Uniquely Bad Is AI, Really?

Disclosure: this blog uses AI heavily, and says so: the meter at the top of every post estimates the human share of each stage, from the idea to the final text. That is a reason to be careful with this question, not a reason to avoid it.

The question

A lot of people have concluded that AI is not just risky but bad in a special way: thirstier, dirtier, more dishonest and more dangerous than other technologies. Others think all of that is panic, like the fears that greeted the printing press, the telephone and the internet.

My answer: AI is not uniquely bad across the board, but it is unusual, and the unusual part is not where most of the public argument is. Both camps tend to treat “AI’s harms” as one lump. Sorting them takes four questions:

1. Scale: is the harm large per use, or only large because the use is enormous?
2. Likelihood and severity: how probable is it, and how bad if it happens?
3. Coverage: has an older technology caused the same kind of harm, and do the safeguards we built then actually cover this case?
4. Recoverability: if it goes wrong, can it be stopped, compensated or undone?

None of these alone decides whether to worry. A familiar harm can still be severe; a novel one can be trivial. The questions are there to separate the cases.

Per prompt, the energy charge is weak for Google’s text prompts

A common everyday charge is that AI queries waste huge amounts of electricity and water. The most detailed public figures come from Google, which estimated that in May 2025 its median Gemini Apps text prompt used about 0.24 Wh of energy and 0.26 mL of water for data-centre cooling¹. (“Median” here is Google’s method: it ranks its models by average energy per prompt and reports the average of the model serving the middle prompt.)

Axes: each dot is the energy one activity uses, in watt-hours, on a logarithmic scale where each labelled step is ten times the previous one; the values are in the labels. The household figures are illustrative estimates: a kettle heating a litre from 20 °C to boiling at about 93% efficiency, a 9 kW electric shower, an electric car using 17 kWh per 100 km. Look for: the gap between the blue dot and the kettle. Takeaway: boiling a litre of water uses roughly as much energy as 400 median Gemini text prompts.

That comparison has firm limits. It is Google’s own figure, for one provider’s median text prompt, under Google’s accounting: it excludes training the model, networking and your device, and the water figure leaves out water used to generate the electricity and to make the hardware. Long reasoning tasks, image and video generation and agents that run for minutes use many times more. So the fair verdict is narrow: for Google’s median text prompt, the energy used per prompt is small. Whether any use is worth its energy is a separate judgement, and the figure says little about heavier uses.

The total is a different question. The IEA estimates data centres used about 415 TWh in 2024, around 1.5% of world electricity, and projects about 945 TWh by 2030, driven mainly by AI². That is a real increase, and it is concentrated: a cluster of data centres can strain one region’s grid and water supply even when the global share is small. But it is a familiar kind of problem, comparable to other growing industries, with known remedies (clean electricity, efficiency, siting, dry cooling). It also connects to a question I look at in another piece: data centres want power around the clock.

Real, with precedents: harms we have met before

Several serious charges against AI have close historical parallels:

• Jobs disrupted. Precedents: mechanised farming, factory automation, computers in offices. Tools we have: retraining, unemployment insurance, time. For the people in those transitions, the losses were often lasting.
• Cheap misinformation. Precedents: pamphlets, tabloid press, social media. Tools: provenance, reputation, platform rules, media literacy.
• Biased decisions. Precedents: credit scores, actuarial tables, biased human judges. Tools: anti-discrimination law, audits, rights to an explanation.
• Copyright and data. Precedents: photocopiers, video recorders, music file-sharing, search engines. Tools: court cases, licensing schemes, new law.
• Market concentration. Precedents: railways, Standard Oil, telecoms, big tech platforms. Tools: competition law, regulation, open standards.

A precedent tells us where to start, not that the old tools will be enough. The US standards body NIST, for example, lists ways AI risks differ from those of ordinary software, such as harmful bias that is hard to detect and systems that change after deployment³. And job disruption could be faster than past waves, because AI touches cognitive work in many occupations at once. These harms are serious; what they don’t require is a wholly new category of thinking.

Genuinely different: breadth, speed and delegated action

What sets AI apart is not that any single property is unprecedented. Automation has acted without humans before: NASA’s Remote Agent software planned and executed spacecraft operations on Deep Space 1 in 1999⁴, and in 2012 a faulty automated trading system at Knight Capital sent millions of unintended orders within about 45 minutes⁵. What I think is new is a combination of three things, each a change of degree, that earlier general-purpose software and automation did not bring together at this breadth.

1. Breadth. Most technologies do one thing. A nuclear plant makes electricity; a car moves people. AI is a general-purpose technology like electricity or the computer, but one that applies to thinking tasks: writing, coding, analysing, planning, persuading. That makes its harms hard to list in advance, because its uses are.

2. Speed of improvement and spread. Capabilities have jumped noticeably from year to year, and tools reach hundreds of millions of users within months. Institutions that normally adapt over decades (education, law, labour markets, scientific publishing) are being asked to adapt within years.

3. Delegated action for everyone. Remote Agent took a team of NASA engineers to build for one mission. Agentic AI lets anyone hand over multi-step digital tasks (writing and running code, browsing, sending messages, operating software) and check only the result. Its reliability on long, open-ended tasks is still limited, but the range is growing. Two concerns follow:

• Misuse with less skill. A system that can walk someone through a complex task step by step can lower the expertise needed for harmful tasks too: cyberattacks or dangerous biology, a major concern in several frontier labs’ safety policies. This is a dual-use problem, like chemistry or nuclear physics, with the expert built in.
• Systems that pursue the wrong goal. The alignment problem is that a system can learn to pursue something that looked right during training but isn’t what its makers wanted. With a chatbot, that produces a wrong answer. With an agent that has access to money, code or infrastructure, it produces wrong actions.

How seriously do experts take the worst case? In a large survey of researchers who publish at top AI venues, run in 2023, the 1,321 who answered the relevant question gave a median 5% probability that advanced AI leads to outcomes as bad as human extinction or permanent severe disempowerment of humanity⁶. That is a subjective forecast about technology that doesn’t exist yet, and answers shifted noticeably with how questions were worded. Most of the same researchers expected good outcomes to be more likely than bad ones. It is not a prediction of doom; it is evidence that many people with relevant expertise think the risk is not negligible.

Recoverability: a separate question from novelty

It is tempting to say the familiar harms are reversible and the new ones are not. That isn’t true. A worker displaced by automation may never recover their earnings; a pathogen release or an attack on a power grid was possible long before AI. Recoverability is its own question, and it applies to every harm on the list.

What AI may change is how likely or how widely accessible some irreversible harms become: if it lowers the skill needed to cause them, or concentrates power in a way that removes the means to correct it. That is the same reasoning used for nuclear power in the safety piece: a technology can be very safe on average and still deserve special rules because its rare failures are catastrophic. Averages cannot settle how to weigh a fat tail.

The four questions, worked through once

Take one charge: AI will help someone create a dangerous pathogen.

1. Scale: a single successful misuse could affect thousands or millions of people, far beyond the person who asked.
2. Likelihood and severity: almost every query is harmless, and malicious users are very rare, so the likelihood is uncertain. In May 2025 Anthropic switched on stricter safeguards for one of its models because it could no longer clearly rule out meaningful help with such weapons⁷; forecasts of real-world risk differ widely. The severity, if it happened, could be enormous.
3. Coverage: biosecurity rules control labs, materials and some technical assistance (export controls already cover expert help, for example), but they were not built for expert-level advice available instantly to anyone; model safeguards are new and their effectiveness is hard to verify from outside.
4. Recoverability: a released pathogen cannot be recalled.

Uncertain likelihood, extreme severity, partial coverage, no recovery: that combination justifies precaution (testing models before release, restricting the most dangerous help) even without knowing the exact odds. Run the same questions on “AI text prompts waste energy” and you get a small per-prompt harm, an estimated total for all data centres (the share from text prompts is unknown), existing tools, and a harm that can be cut going forward even though past emissions can’t be undone: worth managing, not worth alarm.

So, how uniquely bad is AI?

• Not uniquely bad across the board. It is a tool built and used by people. Some popular charges, like the energy cost of Google’s median text prompt, are weak per use, and most others have clear historical parallels, even if AI may amplify them.
• Not just another technology either. It combines breadth across thinking tasks, fast improvement, and delegated action available to almost anyone. Each existed before; in my assessment, their combination at this breadth is what’s new.
• Where I put my worry, as a judgement under real uncertainty: misuse that could cause mass harm, such as help with dangerous pathogens or attacks on critical infrastructure, and the loss of human oversight as more consequential actions are delegated to systems that are not reliably aligned. Some of the capabilities involved are demonstrated today; the most severe scenarios are forecasts. What puts them first for me is the combination of extreme severity, safeguards that are still unproven, and no way back.

The four questions work for any claim, not only for AI: how big per use and in total, how likely and how bad, already covered or not, and recoverable or not? A harm that is severe, poorly covered and hard to undo deserves attention however familiar it sounds, and even when its likelihood is uncertain.

---

1. Elsworth, C. et al. (2025). Measuring the environmental impact of delivering AI at Google scale. arXiv:2508.15734 ↩

2. International Energy Agency (2025). Energy and AI. iea.org/reports/energy-and-ai ↩

3. NIST. AI Risk Management Framework, Appendix B: How AI risks differ from traditional software risks. airc.nist.gov ↩

4. NASA Jet Propulsion Laboratory. Remote Agent: frequently asked questions (Deep Space 1, 1999). jpl.nasa.gov ↩

5. U.S. Securities and Exchange Commission (2013). In the Matter of Knight Capital Americas LLC, Release No. 34-70694. sec.gov ↩

6. Grace, K. et al. (2024). Thousands of AI authors on the future of AI. arXiv:2401.02843 ↩

7. Anthropic (2025). Activating AI Safety Level 3 protections. 22 May 2025. anthropic.com ↩