Q. Everyone’s talking about AI now. Do you think I should start using it?
A. That depends. You can, of course, but should you? There are benefits, yes, but there are worrying costs that most people don’t see, and fanboys (and fangirls) don’t want to talk about.
Q. Hang on. When you say “AI” and I say “AI”, are we talking about the same AI?
A. A good starter. AI covers a lot of ground. Some of it is task-specific and practical, like the systems the NHS uses: analysing X-rays to support radiologists, predicting patient flow, or organising hospital schedules so care runs smoothly. These are Narrow AI tools, which don’t create or invent. Instead, they analyse data and help people make faster, better-informed decisions. Others, e.g. Generative AI, create content: text, images, music or video. And that’s what we’re really talking about here, because that raises questions for anyone whose work depends on originality: writers, artists, designers and more.
AI costs: small numbers, big picture
Q. Costs? What costs?
A. The first is environmental. Every time you type a query into ChatGPT, it uses upwards of 0.003 kilowatt-hours of electricity. That’s around fifteen times the power used by a standard search engine.
Q. That doesn’t sound like much.
A. Not for one question, no. But scale it up. Globally, ChatGPT is asked around 250 billion questions every day. That adds up to roughly 750 gigawatt-hours daily, enough electricity to power 85.5Mn UK homes for a day.
And the UK has only around 28 million homes. The UK could keep the lights on and kettles boiling for three days with the energy currently spent daily on people asking AI to write poems about their cats. And that’s just ChatGPT. Then there’s Gemini, Claude, Copilot, DALL-E, Llama, Perplexity, MidJourney… and more.
Q. That’s a bigger number than I expected. And a lot more energy too!
A. That’s just the electricity. Every query run also requires between two to five litres of water to keep the servers cool. Plus, on your device, a single session can use around 100 to 300MB of memory. One prompt may not seem much, but billions of them a day add up fast.
AI and employment: whose job is it anyway?
Q. Fair enough. What about the effect on people?
A. That’s already happening. Jobs are being lost or reshaped. Copywriters, designers, customer service agents, skilled masons; many of these roles are being undercut by AI tools. Not because AI is better, but it’s faster and cheaper, although not always right (see later).
Q. Isn’t that just how technology works?
A. Sometimes. But this shift to AI has been abrupt and the results are mixed. Look at Klarna, the payments company. They laid off around 700 customer service staff in 2023 and replaced them with AI. At first it looked like a big saving. But customer satisfaction dipped, badly. They are now rehiring people to restore quality, but in more precarious, flexible roles. So real people lost stable jobs, customers had a worse service, and the supposed efficiency wasn’t.
Q. So it backfired.
A. Exactly. The other problem is de-skilling. If AI always writes the first draft or generates the quick design, when do the juniors learn? Expertise only grows with practice. Take that practice away, and you hollow out the very skills you think you are saving.
Q. And I guess access is uneven too?
A. Yes. The best AI models are locked behind subscriptions. Those who can afford them get more speed and features; those who cannot are left behind.
Enter your data, surrender your soul
Q. Let’s say I used it for harmless stuff. Do I still need to worry about privacy?
A. Yes. In 2023, Samsung engineers pasted proprietary source code into ChatGPT to get debugging help. That code left their control the moment they pressed Enter. Companies say they anonymise or protect data, but the details are vague. The fact is, once you give something to AI, you’ve handed it to someone else’s system. You cannot pull it back.
Q. I wouldn’t put secrets in it though.
A. You say that, but even ordinary text can reveal more than you think. And what feels harmless today may not feel the same later.
Does AI work?
Q. Alright, but if I keep it general, does it at least work well?
A. Only if you ask properly. How you phrase a prompt makes all the difference. Ask vaguely, you get vague answers. Ask clumsily, you get clumsy results.
Q. So I just need to practise?
A. Yes, but that’s the point. There’s now a whole industry being built around ‘prompt engineering’ to teach people how to ask the right way. If the tool that promises to make life easy still demands you learn a specialist skill, is it really working?
Q. But if I get it right, the answer should be accurate?
A. Not always. AI has a habit of ‘hallucinating’ – or what you and I know as ‘lying’. And it does so with complete confidence. Unless you know the subject well, you may not spot the mistakes.
AI bias: old stereotypes, new paint job
Q. And bias? I keep hearing that word with AI.
A. That’s real. AI systems learn from human data, and that data contains human prejudice. Bloomberg tested image generators in 2023. They asked for a picture of a ‘CEO’ or a ‘doctor’, the majority of outputs were white men. Asked for a ‘nurse’ or a ‘cleaner’? They got women.
Q. So it just generates stereotypes?
A. Yes. Not necessarily because someone coded them in deliberately, but because they’re in the data. And if you use AI enough, those stereotypes get fed back through, until they become normalised all over again.
…And on the bright side?
Q. This is sounding pretty bleak. Are there any sunny uplands?
A. There are. AI can be genuinely useful for accessibility. It can simplify text, read things aloud, or translate quickly. For people with disabilities, those are real gains. It can also help anyone stuck with a long report to summarise or generate a list of ideas to get moving again.
Q. So there are good uses?
A. Yes. But for most of us the question is not whether AI can do something; it is whether it should.
And the answer is?
Q. So should I use, or should I no’ noo?
A. My view: only sometimes. If it genuinely adds value, fine. If it saves you from drudgery you cannot avoid, or provides accessibility you need, then yes. But don’t treat it as the default. Every query carries a cost: in energy, in water, in jobs, in privacy, in trust.
Q. And if I decide not to use it?
A. Then you’re not missing much. There’s nothing wrong with our pre-existing tools. They may be slower, but they’re reliable, they don’t (always) lie to us, and you’re more likely to remember the answer.
Q. So where do you stand on it?
A. On the side of caution. AI is powerful, but it’s not magic, it’s not neutral, and it’s not free. I’d still rather think slowly with my own words than outsource my voice to a machine. This way of working may not be on trend, but it’s more honest and I can do it offline. And if I’m going to write, then I want it to sound like me, not like I ‘prompted’ it.
