Why Is AI Bad? The Real Problems, Without Panic

Why is AI bad? In five ways you can document today, and in a longer list you can only forecast. The documented five are confident wrong answers, scams built on cloned voices, bias that reaches decisions without anyone intending it, an electricity and water bill that's growing fast, and a reshaping of work in exposed occupations. The forecast list, the one that fills headlines, runs from concentration of power to weapons to the extinction scenario, and the experts who rate those risks highest also attach wide error bars to them.

The reason "why is AI bad" gets such messy answers is that most explainers pour all of that into one list. I'm going to grade each harm by the kind of evidence behind it, name the source, and say who actually bears it. Then you can decide which ones deserve your worry this week and which belong to the decade.

The Negatives of AI, Graded by Evidence

The grade column is the point of the table. The first three rows are things you can reproduce this afternoon. The last two are informed guesses by people who disagree with each other, which doesn't make them wrong, but does make them a different kind of claim.

What the Experts Rank Highest, With the Error Bars

In late 2025, researchers at MIT and the University of Queensland ran a three-round Delphi study with 272 international AI experts, published in June 2026. The experts rated 24 risks for severity and for the chance of a catastrophic outcome, which the study defined as more than a million deaths or more than 100 billion dollars in losses, over the five years from 2025 to 2030.

The news write-up that ranks for this search gives the top five. The report gives the intervals, and the intervals matter.

A few things in the full report didn't make the headlines. Under business as usual, the experts gave 18 of the 24 risks more than a 10 percent chance of a catastrophic outcome in five years. With what the study calls pragmatic mitigations in place, five stayed above 10 percent, and two of those five, environmental harm and inequality with unemployment, aren't in the top-five severity table at all. Every one of the 24 kept at least a 5 percent chance.

Then the line I'd underline. The experts judged AI users and the general public the most vulnerable to these risks, and assigned the highest responsibility for fixing them to general-purpose AI developers and to governments and regulators. The people with the least say bear the most exposure. That mismatch is, to my mind, the most useful single finding in the study, because it tells you where pressure has to go.

Does AI Lower Your IQ?

No study I could find measured anyone's IQ before and after using a chatbot. Two studies measured something adjacent, and they're the source of every "AI is rotting your brain" headline from the past year.

The first is Your Brain on ChatGPT, a preprint from MIT Media Lab submitted in June 2025. Fifty-four people wrote essays across three sessions while wearing EEG caps, split into a chatbot group, a search-engine group and a group with no tools. The abstract reports that the no-tools group "exhibited the strongest, most distributed networks," the search group showed moderate engagement, and the chatbot group "displayed the weakest connectivity." The chatbot group also "struggled to accurately quote their own work," and reported the lowest sense of ownership of what they'd written. Only 18 people took part in the fourth session.

The second is a survey of 319 knowledge workers from Microsoft Research and Carnegie Mellon, presented at CHI 2025. Its finding is a correlation from self-reports. Higher confidence in the AI went with less critical thinking, and higher confidence in oneself went with more.

So here's the honest reading. When you hand a chatbot the whole task, you engage less with it, remember less of it, and feel it's less yours. That's what was measured, on one task, in small groups, over months rather than years. Nothing in either study shows a permanent decline, a change in general intelligence, or that using AI to check a draft you wrote yourself does any harm at all. The direction of use is the variable neither headline mentions.

That's also the answer to "why is AI bad for students" and "why is AI bad for creativity." The ownership finding is the one I'd take seriously. If the essay never felt like yours, you didn't learn to write it. My rule is draft first, ask second, and I'd give the same rule to a fourteen-year-old.

Why Is AI Bad for the Environment?

Because the buildings that run it draw electricity and evaporate water, and both are growing faster than the grid or the local utility planned for. The Berkeley Lab report for the US Department of Energy puts US data centre electricity at 176 TWh in 2023, 4.4 percent of the national total, with a 2028 projection of 6.7 to 12 percent. The same report counts 66 billion litres of water evaporated on site in 2023 and nearly 800 billion litres more consumed at the power plants feeding those sites.

Per prompt it's tiny. Google's own estimate for a median text prompt is 0.24 watt-hours and 0.26 millilitres of water. The problem is aggregate and local, and the full accounting, including why the "gallon per prompt" meme is off by a factor of at least 76 against any published figure, is in how AI uses water.

One detail from the Delphi study belongs here too. Environmental harm was one of the five risks the experts kept above a 10 percent catastrophic chance even after pragmatic mitigations. It didn't make their top five for severity. It did make the list of things they don't think will simply be fixed.

Why Is AI Bad for Jobs?

The best measurement I know of isn't a forecast of job losses. It's an exposure index. The International Labour Organization's May 2025 working paper finds that one in four workers worldwide is in an occupation with some degree of generative AI exposure, and that 3.3 percent of global employment sits in the highest exposure category. Women are over-represented in that top band, 4.7 percent against 2.4 percent for men globally, and 9.6 against 3.5 in high-income countries.

Exposure means the tasks overlap with what the tools can do. It doesn't mean the job disappears. The ILO's own reading is that transformation of jobs is the most likely outcome, and I'd trust that over any confident replacement figure.

People also search for which three jobs will survive AI. I won't name three, because nobody can, and any list you've seen came from an interview rather than a measurement. The calmer version of the question, with what to actually do about your own role, is in will AI take my job.

The Scam Problem Is Already Here

Of all the harms in the table, this is the one I'd put on the fridge.

The US Federal Trade Commission's consumer alert from March 2023 describes the mechanism plainly. A scammer needs a short audio clip of a family member's voice, which can come from anything posted online, and a voice-cloning program. Then the call comes. Someone who sounds exactly like your child or grandchild is in trouble and needs money now.

The FTC's advice is short enough to memorise. Don't trust the voice. Call the person who supposedly contacted you, on a number you know is theirs, and verify. If you can't reach them, try another relative or a friend. I'd add one thing the alert doesn't say, which is to agree a private word with your family now, while nobody is panicking, and ask for it on any urgent call.

The tools are three years better than they were when that alert was written. The pattern is identical.

Bias Doesn't Need a Villain

A model learns from records of past decisions. If those decisions were skewed, the model learns the skew and applies it at scale, quickly, with a straight face.

The US National Institute of Standards and Technology published Special Publication 1270 on exactly this in March 2022. Its abstract says biases "remain endemic across technology processes and can lead to harmful impacts regardless of intent." That last clause is the one to keep. Nobody has to want the outcome for the outcome to happen.

My concern with this harm is that it's quiet. A rejected loan or a filtered CV doesn't announce that a model was involved, and the person affected rarely finds out. Of the documented harms, it's the one with the least natural feedback, which is why it needs rules rather than goodwill.

Why Should AI Be Stopped?

The strongest public statement on the far end of the risk spectrum is one sentence. In 2023 the Center for AI Safety published a statement that reads, in full, "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." The signatories shown on the page include Geoffrey Hinton, Yoshua Bengio, and the chief executives of OpenAI, Google DeepMind and Anthropic.

Surveys point the same way without agreeing on a number. In the largest one, covering 2,778 researchers who had published at top AI venues, between 38 and 51 percent of respondents gave at least a 10 percent chance to advanced AI leading to outcomes as bad as human extinction. Read that as it's written. Roughly half of a large expert sample think the worst case is not negligible, and roughly half don't.

Notice the verb, though. The statement says mitigating. The Delphi study assigns responsibility to developers and governments, which presumes the work continues under rules. None of the sources in this article call for stopping, and I couldn't find a serious one that does. "Stop AI" is a slogan; "who is responsible for which risk" is the question the evidence can actually answer.

Bill Gates published an essay whose title calls the risks of AI real but manageable. His page blocked my fetch on the day I wrote this, so I'll leave his argument to him rather than paraphrase it from memory.

What to Worry About This Week, and What Belongs to the Decade

This is judgment, not evidence, so I'll label it.

This week, three habits cover most of the documented harm. Verify before you act on an answer, using the five-minute method in how to fact-check AI, because fluent and true are different qualities and the hallucinations explainer shows how often they come apart. Treat any urgent voice call about money as unverified until you've called back. And do your own first draft, then use the tool to check it, so the work stays yours and the engagement finding stays somebody else's problem.

This decade, the harms that matter are the ones you can't fix at your desk. Where the power comes from and where the water goes. Who owns the most capable systems. Whether the rules catch up with the bias that nobody intended. Those belong to votes, budgets and regulators, and the useful thing a beginner can do is know which is which.

AI is bad in specific, checkable ways and in large, uncertain ones. Keep the two lists separate and neither one will panic you.