Why is AI bad?
The strongest complaints against AI aren't about robots — they're about confident errors, work taken without payment, false accusations from detection tools, energy demand doubling to about 945 TWh by 2030, and concentration of power in a few companies. The weakest complaint is that it's 'not really intelligent.' That one doesn't matter.
Why — the first-principles explanation
Most "AI is bad" arguments collapse into one mechanical fact: a language model is trained to produce text that looks right, not text that is right. Truth isn't in the objective function. The model learns the statistical shape of correct-sounding answers, and a fabricated citation has exactly the same shape as a real one. That's not a bug that gets patched — it's what the training procedure rewards. So the model is fluent and confident whether or not it knows, and human beings are wired to read fluency as competence. That mismatch is the root of a huge share of real AI harm.
The second real complaint is economic, not technical. These models were trained on text and images made by people who weren't asked, weren't paid, and can't opt out retroactively. Then the output competes with those same people. Whether that's legal is being fought over in court; whether it's a transfer of value from many creators to a few companies is not really in dispute. Add the capital requirement — training frontier models costs hundreds of millions in compute — and you get an industry where about three to five companies can play. Concentration is a structural outcome of the physics of training, not a conspiracy.
The third is misplaced trust in AI systems that judge people. AI detectors are the cleanest example: seven of them flagged 61% of English essays by non-native speakers as machine-written, while performing near-perfectly on U.S.-born eighth graders. Same tool, same score, wildly different reliability depending on who you are. That pattern — an average accuracy that hides concentrated failure on a specific group — repeats in résumé screening, fraud flags, and risk scoring.
And the resource cost is real but routinely exaggerated in both directions. The IEA puts data centers at about 415 TWh in 2024, roughly 1.5% of world electricity, roughly doubling to 945 TWh by 2030. That's a genuine strain on specific local grids and a fast-growing emissions source — and it's still a small slice of global energy. Both halves of that sentence are true, and most articles print only the half that fits.
What's not a good argument: "it's just autocomplete" or "it doesn't really understand." A tool doesn't need understanding to displace your job or get your loan denied. Judge it by what it does.
An example that makes it click
Imagine a coworker who has skimmed every book in the library, never says "I don't know," and speaks with total confidence in a lovely voice. Ninety percent of the time he's right and genuinely useful. Ten percent of the time he invents a court case, a dosage, or a statistic — in the exact same tone as the true stuff.
He isn't lying. Lying requires knowing the truth. He's doing the only thing he was ever trained to do: produce the sort of thing that usually turns out to be correct. Now imagine your company fires the fact-checkers because he's so fast. That's the actual problem — not the coworker, but the ten percent, times the speed, times nobody double-checking.
Key facts
- Data centers used about 415 TWh in 2024 — roughly 1.5% of global electricity — and the IEA projects a rise to about 945 TWh by 2030, near Japan's total consumption today (IEA, Energy and AI, 2025).
- Data center CO2 emissions are about 180 million tonnes today, projected to reach 300 Mt by 2035 in the IEA Base Case, staying below 1.5% of total energy-sector emissions.
- Seven GPT detectors classified 61.22% of TOEFL essays by non-native English writers as AI-generated, while performing near-perfectly on essays by U.S.-born eighth graders (Liang et al., Stanford, 2023).
- In a 2023 survey of 2,778 AI researchers, the median chance given to AI causing human extinction or severe permanent disempowerment was 5%; 38–51% gave at least 10%.
- The same survey found no researcher consensus that AI is net bad: 48% of those who expected good outcomes still gave at least a 5% chance of extremely bad ones.
▶ The 60-second explainer (script)
Why is AI bad? Skip the robot stuff. Here's the real list. Number one, and it's mechanical: a language model is trained to produce text that looks right — not text that is right. Truth isn't in the objective. A made-up citation has the exact same shape as a real one, so the model sounds equally confident either way. And humans read confidence as competence. That single mismatch causes most real AI harm. Number two: the training data came from people who weren't asked and weren't paid, and now the output competes with them. Number three: tools that judge people. Seven AI detectors flagged sixty-one percent of essays by non-native English speakers as machine-written — while being nearly perfect on American eighth graders. Same tool, different reliability depending on who you are. Number four: energy. Data centers were about one and a half percent of world electricity in 2024, doubling by 2030. That's a real strain on real grids — and still a small slice of the total. Both are true. Now the bad argument — it's just autocomplete, it doesn't really understand. Doesn't matter. It doesn't need to understand to get your loan denied. Judge it by what it does.
What authoritative sources say
People also ask
Is AI actually bad, or just new?
Both framings are lazy. It's a tool with unusually high fluency and unusually low reliability signaling — a combination that hasn't existed before. The novelty is real; so are the harms.
Isn't 'AI uses tons of water and power' the main issue?
It's real but oversold. Data centers were ~1.5% of global electricity in 2024. The honest concern is the growth rate and local grid strain, not that your chatbot session is melting the planet.
Do AI researchers think AI is bad?
No. The 2023 survey found optimism and alarm living in the same people — most expected net good outcomes while still assigning meaningful probability to catastrophe.
What's the single most underrated harm?
Automation of judgment about people — hiring, cheating accusations, credit, benefits — where average accuracy hides concentrated failure on specific groups, and there's no appeal.
Is it bad that AI 'doesn't really think'?
No, and it's a distraction. A calculator doesn't think either. What matters is what the system does in the world and who's accountable when it's wrong.
The same question, asked other ways
- Is AI bad?4,400/mo
- How is AI bad?590/mo