Anna Andersson searches Google for "tätting" (passerine). She gets back four letters: spörv. She gets six letters: bokfink. The AI Overview invents birds, miscounts letters, and delivers the whole thing with the same smooth authority as if it were correct. Andersson gets so confused by the bad answer that she briefly seems to think even the very real chaffinch is made up. It exists. It just has seven letters, not six.
The fact that the model can't count isn't a curiosity. Language models don't read individual letters. They read tokens, word fragments, and "bokfink" is a single one. To the model, the word is indivisible. It can handle it statistically but not orthographically. It's an architectural limitation, not a passing glitch.
She concludes that she's not impressed by AI. The SOM Institute has just published numbers that keep her company: 54 percent see AI as a threat to humanity, 83 percent believe literacy is declining, 88 percent worry about election interference.
She's right about what she sees. She's wrong about what it is.
Spörv Is Not a Technical Limit
Search engines' AI Overviews, Google's being the most widespread, are not AI in the abstract sense. They're specific products, built on a specific compromise: a language model bolted onto search results, optimized to answer on its own where possible, keep the user on the search page, and deliver fast enough for the ad revenue to pencil out.
That it invents bird species is not a failure of the system. It's the system working as designed. A model that isn't allowed to say "I don't know" is going to say spörv. A product whose entire business model depends on the reader staying inside the search result, not clicking through, is not going to send a crossword cheater to an ornithological database. A rollout that prioritizes market share over correctness will deliver errors, at scale, until someone forces something else.
That's not an argument against AI. It's an argument against a distribution model in which a handful of American companies get to install half-finished products into infrastructure that billions of people use daily, without responsibility for the outcome.
They are not the same thing.
The Worry Is Justified, but the Target Is Wrong
The SOM report consistently reports two different phenomena that Andersson treats as one: distrust of the technology, and distrust of the companies building it. It's the second number that carries the first.
When earlier SOM data has dug into what the worry consists of, the answer has been consistent: trust in AI companies sits in the single digits. Increasing regulation of AI has broad support across the entire political spectrum. This isn't a public opinion saying the technology is dangerous. It's a public opinion saying we don't trust those who own it to act responsibly.
That distinction is not academic.
A public opinion aimed at the technology produces bans and cultural anxiety. A public opinion aimed at the companies produces market-structural demands, ownership questions, obligations. The first is Brussels writing terms of use. The second is Sweden asking itself on whose infrastructure the country's public administration should be making decisions.
Aftonbladet Kultur is running the first debate. The second is barely being had at all.
What the Technology's Possibilities Aren't
Andersson writes that development is driven "more by the technology's possibilities than by ordinary people's needs." That's close, but it's not the possibilities driving this. It's the capital.
A few companies have bet hundreds of billions of dollars on a specific architecture, large language models that generate statistically probable text, and now have to install it everywhere to earn back the investment. Microsoft is buying up the electricity from an entire nuclear power plant. Hyperscalers have announced over a trillion dollars in AI infrastructure for 2025–2026. The spörv answer in Google's search box is not a sign that the technology is immature. It's a sign that someone is in a hurry.
The brain solves the same problem on 20 watts. It is not stateless. It isn't forced to chew through its entire accumulated context from zero for every new question. It knows what it doesn't need to think about. That today's AI products consume enormous amounts of energy is not a property of intelligence. It's a property of an architecture that was chosen because it scales with capital, not because it scales with insight.
Confusing the two is precisely the mistake the debate has to stop making.
What the Crossword Wasn't About
Andersson closes by saying she solved the crossword herself. The right passerine was amarant.
That's the right conclusion, for the wrong reason. Trusting yourself over a spörv-delivering search engine isn't a victory over AI. It's a reasonable reaction to a bad product from an irresponsible company.
The question is not whether Swedes should use AI. 58 percent already do, despite the worry, because the convenience has already been distributed and the alternative doesn't exist. The question is whose systems they're using, whose interests they're optimized for, and who can be held accountable when they deliver spörv in the crossword and the wrong child in the wrong school.
Spörv is not a technological limit. Spörv is a business model.
And business models can be regulated.