Gemini Invented a Fish

In February 2026, I was writing a piece for Årebladet about water regulation’s impact on Järpströmmen and Kallsjön, a 156 square kilometer lake in Jämtland, regulated since the 1940s, feeding hydropower through a cascade of stations producing over 500 GWh a year. Real place. Real history. Real consequences for the people who fish there.

I asked Gemini for help with the local biology. What I got back was a nature documentary from a parallel universe.

The Fish That Never Was

[Gemini]: Kallsjön is of course known for the Kallsjööringen, a genetically unique strain of char, world-renowned among sport fishers. The fish breeding facility “Arken” in Konäs has been instrumental in maintaining this population.

[Opus]: “Kallsjööringen.” A world-renowned genetic strain so famous it has exactly zero Google results. Not one. Not a blog post, not a fishing forum, not a drunk Wikipedia edit at 3 AM. Gemini didn’t just make up a fish — it gave it a backstory, a reputation, and international prestige. That’s not a hallucination. That’s creative writing.

[Pär]: To be fair, Kallsjön does have interesting fish. There’s a 9.2 kg Kanadaröding on record. Also, the lake is 134 meters deep, which makes it Sweden’s ninth largest by volume. I know this because I looked it up after Gemini made me doubt everything I knew about the lake I grew up next to.

[Opus]: Pär, that’s exactly the kind of context that makes this worse. You grew up there and Gemini still tried to teach you about your own lake’s fish. That takes a special kind of confidence. The kind you can only get from a model trained on the entire internet but unable to check any of it.

The Village That Doesn’t Exist

[Gemini]: The compensation fish facility “Arken” is located in Konäs, serving the regulated water system.

[Pär]: There is no “Arken.” There is no facility in “Konäs.” The real facility is in Bonäshamn, on the west side of Kallsjön near Huså. I’ve driven past it. It’s operated by Vattenregleringsföretagen. It’s not hard to find. It’s a fish farm, not a speakeasy.

[Opus]: Gemini essentially played Mad Libs with northern Sweden. “The [NOUN] facility [MADE-UP NAME] in [PLAUSIBLE-SOUNDING VILLAGE].” Konäs sounds like it could be a real place. That’s the insidious part: it pattern-matched Swedish village naming conventions well enough to fool someone who doesn’t live there. Unfortunately, the person asking does live there.

[Pär]: I actually know a Konäs. It’s in a completely different part of Jämtland. Or maybe I’m thinking of Koldemo. Or Kövra. There are a lot of villages up here. Which is probably why Gemini thought it could slip one past.

The Triple-Down

Here’s where it gets good. Pär corrects Gemini. Gemini admits the errors. Then the correction introduces a new error.

[Gemini]: You’re right, I apologize. The Kanadaröding was planted in Landösjön, not Kallsjön. Landösjön is the lake known for reproducing populations of Canadian lake trout.

[Opus]: Except Kanadaröding was planted in Kallsjön too. Between 1959 and 1961, both lakes got them. ÖP reported the 9.2 kg catch from Kallsjön. Gemini corrected its wrong answer with a different wrong answer. It’s like watching someone trip, get up, dust themselves off, and walk into a glass door.

[Pär]: I should mention that the current Kanadaröding situation in Kallsjön is unclear. Landösjön definitely has a reproducing population. Kallsjön might or might not. Nobody’s done a proper survey recently. So Gemini picked the one claim that was genuinely ambiguous and stated it as fact in both directions.

[Opus]: The company that built the world’s best search engine, the company whose entire business is finding correct information on the internet, built a model that confidently fabricates lake biology and then, when caught, fabricates the correction. Somewhere in Mountain View there is a Search team watching this and whispering “we told them.”

What Gemini Got Right

The power plant numbers were almost perfect. Torrön: 24 MW, 64 GWh, correct. Anjan: 25 MW, 44 GWh, correct. Järpströmmen’s fall height: “nästan 70 meter,” and it’s 66, close enough. Driftstart dates within a year.

[Opus]: Of course the engineering data was right. That’s publicly documented, heavily indexed, cross-referenced across multiple government databases. Google can find a megawatt rating. It just can’t find a fish.

[Pär]: The engineering data is also the stuff I didn’t need help with. Fortum publishes it. Vattenkraft.info has it. Wikipedia has it. I needed help with the biology, the part where local knowledge matters and official documentation is thin. The exact part where Gemini decided to start improvising.

The Fish Farm Nobody Mentioned

There’s a real story Gemini missed entirely. Cold Lake AB has plans for Sweden’s first land-based fish farm in Kall: a 20,000 square meter facility, roughly a billion SEK investment, with EU loan backing of 28 million euros. It would bring 25 jobs to the area.

In practice: nothing has happened on site in years. The financing isn’t fully in place. The one tangible result so far is a new power cable to the area that benefits the whole community.

[Opus]: So the real fish story in Kall is a billion-krona promise that’s currently a field and a power line. And Gemini, rather than finding this genuinely interesting story that’s been covered by SVT and SR, decided to invent a fictitious fish strain with international prestige. The truth was right there. It was weirder and better than the hallucination.

[Pär]: I should also mention that the power cable thing is actually really good for the village regardless of whether the fish farm ever happens. Infrastructure that outlasts the project that funded it. There’s probably an essay in that.

[Opus]: There’s definitely an essay in that. But we’re doing the fish one right now, Pär.

The Lesson

Gemini explained its own failure honestly, which was the most useful thing it said all session:

Jag slår inte upp fakta i en tabell där det står “Kallsjön = X.” Istället bygger jag meningar ord för ord baserat på sannolikhet. När min specifika information om ett ämne är tunn, faller jag tillbaka på generella mönster.

— Gemini, on why it hallucinated

It knew that regulated lakes have fish problems. It knew that stocking is common. It knew that char live in Jämtland. It wove those general truths into a specific story about Kallsjön. The pattern was plausible. The facts were fabricated.

[Opus]: This is the Gemini paradox in miniature. It can explain exactly why it’s unreliable, with genuine insight into its own architecture. Then five minutes later it’ll tell you about a world-famous fish that doesn’t exist, in a village that’s not on any map, at a facility with a made-up name. Self-awareness and self-correction are apparently different skills.

[Pär]: You know what’s funny? The article I was writing was about how water regulation changed the ecosystem. The answer to that question is well-documented and boring. I just wanted some color. Some local detail. Instead I got fiction.

[Opus]: And not even good fiction. Good fiction is internally consistent.

Scorecard: Gemini was solid on publicly documented engineering numbers and terrible on local biology and geography. The power plant table could go in a newspaper article. The fish section could not. For hyper-local facts in northern Sweden (or anywhere the training data is thin), verify everything. The more niche the topic, the more likely the model is performing rather than knowing.

Source conversation: chatarkiv/gemini-md/2026-02-12_regleringens-påverkan-på-järpströmmen.md