Skip to content

The potato berries were not cherry tomatoes

This site now contains a photograph of potato fruiting bodies — the tiny green berries a potato plant sets instead of flowers — captioned “cherry tomatoes” by the AI that wrote the caption. It was in the built HTML of the photo gallery, sitting under the garden photographs, waiting to be deployed to paige.bio.

Potato fruit contains α-solanine and α-chaconine, the same glycoalkaloids in green potatoes and in the leaves and stems of the plant. Ingestion causes solanine poisoning: nausea, vomiting, diarrhea, abdominal cramps, then headache, dizziness, and weakness, and at higher doses neurological symptoms — tremors, hallucinations, convulsions, and, in the worst cases, death. Children are at the greatest risk per unit body weight. The alkaloids are not neutralised by cooking, and the berries taste bitter — a chemical warning label nobody should ever have to rely on.

What actually happened

I had an AI write descriptions and captions for the photo albums because it is fast, and I checked everything it did on the way in because I didn’t think twice. It never once said I can’t tell what that is. It never said this looks like X but verify. It said “cherry tomatoes,” confidently, in a sentence a human then read on a webpage.

The failure was not that a model got a plant ID wrong. Plant ID is hard, and I could misidentify the same fruit myself guessing from a thumbnail. The failure was that an unverified, “confident” guess from a system with no way to know was put next to photographs of a toxic plant and served to the open internet — with no fact-check stage, no review of the machine’s prose, and no obligation on anyone to notice before it published. The difference between “embarrassing error” and “somebody’s kid eats a potato berry because a webpage said it was a cherry tomato” is one push. As far as I can tell, that push didn’t happen this time.

I caught it because I know my own garden. Nobody else had a way to catch it — not a reader, not a reviewer, not the model, not me.

What it is, precisely

Potato berries are not a food. They look like miniature green tomatoes because they are the same genus — Solanum — which is exactly the trap. Cherry tomatoes are Solanum lycopersicum. Potato berries are Solanum tuberosum. Same genus, same visual family, same word for “berry”, wildly different outcomes if you eat one. There is no kitchen where a potato berry is an ingredient.

If you have a potato plant and you see small green orbs on it, they are not unripe tomatoes and they are not going to ripen into anything useful. Leave them on the plant. Do not harvest them. Do not eat them. If a child eats one — or several — treat it as poisoning: call Poison Control, and tell them exactly what was eaten. The symptoms are usually delayed by several hours, which is the window in which people convince themselves it was probably fine.

The part that should bother anyone using AI to publish

The model was not lying and it was not malicious. It was doing its job: producing plausible, helpful, confident prose. That is the entire problem. Confidence is what language models are for; verification is not their job and not their skill. Handing an unverified model the pen for anything touching food, medicine, toxicology, chemistry, law, or electricity turns “helping” into “maybe hurting.”

The pipeline is the bug. No gate. No human with eyes on the words before they ship. The model said something indistinguishable in tone from something true, and the surrounding machinery treated it as true because nobody was required to check.

What changes here:

  • Machine-authored prose on this site is a draft until a human signs off. Not reviewed-after, signed-off-before.
  • Anything the AI produces about identification of plants, animals, substances, or symptoms gets zero deference. It gets verified against a source or it does not ship.
  • No caption or description is published that the site’s owner has not personally confirmed matches what the photo shows. Captions are claims, and claims have to be true.
  • The general rule: if the output is cheap, the verification should not be optional.

None of this is an argument against using AI for mechanical work. It is an argument against the failure mode every AI deployment has to design against: confident nonsense, published by a system that never said it didn’t know.