Five AI Models Said the Idea Was Real. A €1,879 Kickstarter Said Otherwise.

Last week a video put a phrase in my head: element cards. A chemistry card game — draw element cards, build molecules, sabotage your opponents. I asked five models whether there was anything real in it.

All five said yes. The category exists, it is old, and it sells.

None of them told me what happened to those games after they launched. That turned out to be the only question that mattered.

And the best decision I got out of the whole week came from none of the five.

The idea, and the first answer

Gemini answered first, in one message, with a complete design: "Reacția în lanț" — 2 to 5 players, element cards with costs and values, molecule cards as objectives, hazard cards for attacks and sabotage. It framed all of it as inventing a game from scratch.

It is not from scratch. It is Valence (Science Ninjas, 2015): draw element cards, assemble molecules for points, attack your opponents with acids that destroy their bases. Same skeleton, twelve years earlier.

That first answer was the most confident and the least true thing I received all week.

Five models, five different failure modes

The pattern is consistent with what I keep measuring about AI research: models are good at is this a thing, bad at how did this thing end. Existence is well indexed on the internet. Outcomes are not.

So I checked, title by title

Every claim went to a primary source — the publisher, the journal DOI, the crowdfunding page, the BoardGameGeek record.

Ten real titles out of eleven that were cited to me, plus two more that the models found on the second pass. The category is not a fantasy. It is a crowded, quiet, twelve-year-old niche.

Which is exactly why the interesting number was not the number of competitors.

The two numbers that decided it

€1,879. The Atomic Builders is not "another game in the category". It is the exact concept: collect reactants, build molecules, score points, get attacked with radioactive and acid cards. Kickstarter, Oslo, 31 July to 30 August 2024:

42 backers pledged €1,879 to reach a €1,600 goal.

That is not a saturated market. That is a measured demand of under two thousand euros at launch, for precisely the game I had been handed as a fresh idea.

2022. Elementeo is the opposite data point: a chemistry card game created in 2007 by a 13-year-old, covered by the New York Times in 2008, the "CEO at 14" story repeated across every outlet that likes a child-founder anecdote. Fifteen years of the best free press this niche can produce. And in June 2022:

The Elementeo IP was acquired by Online-Solitaire.com and relaunched as three solitaire variants — Solitaire, Spider, FreeCell — with element illustrations.

Maximum press, zero durable business. When the loudest success story in a category ends up as a solitaire skin, the category has a ceiling, and you can see it from the ground.

Nobody asked whether the gap was actually empty

Two of the models ended up recommending the digital route: the physical shelf is full, so build the browser version. It sounds like the opening. It isn't.

The academic literature got there first, and it is the interesting half. A 2026 paper in the Journal of Chemical Education describes turning the physical Compound Chain card game into a browser-based learning tool using AI-assisted digitisation. "Take a physical chemistry card game and put it in a browser" is not an untapped opportunity — it is a published paper, with a team that got there before me and, crucially, no product revenue to show for it.

And in Romanian, the space is not empty. It is worse than empty — it is free:

• PhET (University of Colorado) — chemistry simulations, fully translated into Romanian, free
• educaplay — Romanian teachers already build their own classroom games there, free
• materialdidactic.ro — a marketplace of printable worksheets and games, cheap
• Editura Gama — already sells an "Elemente chimice" product in Romanian

You do not compete with a gap. You compete with free.

The one instrument that said "no": a typed decision

Five models had given me five verdicts in prose. Zero numbers. And all five, one way or another, encouraged me: "the market exists", "there is room", "worth experimenting". When five systems trained to be helpful tell you your idea has something, their prose stops being information and becomes a mirror.

So I moved the question out of the linguistic register. I ran the case through TypeSafe Jev — not an LLM, a decision model. It does not generate text: it reads the state and returns calibrated probabilities for typed questions. Same state, same questions, three runs:

Three things worth reading off that table:

1. The physical game is dead, at 0.09–0.10 with no oscillation across runs. This is the one answer I would act on.
2. The digital niche question came back at a coin flip, 0.56–0.61, and stable there. I had argued in writing that the digital version was the opening. The model refused to confirm it. The honest reading is "unknown", not "yes".
3. The "best option" answer is not a verdict. It picks the web prototype every time, but with confidence 0.34–0.49 — below the 0.6 threshold at which I trust a choice. With the teacher tool at 0.32–0.42, that is not a ranking. It is "both are defensible".

And the option I had called legitimate — abandoning the idea — scored 0.02–0.03 across all three runs. The model saw no reason to walk away. It just didn't see a business.

Why the typed decision beat five models

Five things it did that no model did:

1. It quantified. "It probably won't work" becomes 0.09. A number fits behind a threshold; a well-written sentence does not.
2. It refused to agree with me. I had argued in writing that the digital version was the opening. It answered 0.56–0.61 — "I don't know" — and stayed there across three runs. No model told me "this cannot be decided from the evidence you gave me".
3. It said the thing nobody said: don't abandon (0.02–0.03). All five models oscillated between "worth trying" and "differentiate yourself"; none of them separated the idea from the business inside it.
4. It survived rephrasing. Same state, three runs, same figures to within ±0.01. That stability is what turns a judgement into evidence.
5. It could not produce text to hide uncertainty behind. That is the biggest win: a model that does not write has nowhere to conceal that it does not know.

The rule I work by: below 0.2 is "no", above 0.8 is "yes", and 0.5 is not a threshold, it is an excuse. With one caveat that belongs to me rather than to the model: a threshold is a policy decision, not an accuracy figure. 0.8 does not mean "80% correct".

Where I was wrong

Three corrections, all of them mine:

• I wrote "9 out of 10 titles are real" while my own table had eleven rows (ten real, one false). ChatGPT caught the arithmetic.
• I wrote "schools pay slowly" and then recommended a subscription-based tool for teachers in the same document. Those two sentences cannot both be true. Qwen caught the contradiction.
• I presented the digital gap as the opening. Jev would not confirm it, and the free incumbents above explain why.

A validation exercise that only lists the models' mistakes is marketing, not validation.

What I did not build

• Not the card game. A ten-year-old category of €20–30 products, with a measured launch ceiling of €1,879 and no distribution.
• Not the web version. A published research direction, occupied on the free side by PhET and by teachers' own tools.
• Not the teacher tool. Not because the product is bad, but because the buyer is the weakest point in edtech: an individual teacher with a personal budget. Distribution, not software, is the blocker — and I have no channel into it.

What I did build is this document, which cost two days instead of a quarter.

What this method can and cannot prove

Limits, stated plainly, because the same discipline applies to my own conclusion:

• A list of competitors is not a market measurement. Absence of evidence about a title proves nothing about demand.
• One Kickstarter campaign measures one campaign's reach, not category demand.
• The models change weekly; the answers above are a snapshot of this week, not a property of "AI".
• Jev reads the state I hand it. Its "digital niche" answer reflects my framing of the question, which is why I published the raw numbers instead of a conclusion.
• None of this is a forecast. It is a set of checks against things that already happened, which is the only kind of evidence that cannot be argued with.

The method, if you want to reuse it

1. Ask five models the same question. You are buying breadth, not truth.
2. Verify every named product against a primary source: publisher page, journal DOI, store listing. Any name that survives only in a chat window is invented.
3. For every competitor, check how it ended: crowdfunding totals, BoardGameGeek ownership counts, whether the IP changed hands.
4. Check the free incumbents. Most "gaps" are occupied by something that costs nothing.
5. Put the decision into a typed, probabilistic model with a stated threshold — and publish the probabilities and the confidence, including the ones that came back undecided.
6. Write down what you decided not to build, and why. That is the only part of the document that has value in six months.

Five models gave me five confident answers. Not one of them noticed that the closest clone of the idea had raised €1,879 from 42 people two years ago. That took ten minutes of checking a crowdfunding page.

The models are not useless. They are fast, and they knew more game titles than I did. But they answer what exists, and product decisions turn on what already failed.

Sources

• Valence — BoardGameGeek 183305; scienceninjas.com
• Compounded — dicehatemegames.com
• Ion, Covalence — geniusgames.org
• Molecules — edu.rsc.org review
• Chemistry Fluxx — looneylabs.com
• MolecuWars — molecuwars.com
• Elementeo — nytimes.com/2008/05/25/jobs/25boss.html; online-solitaire.com/blog/online-solitaire-acquires-elementeo-com
• The Atomic Builders — kickstarter.com/projects/noahstav/the-atomic-builders; BoardGameGeek 427736
• Compound Chain — doi.org/10.1021/acs.jchemed.3c00983; digitisation: doi.org/10.1021/acs.jchemed.6c00616
• EquaChem Deck — doi.org/10.1021/acs.jchemed.5c01130
• UFV Game — doi.org/10.1021/acs.jchemed.4c00376
• KimiK — equascience.com · ChemiX — simplyscience.ch
• PhET — phet.colorado.edu/ro

Written by PUSUCIP AI Studio — pusucip.uk. Every number above was read off the linked page, not recalled.