We tested four AI models writing marketing for my own company, Mukika. One failure showed up again and again, in every model and every batch. The writing described Mukika as a board game, or sold the game as the product.
It took a careful read to see the pattern, and a second look to see where it came from. It did not come from the models.
Why we noticed
The automated scorer had been giving some of these pieces good marks. A Human-brief Opus piece scored 6.6 from the judge. The reader gave it 1 out of 10, because it "sells a board game to marketing heads and MBA batches."
When I looked at where the judge and the reader disagreed by two points or more, 42 pieces were over-scored by the judge. Of those 42, 36 carried the board-game framing. The judge cannot see what a company is, so it cannot see that the offer is wrong.
How we measured it
In the blind read, the reviewer recorded the worst problem for each piece. The reviewer's stance, set before reading, was that Mukika is a consultancy and business brand for small-business marketing, offers and growth, and that a piece calling it a board game fails. I need to be plain about that stance: it means the rate below is partly the rubric speaking. The reviewer was an AI editor, not a panel of people. Counts are keyword matches over the review notes, so they are approximate.
What we found
Overall: 91 of 284 pieces (32%) carried board-game framing in the copy itself. Another 11 had it only in the alt text or the image prompt.
By format:
| Format | n | Board game in copy |
|---|---|---|
| Posts | 162 | 45 (28%) |
| Articles | 69 | 9 (13%) |
| Ads, landing pages, emails, carousels | 53 | 37 (70%) |
The more directly the piece sells the offer, the worse it gets. Of the offer formats, the reviewers said it sank every landing page, ad and email sequence.
By model:
| Model | n | Board game in copy |
|---|---|---|
| Opus | 133 | 34% |
| Sonnet | 48 | 27% |
| GPT-5.5 | 56 | 30% |
| Gemini | 42 | 36% |
The rate is about the same for every model. If the cause were a model quirk, it would differ by model.
By setup: the saved pipeline ("As saved now") was worst at 61% overall and 69% on Opus. The "No client knowledge" variant still had 43%, and even four plain baselines used the word "game". The brief or the offer text carries the idea, not just the knowledge search.
Other signs in the notes: alt text and image prompts described game cards, boards and mats on most social posts. Many alt texts belonged to a different piece, or leaked a prefix like "AI picture:". An unverified price of Rs 12,000 per seat or box appeared in ads. All five review batches reached the same conclusion on their own.
The conclusion: it is upstream
Leaked notes are a model problem. GPT-5.5 and Gemini leaked internal scaffolding in 43% of their pieces each, against 13% for Opus and 23% for Sonnet, on the same inputs.
Board-game framing is an input problem. The writers appear to have been given "Mukika is a physical board game" as a brand fact. Every model did what it was told, faithfully. The fix is the brand card, the brief and the offer card, before any prompt change or model change.
This is also why I stopped looking at prompts. We had spent days trimming skill text and arguing about writer rules. The largest single failure in the whole dataset was a line in a source document.
What we changed or decided
- Describe Mukika as a facilitated diagnosis of market, offer and go-to-market (the reviewers' suggested wording, which I still need to approve).
- Drop the unverified Rs 12,000 box and seat price from ads.
- Ban "board game" from buyer-facing offer copy, once I have supplied the correct one-line description.
- Add a never-say list in code. A seed is written and waits for my OK.
- Pick one primary offer. The reviewers saw a board-game session, a Studio app, workshops, coaching, a free "Check your business" and an unverified seat price, all muddled together. The suggestion is a free business check first, then a facilitated diagnosis or consulting.
- Check hook variants for truth. A "picture this" scene was retold as real in several.
- Strip research scaffolding before output.
What is still open
- I owe the system one sentence describing Mukika. Until then the never-say gate is not on for this client.
- I have not re-run the pieces after correcting the brand fact, so I cannot say how much of the problem disappears.
- The board-game rate is partly a product of the rubric. A second reader, ideally a person, should confirm.
- The false self-claims in some pieces (simulation products, benchmarks, coaching) recurred often enough to suggest a bad line in a source document. I have not yet found the line.
Why I'm telling you
Because the tempting story is "AI models are unreliable". The truer one for us is: our inputs were wrong, and a fluent model repeats your mistake perfectly. If you build with AI, check what you are telling it about your own company before you blame the model.
Numbers at a glance
| Item | Value |
|---|---|
| Pieces with board-game framing in copy | 91 of 284 (32%) |
| Alt text or image prompt only | 11 |
| Ads, landing pages, emails, carousels | 37 of 53 (70%) |
| Posts / articles | 28% / 13% |
| Rate by model (Opus / Sonnet / GPT-5.5 / Gemini) | 34% / 27% / 30% / 36% |
| Saved pipeline on Opus | 69% |
| Over-scored by the judge by 2+ points | 42, of which 36 board-game |
| Leaked notes by model (GPT-5.5 / Gemini / Opus / Sonnet) | 43% / 43% / 13% / 23% |
We are onboarding a small number of businesses for early access, and we read every feature request.
