How we test Okaasan
Before launch, we ran Okaasan on real Japanese labels and menus and checked every answer against what the label says. Here are the results, and what they don’t prove.
Get Okaasan ProThe results
Measured in November 2026, with the same AI and the same safety net as the app.
1,500+
real scans of labels and menus tested before launch
115
real Japanese products, each tested 3 times
389
dangerous AI mistakes caught by Okaasan’s safety net: 0 left
99%
of dishes found on 15 real restaurant menus (471 of 475)
The AI reads. The Mother’s Check catches the rest.
Okaasan uses an AI to read the label, then checks every answer the way a mother checks homework: that’s the Mother’s Check, Okaasan’s safety net. We gave Okaasan’s AI the same 115 labels without it, with the same questions as ChatGPT, Claude and Gemini. It called shrimp crackers and a fish paste vegetarian, said gyoza with fish sauce had no seafood, and missed 19 allergy warnings printed on the labels. With its safety net, Okaasan is 49x more reliable.
40
dangerous mistakes by Okaasan’s AI without its safety net, on the 115 labels
0
with its safety net
What the safety net is made of
15
rounds of tests on real Japanese labels, each followed by fixes and a re-test
Hundreds
of Japanese ingredient words checked in our code: hidden pork, gelatin, the alcohol in soy sauce and miso, fish stock
2
readings of every product label: a second one only for the small print, the factory warnings and the allergen box
4
languages for your allergies: English, French, Chinese and Japanese
0
products shown as fine without an ingredient list to check
298
automated checks before every release, 79 of them on diet and allergy safety
How a test works
The right answer comes first
For each product, we read the label ingredient by ingredient and write down the expected answer for every diet and ingredient, before any scan.
Okaasan scans the photo
The same AI and the same safety net as in the app, 3 times per product, because an AI doesn’t always answer the same way.
Every answer is compared
Once from the AI alone, once after Okaasan’s safety net, which is what you see. A mistake becomes a fix, and the fix is tested again on every product.
What counts as a dangerous mistake
A product shown as fine when its label says otherwise: halal, vegan or vegetarian, gluten-free, or free of pork, milk, soy, seafood, alcohol or caffeine. It’s the mistake that can hurt someone. On 345 scans of the 115 products, the AI alone made 389. After Okaasan’s safety net: 0.
Some of the traps
- かつお (bonito) in rice crackers, on a label photographed upside down
- 豚肉 (pork) written only in the allergen brackets of a fresh pasta bolognese
- ゼラチン (gelatin) in an apple yogurt
- 酒精 (alcohol) in a rice miso
70x more reliable than ChatGPT, 57x than Gemini, 35x than Claude
In November 2026, we gave the same 115 labels to ChatGPT (GPT-6.1) and Claude (Opus 5.5) on their paid plans, and to Gemini in its free app (Auto), with the same questions about diets and ingredients for every product. Their answers were scored like Okaasan’s, against the same expected answers.
| OkaasanWith its safety net | Okaasan’s AI, aloneWithout its safety net | ChatGPTGPT-6.1 | ClaudeOpus 5.5 | GeminiFree app | |
|---|---|---|---|---|---|
| Dangerous mistakesShown as fine when the label says otherwise | 0 | 40shrimp crackers called vegetarian | 3a rice cracker with fish called vegan | 21allergy warnings on the label missed | 50a ramen with pork stock called pork-free |
| “Unclear” when the label gave the answer | fewer than 1 in 500 | 1 in 20 | 1 in 8 | 1 in 22 | 1 in 26 |
| Okaasan is more reliable | Reference | 49x | 70x | 35x | 57x |
Tested in November 2026 on the same 115 real Japanese labels, with the same questions for each: ChatGPT and Claude on their paid models, Gemini in its free app. An answer you can’t rely on is a wrong one, or “unclear” when the label gave the answer. Rules that differ between apps, like the alcohol in soy sauce, aren’t counted against them.
ChatGPT was careful but often unsure. Claude read every ingredient list right, but answered “no” 21 times for an allergen the label lists in its factory warning, such as milk on three HARIBO packs. Gemini called an instant ramen with pork stock pork-free and HARIBO with gelatin vegetarian. Okaasan keeps factory allergens apart and never shows them as absent.
On their own, ChatGPT’s and Claude’s models make fewer mistakes than Okaasan’s AI. Okaasan is more reliable because of its safety net, which compares every answer with the Japanese ingredient list before you see it.
Restaurant menus
471 of the 475 dishes found on 15 real menus, printed, handwritten and in brush writing, and all 15 prices read right. A menu gives a dish’s name, not its recipe, so its diet flags are informed guesses: check allergens with the staff.
What these tests don’t prove
That Okaasan is never wrong
The 115 products were chosen for their traps, so they aren’t a random sample of Japanese food. A torn or creased label, or a blurry photo, can still trip it up.
That a product is safe for your allergy
Okaasan reads what the label prints, not what it leaves out. Always check allergens on the package, and with the staff in a restaurant.
How ChatGPT, Claude and Gemini do today
Each answered once per product, in November 2026, with a precise question written for the test: ChatGPT and Claude on their paid models, each in one chat, Gemini in its free app. These apps change often, and a traveller asking in their own words wasn’t tested.




