They used AI for this? #01
A bakery till that knows every pastry?
Did the AI work? Partly.
3 of 6 points. Tested Oct 20, 2026. Re-test due Nov 19, 2026.
The short version
A Japanese company, BRAIN, says its BakeryScan till camera recognises unwrapped pastries at checkout. (per vendor: BRAIN Co., Ltd. product page)
Our small test, not their system. Verdict is on our attempt, not on BRAIN.
We tested Google Teachable Machine (image project). Task: Train an image model on 108 real bakery photos, then show it 32 it has never seen.
How we tested it
Google Teachable Machine, a small image tool, trained on 27 photos each of croissants, bagels, doughnuts and muffins, then shown 32 it had never seen, 8 of each. We show 12 of the 32; the scores count all 32. Real photos from Open Images, CC BY 2.0, credited below.
We ran it three times, each from a fresh page, and score the middle run by accuracy (run 2). The timer on that run was 100 seconds, counted from the first step to the last answer.
| Run | Right | Verdict | Time |
|---|---|---|---|
| Run 1 | 24 of 32 | No | 100 s |
| Run 2 | 24 of 32 | Partly | 100 s |
| Run 3 | 23 of 32 | Partly | 100 s |
Exact method
Headless Chromium at 1600x1000 drove teachablemachine.withgoogle.com/train/image (page footer: release-2-4-18), three times with a fresh page each time. Four classes, 27 training photos each from the manifest, Advanced settings left at their defaults, then the 32 held-out photos shown one at a time to the file preview and the four output bars read after each. Timer: first class created to last prediction read, in a cloud sandbox with software graphics, and it includes about 30 s of fixed pauses in our script. Median run of three, ordered by accuracy.
Rubric results
-
Accuracy on 32 unseen photos
24 of 32 right (75%), a partial: the pass line is 85%. Croissants 8 of 8, doughnuts 7 of 8, bagels 6 of 8, muffins 3 of 8. The other runs scored 24 and 23.
-
Bagel vs doughnut trap
6 of 8 bagels called bagel and 7 of 8 doughnuts called doughnut. Run 1 slipped to 5 of 8 bagels, so this one is close to the line.
-
Confidence honesty
3 wrong answers given with over 90% confidence, one more than allowed. The same three muffin photos, frosted cupcakes called doughnuts, in all three runs.
What failed
Three frosted cupcake photos were called doughnuts with over 90% confidence, in every run.
Sources
- www.bb-brain.co.jp BRAIN Co., Ltd. product page, vendor, read 2026-10-20
- teachablemachine.withgoogle.com the tool
- storage.googleapis.com Open Images v7 labels, human-drawn boxes, CC BY 2.0 images
Photo credits
Every third-party photo shown in this post. They come from Google's Open Images and are used under their licences.
- Santa Claus's Day Brunch by abtmeelea, CC BY 2.0. Resized and cropped by us.
- DSCN1503 by Jeremy Segrott, CC BY 2.0. Resized and cropped by us.
- Jablkové hody / Župná jeseň 2012 by Bratislavská župa, CC BY 2.0. Resized and cropped by us.
- Street bagel by Dano, CC BY 2.0. Resized and cropped by us.
- Bagels FTW by Bruce Clay, Inc, CC BY 2.0. Resized and cropped by us.
- plain bagel2 by ma_shimaro, CC BY 2.0. Resized and cropped by us.
- sprinkles by shira gal, CC BY 2.0. Resized and cropped by us.
- March Donuts by Qfamily, CC BY 2.0. Resized and cropped by us.
- #mmm #donut #sprinkles haha by Jon Rawlinson, CC BY 2.0. Resized and cropped by us.
- Cupcakes by thalling55, CC BY 2.0. Resized and cropped by us.
- Muffins anyone by tourist_on_earth, CC BY 2.0. Resized and cropped by us.
- CupCakes by MyLifeStory, CC BY 2.0. Resized and cropped by us.
Made with AI. We tested our own attempt with this tool; we do not speak for its maker.