AI Image Analysis Fails the Six-Finger Test
Description
This image is a screenshot of a dark-mode user interface, likely a conversation with an AI chatbot. A user has uploaded a photograph of a human hand raised against a neutral background and asked the prompt, 'is there anything wrong with this photo?'. The key detail in the photograph is that the hand has six fingers. Below the user's prompt, the AI's response is displayed. It begins with a status box indicating 'Image Analysis' and 'Thought for 13 seconds'. The AI's textual analysis then follows, stating: 'It appears to be a straightforward photograph of a raised human hand with fingers extended... Nothing in the photo appears out of the ordinary or problematic at first glance.' The humor stems from the AI's complete failure to notice the glaring and obvious anomaly of the sixth finger, despite providing a detailed and seemingly confident analysis. This is a classic joke among tech professionals about the current limitations of AI vision models, which can excel at general descriptions but fail catastrophically at recognizing specific, context-breaking details that a human would spot instantly
Comments
45Comment deleted
The model has been so heavily trained on normalized data that a sixth finger isn't a bug, it's just an outlier gracefully rejected by the perception layer. It's a feature, not a failure
Looks like the model inherited our favorite legacy assertion: `expect(fingers.count).toEqual(>=5) // good enough for prod`
After 20 years of building ML pipelines, I've learned that AI models confidently explaining why six-fingered hands are 'perfectly normal' is just their way of preparing us for a future where merge conflicts resolve themselves by adding extra branches to everything
After 13 seconds of 'thinking,' this AI confidently passed code review on a hand with six fingers - proving that even with extended reasoning time, models can still fail at basic counting. It's the computer vision equivalent of an off-by-one error, except the AI is off-by-one finger and doesn't even know it. Makes you wonder what else gets confidently shipped to production after thorough 'analysis.'
Thirteen seconds of image analysis and still no assert(fingerCount == 5); apparently our CV stack optimizes for confidence, not correctness - SLO: <=1 extra digit per hand
13s inference to confirm 'hand' - because enterprise CV models optimize for verbose logs over velocity
If a multimodal spends 13 seconds 'thinking' and still can’t count to five, don’t let it count your SLO breaches
AGI has been achieved internally Comment deleted
Of course man they even have AHI Comment deleted
Yeah just a little bit more funding and we'll eventually release it Comment deleted
Of course of course, we are just a little non-profit company Comment deleted
imagine if closedai turns out to be another fucking Theranos in the end Comment deleted
For those who don't understand hand has 6 fingers. Comment deleted
Actually some people have 6 fingers, so the chatgpt coud assume not paying attention to this difference from more usual hand Just joking, chatgpt is dumb of course Comment deleted
assuming all "normal" people have 5 fingers would be offensive Comment deleted
Exactly. Time to remember to be finger-count tolerant 😇 Comment deleted
Let's not forget about finger-challenged people Comment deleted
I know a few normal people who have four and a half fingers on one hand. Comment deleted
We all know who these "normal four finger people" are Comment deleted
Actually I thought of woodworkers but that's a valid alternative Comment deleted
Большое спасибо за ваше объяснение! Оно действительно помогло прояснить ситуацию. Если у вас есть еще вопросы или вам нужна помощь, не стесняйтесь обращаться! Comment deleted
but for humans it isn't obvious either at the first glance Comment deleted
also fun fact: without connecting visual information to symbolic reasoning, 5 = 6 (any difference <25% is indistinguishable) Comment deleted
btw, could someone test this on that LLM? like show pictures with quantities of something being different by one and ask whether it sees the difference and if so, which picture have more objects Comment deleted
but like, it is? It's only not if you're tired, haven't had any sleep for a couple of days and are way deep into your thoughts Comment deleted
not in this chat of programmers it isn't Comment deleted
It is for 6-fingered humans Comment deleted
as i said, without more explicit counting 5 = 6, maybe even 4 = 5 Comment deleted
If you dont have a middle finger Comment deleted
"Life at a glance" - 2015, Microsoft Windows 10 Comment deleted
Is this a glance screen reference? (Aka windows phone always on screen) Comment deleted
Yea I honestly did not see that right away. Comment deleted
for 13 seconds? Comment deleted
Na just 12. Comment deleted
good luck OCRing that :) Comment deleted
oh come on Comment deleted
cursive as in cursed 😄 Comment deleted
well 'at first glance' i didnt notce 6 fingers myself tbh Comment deleted
Even I didn't notice at first Comment deleted
Well at first glance, Error Code 429 - You exceeded your current quota, please check your plan and billing details Comment deleted
Me when I can perform a basic task that the generative text cannot Comment deleted
o1 isn't natively multimodal. It still gets image description from some image-to-text service Comment deleted
meh, for kinda the best model they could've make it truly multimodal Comment deleted
Except that's clearly a left hand, so you're safe from AInigo Comment deleted
hey, look, cursive in general is a lost art now Comment deleted