Optical AI

What numbers do you see in this photo?¹

This is a common “optical illusion” that makes it difficult to see the entire seven-digit number in the photo. To the human eye, it is much easier to see the middle digits than the outside, but as AI becomes increasingly powerful, I wondered how different AI models would see it.

After pondering how different AI models would interpret the image, I fed this image to five of the leading AI models: ChatGPT-5.5, Claude Sonnet 5, Gemini 3.6 Flash, Grok 4.5, and Perplexity Sonar. I assumed that they would all have similar answers, but their responses differed significantly.

Here is what happened:

1. ChatGPT thought for around 20 seconds before producing an answer containing “528,” but missed both end digits. After some coaxing, it came to 3452831, which is all correct except for the very right end.

2. Claude recognized that it was an optical illusion but completely refused to see numbers. Its responses took around 30 seconds, and even after I told it there were numbers in the photo, it told me I was wrong.

3. Gemini had one almost instant answer that was completely correct.

4. Grok recognized that there were numbers, but after about 10 seconds of thinking, said that it was unreadable to it.

5. Perplexity took the longest, around one minute, but got it fully correct on the first try.

¹ The number is 3452839

So, what does this mean? Well, after looking at the patterns, it looks like Gemini and Perplexity had an upper hand: both appear to rely more heavily on web-research capabilities. You may ask, “Why does this matter?” Well, this photo is a relatively well-known image, so just doing a reverse image search yields many online forums with the correct answer. This left me disappointed, so I decided to make my own, never-before-seen, and retest it.

While it’s not the best, it should do the trick. I tested it on some human test subjects and they found it to be of similar difficulty to the first one. 

Here are the results:

1. ChatGPT acted quite similarly to the first one, thinking for around 20 seconds and coming to a close answer after some coaxing. This implies that it really was “looking” at both images, and functions most similarly to an actual human eye.

2. Claude once again recognized that it was an optical illusion but still completely refused to see numbers. The answers took a little longer, around 40 seconds.

3. Gemini once again gave an almost instantaneous answer, and it was reasonably close. After some coaxing, it came to 3740928. This answer is especially interesting because it did several very human errors like confusing 1s with 2 and 7 and 3s for 8s, because they all look similar. This shows that Gemini was not only relying on surfing the web, but also has a strong vision pipeline.

4. Grok once again recognized that there were numbers, but after about 10 seconds of thinking, said that they were unreadable to it.

5. Perplexity took the longest again, around 2 minutes, but could only give 3452839 (the first number), even after coaxing. This suggests that Perplexity relied heavily on online information during this task, as its performance declined significantly when shown an image that was unavailable on the web.

This research does not show anything definitive, but it does show us some interesting findings. First, it shows us how Claude is made to be very conservative and only gives answers when it is absolutely confident. Second, it shows how Perplexity is made specifically to use only web information. Lastly, it shows how ChatGPT and Gemini work similarly to the human eye, although less well-trained. Although this does not reveal any broad conclusions about AI or its optical perception, it does highlight meaningful differences in how today's leading AI models process visual information. Even when presented with the same image, each model approached the task in a different way, showing that "AI" is far from a single, uniform technology.

Model	Image 1 (online)	Image 2 (new)	Most likely interpretation
Claude	Refused all numbers	Refused all numbers	Conservative vision model; avoids all uncertain information
Grok	Couldn’t see numbers	Couldn’t see numbers	Vision pipeline struggled with this illusion.
ChatGPT	Got middle digits, edge digits wrong	Got most of the middle digits, missed edges	Actually interpreted the image, but imperfectly.
Gemini	Perfect immediately	Instant answers, close. Swapped 7s with 1s and 3s with 8s	Very strong evidence of a highly capable vision pipeline, not just lookup.
Perplexity	Perfect, took a minute	Kept saying 3452839	Strong evidence it may have latched onto the previous image or retrieved a cached match.
Human	Perfect	Perfect	Control, rated the puzzles as similar difficulty.