 |
| DALL-E's interpretation of the puzzle image |
Today, a LinkedIn contact sent me a challenge that, while I'm not fond of IQ-test-like games, I enjoy using to test ChatGPT's capabilities. The test involved finding six differences in a set of images.
The Test:
The test involves two nearly identical images featuring an octopus and a chameleon, with the task being to find six subtle differences between them.
Tabea Hirzel uploaded this visual puzzle to ChatGPT, using it as an opportunity to test the AI's ability to identify these differences accurately.
ChatGPT initially identified five out of the six differences correctly but struggled with the sixth, even altering a previously correct answer upon further prompting.
Below is the conversation between ChatGPT-4o and Tabea HIRZEL, with observations on the errors made. The icons and bold highlights are added by the author, not ChatGPT.
The original image with the puzzle:
Conversation with ChatGPT-4:
ChatGPT4o: In the "Unterschiede Finden" puzzle, there are six differences between the two images of Leon visiting his friend the octopus. Here are the differences:
The small plant to the left of the chameleon is missing in the top image.
The cloud on the right side of the top image is missing a part compared to the bottom image.
The octopus's tentacle position is different in both images.
The rocks on the top left are arranged differently.
The purple flower is missing in the top image.
The sand on the left side has an extra line in the bottom image. ❌
Tabea HIRZEL: describe point 6 better, I cannot see it.
ChatGPT4o: Upon closer inspection, it seems there isn't an extra line of sand in the bottom image as previously mentioned. The differences are:
The small plant to the left of the chameleon is missing in the top image.
The cloud on the right side of the top image is missing a part compared to the bottom image.
The octopus's tentacle positions differ between the images.
The rocks on the top left are arranged differently.
The chameleon's pattern on its back differs between the two images. ❌
The small leaf on the palm tree in the top image is missing. ❌
Tabea HIRZEL: No, the last difference is that there are more air bubbles coming out of the camaleon in the top photo.
ChatGPT4o: Thank you for the correction. Here are the six differences:
The small plant to the left of the chameleon is missing in the top image.
The cloud on the right side of the top image is missing a part compared to the bottom image.
The octopus's tentacle positions differ between the images.
The rocks on the top left are arranged differently.
The chameleon's pattern on its back differs between the two images. ❌
There are more air bubbles coming out of the chameleon in the top image. ✔️
Why Does This Happen?
This issue may stem from the convolutional neural networks (CNNs) used in image recognition within ChatGPT's underlying architecture. CNNs are powerful for detecting patterns but can struggle with nuanced, less prominent details. This limitation can cause errors, especially in distinguishing subtle differences in complex images.
Overall, while AI like ChatGPT excels in many areas, challenges like this highlight ongoing limitations in visual processing tasks.
For those interested in exploring the technologies behind this test, such as Convolutional Neural Networks (CNN) and Natural Language Processing (NLP), you can learn more through IBM's free learning path "Artificial Intelligence Fundamentals." Check out the course on Natural Language Processing and Computer Vision here: Artificial Intelligence Fundamentals.
