A Reddit user says ChatGPT found his earbud lost in white gravel using three close-up photos

White earbud, white gravel

On Oct. 25, a Reddit user posted “ChatGPT found my lost earbud” on r/ChatGPT. According to NewsGram, the incident took place on the evening of Friday, Oct. 23. The man was leaving a park with his son and sprinted after the boy when he would not stop. His left Bose earbud fell out of his hoodie pocket into loose white gravel.

He searched for about 10 minutes as the sun was setting. He wrote that he did so “knowing there’s basically no way I’m going to find a tiny white earbud in a sea of white gravel.” According to Entertales, the earbuds would not fully activate or connect to his phone until they detected an ear. That ruled out playing a sound to find it. So he photographed the area and asked ChatGPT.

No confident answer at first

According to the screenshots, he sent the first photo with the message “Can you find my headphone? It’s a white Bose uc earbud.” ChatGPT worked for 1 minute 2 seconds. It said it had zoomed in and scanned the gravel section by section but couldn’t confidently identify the earbud. It noted a white and gray object at the base of a reddish shrub on the left but said it looked more like a plant tag.

ChatGPT then asked for a few closer photos of the gravel, ideally with the camera pointed straight down. It said the earbud’s smooth white plastic should be easier to tell apart from the rocks at closer range.

Found in the third photo

The user sent three close-ups of the area he suspected. This time ChatGPT worked for 1 minute 53 seconds and replied “Found it!” It placed the earbud in the third photo, toward the left side and a little higher than the vertical midpoint. It put the spot at roughly 10 to 12% of the way across from the left edge and 44% down from the top. It said it could clearly see the smooth white body and gray silicone piece between the rocks.

ChatGPT also sent back a copy of the photo with the earbud circled in red and labeled “EARBUD.” It added that the earbud sat to the left of the reddish bush, about halfway between the bush and the photo’s left edge. According to the user, the whole thing took less than five minutes. “I’ll be damned, it actually found it,” he wrote.

From Reddit to X

According to NewsGram, the post had more than 3,600 upvotes and over 300 comments by Oct. 27. On X, the story spread through an account named Sashi. On Oct. 25, Sashi posted four images with the caption “ChatGPT found lost earbud,” without crediting the Reddit post. It had more than 1.7 million views as of Nov. 6, with no Community Note.

The story rests on the user’s own account and screenshots and has not been independently verified. A few low-engagement replies on X called it fake, one suggesting ChatGPT “probably hallucinated that earbud,” but none offered evidence. Outlets such as The Tribune and NewsGram based their reports on the Reddit post.

Sashi@sashikantsingh_

ChatGPT found lost earbud 🤯

How the model scanned the photo

The screenshots do not show which model was used. The “Worked for” timer, the section-by-section zooming and the marked-up copy point to a reasoning mode that processes images with tools. A Reddit commenter quoted by Entertales wrote that ChatGPT “can now open an image in its cloud environment” and check different parts separately.

The approach is not new. In April 2025, developer Simon Willison described OpenAI’s o3 running Python code to crop and zoom into his photo while guessing where it was taken. He counted 25 separate cropping operations and guessed that the model’s image input has low resolution. OpenAI’s developer guide says images may be scaled down to fit a budget of 32-by-32-pixel patches. A small object in a wide shot can end up covering very few pixels, and closer photos ease that.

In a December 2023 paper, researchers Penghao Wu and Saining Xie wrote that multimodal language models struggle with details in high-resolution, crowded images. They traced this to the lack of a visual search mechanism and built V*Bench to measure it.

Known limits

OpenAI’s guide says its models struggle with tasks that require precise spatial localization, may give incorrect descriptions and can only approximate object counts. Models can also “see” objects that are not there. A 2023 study that introduced the POPE evaluation found severe object hallucination in most of the vision-language models it tested.

The method does not always work. According to NewsGram, another Reddit user searched for a screw lost outdoors and was told it had been found. That user wrote that ChatGPT had in fact modified the original picture and put a random screw on it. In the earbud case, ChatGPT stated its uncertainty on the first photo and asked for better ones. The user says he got the earbud back in less than five minutes.

Cover image generated with AI.