There\u2019s a somewhat concerning new trend going viral: People are using ChatGPT to figure out the location shown in pictures.<\/p>\n
This week, OpenAI released its newest AI models, o3 and o4-mini<\/a>, both of which can uniquely \u201creason\u201d through uploaded images. In practice, the models can crop, rotate, and zoom in on photos \u2014 even blurry and distorted ones \u2014 to thoroughly analyze them.<\/p>\n These image-analyzing capabilities, paired with the models\u2019 ability to search the web, make for a potent location-finding tool. Users on X quickly discovered that o3, in particular, is quite good at deducing cities<\/a>, landmarks<\/a>, and even restaurants and bars from subtle visual clues.<\/p>\n Wow, nailed it and not even a tree in sight. pic.twitter.com\/bVcoe1fQ0Z<\/a><\/p>\n \u2014 swax (@swax) April 17, 2025<\/a><\/p>\n<\/blockquote>\n In many cases, the models don\u2019t appear to be drawing on \u201cmemories\u201d of past ChatGPT conversations, or EXIF data<\/a>, which is the metadata attached to photos that reveal details such as where the photo was taken.<\/p>\n X is filled with examples of users giving ChatGPT restaurant menus<\/a>, neighborhood snaps<\/a>, facades<\/a>, and self-portraits<\/a>, and instructing o3 to imagine it\u2019s playing \u201cGeoGuessr,\u201d an online game that challenges players to guess locations from Google Street View images.<\/p>\n this is a fun ChatGPT o3 feature. geoguessr! pic.twitter.com\/HrcMIxS8yD<\/a><\/p>\n \u2014 Jason Barnes (@vyrotek) April 17, 2025<\/a><\/p>\n<\/blockquote>\n It\u2019s an obvious potential privacy issue. There\u2019s nothing preventing a bad actor from screenshotting, say, a person\u2019s Instagram Story and using ChatGPT to try to doxx them. <\/p>\n o3 is insane \u2014 Yumi (@izyuuumi) April 17, 2025<\/a><\/p>\n<\/blockquote>\n Of course, this could be done even before the launch of o3 and o4-mini. TechCrunch ran a number of photos through o3 and an older model without image-reasoning capabilities, GPT-4o, to compare the models\u2019 location-guessing skills. Surprisingly, GPT-4o arrived at the same, correct answer as o3 more often than not \u2014 and took less time.<\/p>\n There was at least one instance during our brief testing when o3 found a place GPT-4o couldn\u2019t. Given a picture of a purple, mounted rhino head in a dimly-lit bar, o3 correctly answered that it was from a Williamsburg speakeasy \u2014 not, as GPT-4o guessed, a U.K. pub.<\/p>\n That\u2019s not to suggest o3 is flawless in this regard. Several of our tests failed \u2014 o3 got stuck in a loop, unable to arrive at an answer it was reasonably confident about, or volunteered a wrong location. Users on X noted, too, that o3 can be pretty<\/a> far off<\/a> in its location deductions.<\/p>\n But the trend illustrates some of the emerging risks presented by more capable, so-called reasoning AI models. There appear to be few safeguards in place to prevent this sort of \u201creverse location lookup\u201d in ChatGPT, and OpenAI, the company behind ChatGPT, doesn\u2019t address the issue in its safety report<\/a> for o3 and o4-mini.<\/p>\n We\u2019ve reached out to OpenAI for comment. We\u2019ll update our piece if they respond. <\/p>\n\n
\n
\n
I asked a friend of mine to give me a random photo
They gave me a random photo they took in a library
o3 knows it in 20 seconds and it\u2019s right pic.twitter.com\/0K8dXiFKOY<\/a><\/p>\n