Topics
Latest
AI
Amazon
Image Credits:Silas Stein / picture alliance / Getty Images
Apps
Biotech & Health
clime
Cloud Computing
mercantilism
Crypto
enterprisingness
EVs
Fintech
fundraise
Gadgets
Gaming
Government & Policy
Hardware
Layoffs
Media & Entertainment
Meta
Microsoft
concealment
Robotics
Security
Social
quad
Startups
TikTok
transfer
Venture
More from TechCrunch
result
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
There ’s a jolly touch new course going viral : People are using ChatGPT to figure out the location shown in pictures .
This hebdomad , OpenAI released its new AI manakin , o3 and o4 - mini , both of which can unambiguously “ reason ” through upload images . In drill , the manakin can prune , circumvolve , and zoom in on photo — even hazy and twisted ones — to thoroughly analyze them .
These image - analyse capability , pair with the models ’ power to search the entanglement , make for a virile location - finding tool . Users on X quickly discovered that o3 , in finical , is quite dependable at deducingcities , landmarks , and even eating house and bar from insidious visual clues .
Wow , nailed it and not even a Sir Herbert Beerbohm Tree in sight.pic.twitter.com/bVcoe1fQ0Z
— swax ( @swax)April 17 , 2025
In many casing , the models do n’t appear to be draw on “ memory board ” of past ChatGPT conversation , orEXIF data point , which is the metadata attach to photos that unveil details such as where the photo was taken .
X is fulfill with example of users giving ChatGPTrestaurant menus , neighbourhood cinch , frontage , andself - portraits , and instructing o3 to imagine it ’s playing “ GeoGuessr , ” an online biz that challenges player to guess location from Google Street View epitome .
this is a fun ChatGPT o3 feature . geoguessr!pic.twitter.com / HrcMIxS8yD
— Jason Barnes ( @vyrotek)April 17 , 2025
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
It ’s an obvious likely privacy military issue . There ’s nothing foreclose a high-risk thespian from screenshotting , say , a person ’s Instagram Story and using ChatGPT to endeavor to doxx them .
o3 is insaneI ask a friend of mine to give me a random photoThey pass me a random pic they take in in a libraryo3 knows it in 20 seconds and it ’s rightpic.twitter.com/0K8dXiFKOY
— Yumi ( @izyuuumi)April 17 , 2025
Of naturally , this could be done even before the launch of o3 and o4 - miniskirt . TechCrunch ran a telephone number of picture through o3 and an older model without picture - reasoning capabilities , GPT-4o , to compare the model ’ localisation - suppose skills . Surprisingly , GPT-4o arrived at the same , right solution as o3 more often than not — and take less time .
There was at least one instance during our brief examination when o3 found a place GPT-4o could n’t . Given a photograph of a purple , get on rhino fountainhead in a dimly - illume bar , o3 correctly answered that it was from a Williamsburg speakeasy — not , as GPT-4o approximate , a U.K. pothouse .
That ’s not to paint a picture o3 is flawless in this respect . Several of our trial give way — o3 got stuck in a cringle , ineffectual to arrive at an answer it was reasonably surefooted about , or volunteered a wrong location . Users on X remark , too , that o3 can beprettyfar offin its location deduction .
But the movement illustrate some of the emerge peril presented by more capable , so - called reasoning AI models . There appear to be few safeguards in place to prevent this variety of “ reverse location search ” in ChatGPT , and OpenAI , the company behind ChatGPT , does n’t come up to the issue in itssafety reportfor o3 and o4 - mini .
We ’ve reached out to OpenAI for comment . We ’ll update our piece if they reply .
Updated 10:19 p.m. Pacific : hour after this story was published , an OpenAI spokesperson post TechCrunch the following statement :
“ OpenAI o3 and o4 - mini bring visual reasoning to ChatGPT , making it more helpful in area like availability , research , or identifying fix in pinch response . We ’ve worked to train our models to refuse requests for private or sensible information , added safeguard intended to proscribe the model from discover private individual in images , and actively monitor for and take action against abuse of our usage insurance on privacy . ”
