There is something about the new trend viral: People are using the ChatzP to find the position shown in the picture.
This week, Open has published its new AI models, and 3 and 4-minute, both can “cause” uniquely through uploaded images. In practice, in photos to fully analyze the models – even blurred and distorted people can crop, rotate and zoom.
This image-analysis ability, web searching models efficiently, create a strong position tool to find the tool. X’s users quickly discovered that and 3, especially, is pretty good in terms of discount City, LandmarksAnd even from fine visual clue to restaurants and bars.
Wow, nail it and not even a tree. pic.twitter.com/bvcoe1fq0z
– Swux (@S SUX) April 17, 2025
In many cases, models do not seem to draw “memories” of the past chatzipi conversation or EXIFT DATA – The metadata connected to the photos reveals the details of where the image was taken.
X -chatzipt is full with examples of users to give Restaurant menu, Snaps, Front And Self-styleAnd the 3K is an online game instructed to imagine what is playing “Geography” that challenges players to estimate the locations of Google Street View images.
This is a fun chatzipt and 3 features. Geoguers! pic.twitter.com/hrcmixs8YD
– Jason Burnes (@Vairotech) April 17, 2025
This is an obvious potential privacy problem. There is nothing to prevent a bad actor from screensting a person’s Instagram Story and use ChatzPT to try to doxx them.
O 3 insane
I told one of my friends to give me random pictures
They gave me a random picture in a library
O 3 knows it in 20 seconds and it is exactly pic.twitter.com/0k8dxifkoy– Um (@IGUUUM) April 17, 2025
Of course, it can be done before the launch of O3 and O4-Mini. To compare the position of the TechCrunch models to compare the efficiency of the proponance and through an old model without image-reissing power and run an old model. Surprisingly, the GPT -4O came at the same time, and the correct answer as 3 was often not much -and it took less time.
There was at least one instance during our short exam while 3 no place was found in GPT -4O. An image of a purple color is given, the head mounted in a faded bar, and 3 correctly answered that it came from a Williamsberg’s Specific-such as GPT-4O, the UK.
This is not the same as proposing to be 3 innocent. Several tests failed – and 3 was stuck in a loop, which was unable to reach an answer it was reasonably confident or did any wrong position volunteer. The users of the X also mentioned, and may be 3 Beautiful Far away In case of discounting its position.
The trend, however, depicts some of the emerging risk depicts presented by the more capable, so -called reasonable AI models. There are a few security measures to prevent this type of “reverse location” in ChatzPT and the company behind ChatzPT does not address its problem. Security report For O3 and O4-Mini.
We reached the opening to comment. We will update our piece if they respond.