The latest direction of ChatGPT virus is “reverse search on site” from photos


There is somewhat in relation to the new virus: people use Chatgpt to find out the location shown in the pictures.

This week, Openai released the latest models of artificial intelligence, O3 and O4-MiniBoth can “cause” a unique “cause” through the downloaded images. In practice, models can cut, rotate and enlarge images – even non -fog and dispersed – for precisely analyzing them.

This image analysis possibilities, associated with the ability of the models to search in the web, make the site generation tool. Users discovered X quickly that O3, in particular, is very good in conclusion Citiesand LandmarksEven restaurants and bars from hidden visual evidence.

In many cases, the models do not seem to depend on the “memories” of the previous ChatgPT conversations, or Exif dataIt is the descriptive data attached to the pictures that reveal details like the place where the image was taken.

X is filled with examples of users who provide Chatgpt Restaurant listsand The neighborhood settlesand FacadesAnd Self -imagesO3 guidance to imagine that she plays “GeoguessR”, which is an online game that challenges players to guess sites from Google Street View.

It is a possible and clear privacy issue. Nothing prevents a bad actor from the screenshot, for example, a person’s Instagram story and the use of Chatgpt to try to try.

Of course, this can be done even before the launch of O3 and O4-MINI. Techcrunch has played a number of photos through O3 and an older model without the possibilities of displaying photos, GPT-4O, to compare sites guessing sites for models. Surprisingly, the GPT-4O has reached the same right answer as O3 in the most time-and takes less time.

There was at least one counter during our brief test when O3 found a place that the GPT-4O could not. Looking at a picture of the purple head, installed in a faint-ranging tape, O3 answered correctly that it was from Williamsburg Speakesy-not, as GPT-4O guessed, UK pub.

This does not mean that O3 is not flawed in this regard. Many of our tests failed – O3 stumbled in a loop, unable to reach an answer, was reasonable, or volunteered a wrong position. Users also noticed that O3 could be Beautiful away In its locations.

But the trend shows some of the emerging risks offered by the most capable artificial intelligence models, so -called logic. It seems that there are few guarantees in force to prevent this type of “reverse search for the site” in Chatgpt, and the company that stands behind Chatgpt does not address the problem in it Safety Report For O3 and O4-MINI.

We have contacted Openai to comment. We will update our article if they respond.



Leave a Reply

Your email address will not be published. Required fields are marked *