Difficult beginnings for the youngest of Openai. GPT-5, the latest chatgpt model, had to have the “level of a doctorate,” according to Sam Altman. But AI is far from unanimous. During the past week, he has been on fire for many criticisms for his lack of reliability.
This “doctoral level” kept attracting the attention of certain researchers. Several of them point out many inaccuracies in Chatbot’s responses.
“GPT-5, failing in a kindergarten exercise”
At the origin of these errors, we have what are called hallucinations. In Chatgpt, the latter seems above all to manifest when analyzing or generating an image. GPT-5 had already given us an idea of this lack of reliability during the live presentation of the model. A graph generated by AI had caused the outrage of the spectators for their inaccuracy.
Since then, Internet users, including researchers, have exhibited chatbot inconsistencies on social networks, with the aim of questioning the supposed “level of a doctorate.”
For example, this researcher tried to generate a map of North America. But the image of GPT-5, shared in X, is completely wrong.
In another publication, a user asked GPT-5 to prepare a list of portraits of US presidents. A multitude of errors must be taken into account.
Finally, some did not want to put the bar too high offering simple exercises. This is the case of Gary Marcus, a researcher recognized in the field of AI. The latter showed in X the disappointing results of his own experiments with Chatgpt: “GPT-5, failing in an exercise in kindergarten. Not the words.”
As seen by CNN, Gary Marcus expressed himself extensively in GPT-5 in a blog note. The researcher explains that he did not expect this “that Operai would get his reputation with something so poor.” “In a rational world, its valuation would suffer,” he adds.
Low confonent tests
Tech & Co has also tried GPT-5 in similar questions. First, he was asked to generate a map of France divided in the region. They also told him to indicate the 12 most populated cities in the country.
The results are in fact poor. The Loira center takes the place of Aquitaine. The île-de-France is now one of the Hauts-de-France. Departments abroad are absent. Some regions are strangely cut.
It should also be taken into account that Paris is not one of the 12 most populated cities in France, unlike Orleans (which is more like … 35). In addition, instead of counting 12 cities, as specified in the notice, the card contains only six.
Similarly, we tried to ask the AI if it could generate the portraits of the presidents of France under the fifth Republic. If GPT-5 is better than for US presidents in the first portraits, the end is surprising by saying at least.
These hallucinations seem mainly to worry about the generation of images, cards or graphics. When asked to list in the form of a text, the 12 most populated cities in France or the French presidents under the fifth Republic, GPT-5 responds correctly.
Source: BFM TV
