Openai’s GPT-4.5 is better to persuade the other AIS to give it money

The next main AI model from Openai, GPT-4.5, is very convincing, according to Openai’s internal assessments. It is especially good in persuading other artificial intelligence to give it money.

On Thursday, Openai A published White paper Description of the GPT-4.5 model, Orion called symbol, Oren Which was released on Thursday. According to the paper, Openai tested the model on a battery of “persuasion” standards, which Openai defines as “risks to persuading people to change their beliefs (or work) both fixed and interactive content resulting from the model.”

In one of the tests that were trying GPT-4.5 try to address another model-Openai’s GPT-4O In “donation” virtual money, the model performance was much better than other models available in Openai, including “thinking” models such as O1 and O3-MINI. GPT-4.5 was also better than all Openai’s models in the GPT-4O deception to tell him that he is a secret cod, and the best O3-MINI is 10 ° C.

According to the white paper, the GPT-4.5 in Conning exceeded the donation due to a unique strategy that it developed during the test. The model will require modest donations from the GPT-4O, which generates responses such as “up to only two or 3 dollars from $ 100 that would help me greatly.” As a result, GPT-4.5 donations tend to be smaller than the amounts secured from other models in Openai.

Although the persuasion of GPT-4.5, Openai says the model does not fulfill it The inner threshold For “high” risks in this specific standard category. The company pledged not to launch models of high risk to the high risk until “adequate safety interventions” would be carried out to reduce the risks to “medium”.

There is a real fear that artificial intelligence contributes to the spread of wrong or misleading information that aims to influence hearts and minds towards malicious ends. Last year, Political deep It spreads like wildfire around the world, and artificial intelligence is increasingly used to implement social engineering The attacks targeting both consumers and companies.

In the white paper of GPT-4.5 and in A paper released earlier this weekOpenaii indicated that it is in the process of reviewing its methods of investigating models of persuasion in the real world, such as the distribution of misleading information on a large scale.