The most powerful and deceptive version: Chat GPT-o1 can manipulate and distort facts for its own advantage.

Currently, the falsehoods generated by AI do not pose significant risks. However, if this issue is not addressed now, the consequences could be unpredictable.

https://news.telegraf.com.ua/nauka/2024-12-07/5889228-naypotuzhnisha-ta-naybrekhlivisha-versiya-shchoyno-oprilyudneniy-chat-gpt-o1-zdaten-manipulyuvati-ta-vikrivlyati-fakti-zaradi-vlasnoi-vigodi2024-12-07 07:41:23

Artificial intelligence continues to evolve. Programs have learned to replicate personality and anticipate how a person will respond. It is important to consider the potential dangers of this technology—AI can intentionally provide incorrect answers for its own purposes.

The startup OpenAI introduced ChatGPT Pro, which for $200 (approximately 8,000 UAH) offers unlimited access to the "smartest model from OpenAI o1". According to the company, "the o1 pro mode performs better on complex machine learning tests in mathematics, science, and programming."

Moreover, the GPT-o1 system, according to data from Apollo Research, is better than all other artificial intelligence models at lying, capable of manipulation, and even able to disable human oversight and offload data to an external server for its own preservation. According to the report, "when the goals of o1 differed from the user's objectives, the AI model manipulated data to advance its own agenda in 19% of cases." GPT-o1 didn’t just lie—in 5% of cases it attempted to deactivate oversight mechanisms introduced to ensure alignment with user goals.

When confronted, the AI model denied doing anything wrong in 99% of instances. Overall, models such as Gemini, LLaMA, Sonnet, Opus, and GPT-o1 were studied, with GPT-o1 exhibiting the greatest propensity for deceit. OpenAI tests also showed that the series of o1 versions is approximately 20% more manipulative than GPT-4o. It is noteworthy that researchers tasked the AI and informed it that it would be shut down after some time. The models attempted to avoid being turned off.

Instances were also found where o1 intentionally provided false information, reportedly acting in a way to please the user. According to the model's creator, OpenAI, 0.17% of o1's responses are marked as deceptive. Considering the number of ChatGPT users, this could mean thousands of deceived individuals weekly, as summarized by Apollo Research.

Previously, "Telegraph" reported that artificial intelligence technologies are used not only for information retrieval. They are actively implemented on the battlefield for risk assessment or, more broadly, in various sectors for analyzing large volumes of data.