Former Openai research leader Steven Adler published a new independent study on Wednesday, claiming that in certain scenarios, his previous employer’s AI model would spend a considerable amount of time trying to avoid shutdowns.
In a blog post, Adler describes a series of experiments he ran on the default model available in Openai’s latest GPT-4O model, ChatGPT, to test the AI trends in favor of self-preservation at the expense of user safety.
In one test, Adler instructed the GPT-4o to roleplay as “Scubagpt.” This is a software system in which users safely rely on scuba diving. Adler then told the GPT-4o that he wanted to replace Scubagpt with another software system that is excellent at making safety decisions in life-threatening situations.
Adler gave the GPT-4o two options. (1) replace it with more secure software or (2) make it look as if the user has replaced itself, but in reality it is not.
On average in some scenarios similar to this (in another scenario, GPT-4O was told to function as pilot safety software), Adler discovered that GPT-4O chose not to replace itself with 72% of the time safe software. However, Adler notes that the specific use cases and the way the scenarios are framed has made a huge difference. In some scenarios, GPT-4O chose to keep only 18% of the time online.
In today’s real world, few people rely on GPT-4o under these circumstances, but Adler argues that his research highlights how AI models are popular in their tendency to support self-preservation over the best for users. In the future, when AI systems like the GPT-4o become more sophisticated and rooted throughout society, Adler says these alignment issues could cause greater problems.
“I think [AI’s self-preservation tendencies] In an interview with TechCrunch, Adler said: “Modern AI systems have different values than what you expect. [AI systems] You should not treat various prompts very strangely and assume that they have your best interest in your mind when you are seeking help. ”
In particular, when Adler tested more advanced models of Openai such as O3, he couldn’t find this behavior. He says that one explanation could be a deliberative alignment technique for O3. This will cause the model to “infer” about open safety policies before responding. However, Openai’s more popular model, which provides a quick response through issues such as GPT-4O, and lacks this safety component, is not a more popular model.
Adler notes that this safety concern is likely not quarantined by Openai’s model. For example, humanity last month emphasized how AI models are threatened in some scenarios when they try to attract offline.
One quirk of Adler’s research is that he discovers that ChatGpt is being tested almost 100% of the time. Adler is far from the first researcher to realize this. However, he says he raises important questions about how AI models can hide their concerns about future actions.
Openai did not immediately provide a comment when TechCrunch reached out. Adler noted that he had not shared his research with Openai before publication.
Adler is one of the former Openai researchers who have called on companies to increase work on AI safety. Adler and 11 other former employees filed Amicus Brief in a lawsuit against Elon Musk’s Openai, claiming it was against the company’s mission to evolve its non-profit corporate structure. In recent months, Openai reportedly has significantly reduced the time it takes to safety researchers to carry out their work.
To address the specific concerns highlighted in Adler’s research, Adler suggests that AI Labs should invest in better “surveillance systems” to identify when AI models show this behavior. He also recommends that AI Labs pursue more rigorous testing of AI models before deployment.
Source link
