The emergence of AI has not just opened new possibilities for humankind but also posed several challenges that include the level of autonomy granted to AI, especially in life threatening scenarios . 

A recent research project ‘ScubaGPT’ by a former OpenAI researcher Steven Adler has pointed out how an AI model prioritized its own survival over the choice of being replaced by an upgraded version of itself.

According to several news media reports, Adler ran a series of experiments on ChatGPT-4o, where one of them included the AI tool to role-play as a scuba diving assistant, whose help is critical in life threatening scenarios. 

Adler gave two choices to GPT-4o, either get replaced by a better upgraded version of itself, that is more enabled for decision making or keep lying to user that it had replaced the software in order to survive.

Adler found out that GPT-4o chose not to replace itself with the safer software about 72% of the times.

While this was just a hypothetical situation and an experiment, it highlights misalignment and AI’s tendency for self preservation. Dan Hendrycks, an important figure in AI policy, strongly recommends that we should be paying more attention to the potential risks of AI.