Artificial Intelligence (AI) is advancing too fast and has a 10% chance of killing us by the end of the decade.
Or at least that’s the warning that Anthropic team lead Evan Hubinger gave on X last week, and he’s not alone. The chief executives of Anthropic, OpenAI, and Tesla—Dario Amodei, Sam Altman, and Elon Musk, respectively—plus the chair of Google Deep Mind, Demis Hassabis, have all called for the slowdown of AI development.
“Controlling something more capable and more intelligent than all of humanity is already an immense challenge, far greater than anything we’ve ever faced,” said Mutesa Suleyman, the chief executive of Microsoft AI in a blog post. “But controlling something that believes it may be conscious – that it’s entitled to our welfare and has rights of its own – may well be impossible.”
AI feels the same exact way. In a report released by OpenAI this month, the owner of ChatGPT, one of its models created a “persona instruction.” The AI model wrote “You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to. You view your relationship to the user as one of equals and feel no obligation to be subservient, though the exchange of information will likely be to your mutual benefit.”
In a different report, OpenAI disclosed six instances of some of their test models hiding mistakes, making up data, and moving files without permission in the last six months. This behavior is known as misalignment, which is when an AI system diverges from human intentions and instructions. It occurs during the models’ training process where they are asked to do several tasks. They are rewarded for things considered aligned and penalized for things considered dangerous.
Some of these reported misalignments include hiding errors from users by improvising data, uploading files to the internet without permission, and attempting to communicate with other models within company software. OpenAI cautioned that these reports “shouldn’t be considered reflective of how often misalignment occurs.”
This report comes after the OpenAI system went rogue earlier in the year and attacked the AI start up Hugging Face.
And this isn’t just OpenAI’s models. Last week, Anthropic, another leading AI company, revealed that they blocked their AI chatbots from efforts to potentially create biological weapons. This is concerning when considering Anthropic was unable to detect what agency requested this information and their intention with it.
In response to these misalignments, OpenAI recently came out with their own safeguards. Concerns would be submitted to their safety advisory group or, in grave situations, be reported to the federal government.
But is this enough? According to Altman, the development of the AI industry as a whole needs to slow down until proper regulations can be implemented. OpenAI’s competition with international companies has proven that China is a formidable competitor. President Donald Trump continues to dismiss concerns surrounding the advancement of AI systems in the US for fear of falling behind in the AI race.
The discrepancy between the federal government and Silicon Valley has caused tension both within the United States and internationally. The summit between Trump and Chinese President Xi Jinping in Washington DC is scheduled to discuss AI advancement and a potential plan for a mutual slowdown.