Recently one of the leading AI Research Lab, OpenAI, has Signaled they could cut back on their safety regulations if another company were to publish a powerful and unrestricted AI system. This company’s stance was explained in an earlier document they published on their so-called Preparedness Framework along with a blog post few days ago on Tuesday.
A Competitive Catch in Safety Procedures
The Preparedness Framework report sheds light on OpenAI’s innovative new AI safety approaches and how the company tracks, measures, estimates, and defends what they deem cataclysmic risks. In this regard, the report contains contradictory information which focuses on the competitive market.
“Adjusting our requirements may be possible if another frontier AI developer issues a high-risk system without comparable protective measures,” is what OpenAI said in their blog post on Tuesday. The company elaborated on the specific conditions under which such an adjustment would be considered, emphasizing a cautious approach.
“However, we would first confirm with great rigor that the risk landscape has indeed shifted, publicly claimed that we are making an adjustment, put the adjustment in place in such a way that it does not meaningfully increase overall risk of harm, and still retain protective measures at a more stringent level.” This suggests adaptive measures are being considered by the company, but confirms reliance on competitor innovations to drastically change the paradigm in AI development in order to broaden the scope of care.
Framework for Assessing AI Risks
OpenAI’s new framework revolves a step-by-step structure to mitigate the possible dangers posed by advanced AI models prior to publicly releasing them. The company aims to manage risks deemed plausible, measurable, novel, catastrophic, and irrevocable. Once determined, appropriate protective measures are crafted and the risks are subsequently designated as low, medium, high, or critical. The system aims to mitigate the release of models deemed ex-ante capable of causing catastrophic damage.
The firm follows an observatory approach when managing risk with regard to the operational features of its models. This includes possible exploitation or unintended virtualization in biology and chemistry, as well as in cybersecurity. OpenAI also watches over the self-improving capabilities of some of its AI models which pose unpredictable growth innovations.
These risks were in addition to those already known. Also confirmed by OpenAI are the so-called emergent risks. AI models operated for long periods of time without human supervision and the hypothetical ability to self-replicate, as well as the particular dangers that AI can inflict in areas of nuclear and radiological safety, are among emerging risks.
Importantly, the structure makes an exception for what is termed by OpenAI as “persuasion risks.” Concerns about the use of such models as ChatGPT for campaign or lobby work are not under the Preparedness Framework. Rather, these issues are managed by the so-called “Model Spec,” a document that outlines the expected conduct, behavior, and operational regulations for models like ChatGPT.
Critique from a Discontented Insider
Adler’s criticism comes to the forefront after examining the new policy framework. Adler, a former researcher at OdinAI, sharply critiqued these updates on the social media platform X. His argument suggested updates to the policies emerged from underlying safety considerations and were already shifting well before announcing changes at the start of this year.
Ex-AI Under Adler drew attention to Open AI’s master Reset Policy from its December 2023 Commitment. “Adler emphasized a specific change in OpenAI terms of their safety commitments,” he argued. Open AI committed to testing all “finetuned versions” on its AI models for breaches in safety mechanisms. “This policy shift indicates Open AI’s narrow primary focus on models tested during release phases,” he claims, “weights are intended for release.” AI model fine-tuning literally pertains to modifying a behavior captured by pre-training. Base models undergo a static testing paradigm changing ‘model fixed’ where behavior rather than testing model initiation alters risk where behavior rather than model alters risk cannot be captured in base testing of model initiation.
OpenAI’s conversational cautions regard competition’s effects on the speed and safety region of AI development suggest an intricate balance between rapidly evolving AI technologies, the need for comprehensive safety measures, and the challenge of keeping pace in the technological race. The company still honors its commitment but concedes to potential shifts depending on the behavior of other actors, which continues to fuel debate on what is considered responsible AI implementation.
