Stay updated with our daily and weekly newsletters for the latest in AI industry coverage and exclusive content. OpenAI, the creators of ChatGPT, has introduced its “Preparedness Framework” today, designed to monitor and manage the risks associated with increasingly powerful AI models.
This announcement comes at a challenging time for OpenAI, which has recently been criticized for its handling of the dismissal and subsequent rehiring of its CEO, Sam Altman. The incident has sparked discussions about the lab’s governance and responsibility, especially since it is at the forefront of advanced AI development.
OpenAI’s new Preparedness Framework aims to address these concerns by emphasizing a responsible and ethical approach to AI development. The framework includes processes for tracking, evaluating, and protecting against significant risks posed by advanced AI models, such as their potential use in cyberattacks, mass influence, or autonomous weapons.
A significant part of this framework is the introduction of risk “scorecards” for AI models. These scorecards track various indicators of potential harm, including the model’s capabilities and vulnerabilities. They are updated regularly and prompt reviews and interventions when certain risk levels are reached.
The framework prioritizes rigorous, data-driven assessments over speculative scenarios that often dominate public discussions about AI. OpenAI is investing in these assessments and in developing strategies and safeguards to mitigate risks.
Importantly, the framework is not static; it will continuously evolve based on new data, feedback, and research. OpenAI plans to share the findings and best practices derived from this framework with the wider AI community.
This initiative comes after similar moves from OpenAI’s competitor, Anthropic, which has released its Responsible Scaling Policy. Anthropic’s approach is more formal and prescriptive, setting specific AI safety levels and protocols, and pausing development if safety cannot be assured. In contrast, OpenAI’s framework is more flexible, setting general risk thresholds that trigger reviews when necessary.
Experts see strengths and weaknesses in both approaches. Anthropic’s method might be better at enforcing safety standards, as it integrates safety into the development process. OpenAI’s approach, however, allows for more flexibility and human judgment, which can sometimes lead to variability and potential errors.
OpenAI has been under scrutiny for its rapid deployment of models like GPT-4, which has drawn criticism and concern over safety. Anthropic’s proactive policy might give it an advantage in this aspect.
Despite these differences, both frameworks mark a significant advancement in AI safety. As AI technology continues to grow, collaboration and coordination on safety measures among leading labs and stakeholders are crucial to ensuring that AI is developed and used ethically for the benefit of humanity.