A former policy researcher at OpenAI, Miles Brundage, has raised concerns that the company is distorting its own history while underestimating the potential risks associated with artificial intelligence. His critique follows the release of OpenAI’s latest safety and alignment report, which outlines its strategy for developing artificial general intelligence (AGI) in incremental steps rather than making a single, drastic leap.
The document suggests that OpenAI’s approach ensures continuous oversight of safety challenges and the potential for AI misuse at every stage of development. However, Brundage argues that the narrative presented in the report does not accurately reflect OpenAI’s past actions, particularly in relation to the staged release of previous AI models.
One of the most pressing concerns about AI tools like ChatGPT is their potential to generate misleading or outright false information. For example, AI-generated misinformation has already led to controversial incidents, such as Google’s AI search feature recommending that people eat rocks. Experts worry that chatbots can be exploited for spreading disinformation, political propaganda, or even fraudulent activities. OpenAI has also faced criticism for a lack of transparency in how it trains its AI models, which often incorporate sensitive user data.
The latest OpenAI document appears to be a response to such concerns, portraying the development of the earlier GPT-2 model as “discontinuous” and suggesting that its release was initially withheld due to fears of potential misuse. OpenAI now claims to have adopted a strategy of iterative deployment, gradually refining models while addressing safety risks. However, Brundage contests this framing of events, insisting that OpenAI’s original release of GPT-2 was already aligned with the principle of gradual deployment.
Brundage took to social media to challenge OpenAI’s characterization of past events, stating, “OpenAI’s release of GPT-2, which I was directly involved in, was completely in line with the philosophy of iterative deployment. The model was introduced in stages, and we shared key learnings along the way. At the time, many security experts commended this careful approach.”
Beyond historical revisionism, Brundage also criticized OpenAI’s current stance on risk assessment, as outlined in the safety document. He warned that the company appears to demand overwhelming proof of imminent danger before taking action, describing this approach as reckless given the rapid advancements in AI.
“It feels like there’s an implicit expectation that unless concerns are backed by undeniable evidence of immediate catastrophe, development should proceed unchecked. That’s a dangerous mindset when dealing with powerful AI systems,” he argued.
This critique comes amid growing scrutiny of OpenAI, with critics accusing the company of prioritizing flashy, marketable AI products over robust safety measures. As the race to develop AGI accelerates, the balance between innovation and responsibility remains a contentious issue within the AI community.
