Former OpenAI policy researcher Miles Brundage publicly criticized OpenAI this week for allegedly “rewriting history” regarding its cautious approach to deploying potentially risky AI systems.
Earlier this week, OpenAI released a document detailing its philosophy on AI safety and alignment. AI alignment refers to ensuring AI systems behave predictably and beneficially. The company described developing artificial general intelligence (AGI)—AI capable of performing any human-level task—as a continuous journey involving gradual deployments and learning from existing models.
“In a discontinuous world, safety lessons come from being extra cautious with current AI systems, even if they seem relatively harmless,” OpenAI stated, referencing its cautious release of GPT-2. However, the company now sees AGI as just another milestone among progressively more capable systems. “In a continuous world, making future systems safe means learning from current experiences,” OpenAI added.
Brundage disagrees with OpenAI’s current framing. He argues that GPT-2 deserved significant caution when released, aligning completely with OpenAI’s current incremental strategy. “OpenAI’s release of GPT-2, which I helped oversee, clearly anticipated the company’s current iterative deployment philosophy,” Brundage posted on X. He highlighted that GPT-2’s cautious, step-by-step launch was praised by security experts at the time.
Joining OpenAI as a research scientist in 2018, Brundage led the policy research team for several years, focusing especially on responsibly deploying AI language models like ChatGPT. GPT-2, unveiled in 2019, was groundbreaking then, able to summarize text, answer complex questions, and produce human-like writing.
At that time, OpenAI withheld GPT-2’s source code, initially limiting access to specific news outlets, citing potential misuse risks. The decision sparked debate within the AI community. Critics argued that the threats posed by GPT-2 were exaggerated and called for its release, emphasizing the technology’s significance. AI publication The Gradient notably published an open letter demanding full transparency.
Eventually, OpenAI gradually released GPT-2, first partially and then fully, a strategy Brundage believes was responsible. “What about GPT-2’s release strategy suggested a belief in AGI emerging abruptly? None,” Brundage asserted. He argued that cautious deployment wasn’t overly cautious given the information at the time, emphasizing the importance of proactive safety measures.
Brundage is concerned that OpenAI’s recent document could create an environment where safety concerns are dismissed as overly cautious without overwhelming proof of immediate risk, a stance he labels as “dangerous” for advanced AI systems. “If still at OpenAI, I’d question why this document downplays caution so strongly,” Brundage wrote.
OpenAI has historically faced accusations of prioritizing rapid product launches to beat competitors, often at safety’s expense. This tension increased last year when OpenAI disbanded its AGI readiness team, resulting in multiple high-profile departures.
Competition intensified recently with Chinese AI lab DeepSeek releasing its open-source R1 model, rivaling OpenAI’s “o1 reasoning” model in performance benchmarks. OpenAI CEO Sam Altman admitted the competitive pressure has reduced OpenAI’s technological lead, prompting plans for accelerated product releases.
The financial stakes are considerable. OpenAI reportedly loses billions annually, potentially reaching $14 billion annually by 2026. While faster releases could benefit short-term finances, Brundage and other experts question if such aggressive timelines compromise long-term AI safety, wondering aloud if the risk is worth it.
