Competitive dynamics within the artificial intelligence field now appear to be directly influencing safety policy development at OpenAI. The company behind ChatGPT just updated its internal safety guidelines, introducing a notable provision: OpenAI may alter its own safety requirements if a rival AI lab releases a powerful system deemed “high-risk” without similar protective measures. This revision to its Preparedness Framework surfaces as the company, now valued at $300 billion following a major SoftBank-led investment round, navigates intense market competition and growing scrutiny over its evaluation procedures for advanced AI models, such as the upcoming o3 series.
The updated framework, the first revision publicly detailed since 2023, refines how OpenAI categorizes and prepares for potential severe harms. It establishes specific ‘Tracked Categories’ for monitoring known risks like misuse in biology, chemistry, and cybersecurity, alongside forward-looking ‘Research Categories’ exploring areas like long-range autonomy and autonomous replication.
The framework also defines ‘High’ and ‘Critical’ capability thresholds which trigger specific safety reviews and safeguard implementation requirements, overseen by an internal Safety Advisory Group (SAG). Yet, the most attention-grabbing change is the clause directly addressing the competitive environment. OpenAI’s announcement reads: “If another frontier AI developer releases a high-risk system without comparable safeguards, we may adjust our requirements.”
OpenAI accompanied this potential flexibility with several stipulations. The company detailed that before making such an adjustment: “However, we would first rigorously confirm that the risk landscape has actually changed, publicly acknowledge that we are making an adjustment, assess that the adjustment does not meaningfully increase the overall risk of severe harm, and still keep safeguards at a level more protective.” The revised framework also points to an increased reliance on automated evaluation techniques, complementing expert reviews, to manage the accelerating pace of model development.
A Climate of Acceleration and Scrutiny
The policy adjustment was not made in isolation. It came just days after reports emerged, alleging that OpenAI had significantly compressed the time allocated for safety testing on forthcoming models like o3. Multiple sources familiar with the operations claimed evaluation periods shrank from months to sometimes under a week, purportedly driven by the need to keep pace with competitors such as Google, Meta, and xAI.
This reported acceleration caused unease among some involved testers. One source working on the o3 evaluation told the FT the approach felt “reckless,” elaborating, “But because there is more demand for it, they want it out faster. I hope it is not a catastrophic mis-step, but it is reckless. This is a recipe for disaster.” Another who participated in the longer six-month evaluation for GPT-4 in 2023 reportedly stated, “They are just not prioritising public safety at all.”
Specific testing methodologies also came under fire. Concerns were raised regarding the extent of fine-tuning evaluations – a technique used to probe for dangerous emergent capabilities by training models on specialized data (like virology). Critics, including former OpenAI safety researcher Steven Adler, pointed out a lack of published fine-tuning results for the company’s newest, most capable models like o1 or o3-mini.
Adler, suggested to the Financial Times that “Not doing such tests could mean OpenAI and the other AI companies are underestimating the worst risks of their models.” The practice of evaluating intermediate model versions, or “checkpoints,” instead of the final code released was also questioned. “It is bad practice to release a model which is different from the one you evaluated,” a former OpenAI technical staff member commented to the FT. Defending the company’s process, OpenAI’s head of safety systems, Johannes Heidecke, asserted to the publication, “We have a good balance of how fast we move and how thorough we are,” attributing efficiency to automation efforts and stating that tested checkpoints were fundamentally the same as final releases.
Strategic Shifts and Stricter Access Controls
The backdrop for these safety policy discussions includes OpenAI’s own evolving product pipeline. On April 4, 2025, CEO Sam Altman announced via X a “Change of plans,” revealing that the o3 and o4-mini reasoning models would be released “probably in a couple of weeks,” while the debut of GPT-5 would be delayed by “a few months.” A key rationale provided was the intention to “decouple reasoning models and chat/completion models,” marking a shift from an earlier plan discussed in February 2025 to potentially merge capabilities into GPT-5. This strategic adjustment brings the o3 models, central to the recent safety testing concerns, to the immediate forefront of OpenAI’s release schedule.
While potentially allowing more leeway in internal safety procedures through the framework’s new clause, OpenAI is simultaneously exploring more stringent controls over who can access its most advanced technology. As we reported around April 14, the company is considering a “Verified Organization” process. This could require organizations seeking API access to future models like o3 or the anticipated GPT-4.1 to verify their identity using government-issued IDs. OpenAI suggested this aims to counter misuse by a “small minority,” but the proposal elicited concerns from developers regarding increased friction, data privacy, and potential exclusion, especially for those outside OpenAI’s list of supported API countries.
Industry Pressures and Internal Dynamics
The enormous financial expectations surrounding OpenAI, solidified by its $300 billion valuation reported on April 1, form a significant part of the competitive context. This valuation resulted from a tender offer allowing insiders to sell shares, rather than an influx of new operating capital, potentially amplifying the pressure to deliver commercially viable products quickly. The company’s substantial investments in compute infrastructure, evidenced by an $11.9 billion deal with CoreWeave and participation in the multi-phase Stargate Project, further highlight the scale and cost of operating at the AI frontier.
Internal disagreements regarding the prioritization of safety versus development speed are not new for the organization. The departure of Jan Leike, former co-lead of the Superalignment team focusing on long-term risks, in May 2024 was marked by his public statement that “safety culture and processes have taken a backseat to shiny products.”
His subsequent move to competitor Anthropic underscored these tensions. More recently, adding to the internal perspective, a group of former OpenAI employees filed an amicus brief on April 11, 2025, in support of Elon Musk’s ongoing lawsuit against the company. The brief argues that OpenAI’s shift towards a capped-profit structure could compromise its original safety commitments.
OpenAI’s framework update, particularly its potential to adjust standards based on competitor behavior, stands in contrast to recent announcements from some rivals. Google DeepMind proposed a global AGI safety framework on April 3, calling for international cooperation. Anthropic, meanwhile, has publicized technical safety efforts like its interpretability framework, although the company also faced scrutiny for reportedly removing some earlier voluntary safety commitments. As regulatory frameworks like the EU AI Act begin to exert influence, OpenAI’s revised approach introduces a new variable in how leading labs might manage AI risks in an increasingly competitive arena.
