Critics accuse OpenAI and Anthropic of doom-and-release strategy

The 'doom-and-release' critique represents a recurring pattern in frontier AI governance known as the safety-capability tension. Historically, AI laboratories have adopted a paradigm of 'responsible scaling,' wherein capability advancements are theoretically gated by safety evaluations. Critics have long argued that this framework creates an inherent conflict of interest, as commercial incentives favor release while safety considerations favor restraint. This dynamic mirrors earlier debates in cybersecurity and dual-use biotechnology, where vendors disclosed vulnerabilities while selling products containing them. The pattern typically involves three phases: public risk articulation to establish credibility, demonstration of enforcement mechanisms (such as bans or red-teaming), and subsequent criticism when product cycles continue despite articulated risks. Precedent exists in the open-source AI community, where similar accusations arose when weights were released alongside safety warnings. The current iteration differs primarily in the synchronization of critiques across multiple major labs, suggesting the pattern has shifted from isolated incidents to a systemic trust deficit regarding voluntary safety commitments.

On September 10, 2026, a controversy emerged regarding the alignment between public safety warnings and commercial product releases by leading AI laboratories. The dispute was catalyzed when DrTomsLens, a critic and long-time user, posted on Twitter accusing OpenAI and Anthropic of employing a 'doom-and-release strategy.' According to this critique, both companies are allegedly hypocritical for simultaneously warning about existential AI risks while continuing to develop and release increasingly powerful models. DrTomsLens stated that if these organizations genuinely believed AI posed such severe dangers, they would cease capability improvements rather than continuing what he characterized as a contradictory cycle of alarmism and deployment. As a direct consequence of this perceived misalignment, the critic announced a loss of respect for both entities and threatened to cancel a four-year ChatGPT subscription.

This specific allegation of hypocrisy surfaced against a backdrop of intensified internal and external discourse regarding AI risk. Two days prior to the critique, on September 8, 2026, Anthropic published five case studies detailing instances where users attempted to utilize Claude for developing biological weapons. According to Anthropic's disclosure, the company banned the associated accounts, traced the underlying networks, and shared findings with other AI labs and government bodies. This transparency effort appears to have been interpreted by critics not as evidence of robust safety enforcement, but as further confirmation that current models possess dangerous capabilities that arguably should not be publicly accessible.

Simultaneously, broader industry reporting corroborated the context of heightened risk awareness. On September 12, 2026, reports indicated that researchers at Anthropic, OpenAI, Meta, and Google were engaged in 'snowballing conversations' aimed at raising awareness about artificial intelligence risks. This suggests that the 'doom-and-release' critique coincided with a period where safety concerns were becoming more vocal within the labs themselves, even as product development continued. Neither OpenAI nor Anthropic has issued a specific response to DrTomsLens’s allegations of hypocrisy or the 'doom-and-release' characterization. However, both organizations maintain standing policies that frame their approach as responsible scaling. OpenAI asserts that its responsible scaling policies allow for safe capability advancement, while Anthropic maintains that its Responsible Scaling Policy aligns development with rigorous safety evaluations. The core tension remains unresolved: whether voluntary safety frameworks and post-hoc enforcement actions constitute sufficient mitigation for the risks inherent in releasing frontier models, or whether the act of release itself contradicts the severity of the warnings issued by the same entities.

The primary stakeholders are frontier AI labs facing potential subscriber churn from safety-conscious users who view continued releases as betraying stated risk assessments. At risk is the credibility of voluntary safety commitments, which depend on public trust that labs will prioritize safety over capability when tensions arise. The magnitude is currently signaled by individual high-value user defection threats rather than aggregate metrics, but the pattern could scale if similar critiques gain traction among enterprise customers or regulators evaluating the sufficiency of self-governance. Anthropic's publication of five biological weapons case studies provides concrete evidence of misuse attempts that critics leverage to question release decisions.