In recent months a succession of high-profile AI claims and incidents—ranging from Anthropic’s late-April assertions about Claude Mythos and alleged mathematical breakthroughs to the OpenAI–Hugging Face hacking episode and a viral resignation by Anthropic engineer Jacob Coxon—have generated intense media attention. Technical experts who later examined these cases often found different explanations: negligence, misattribution, or overstated novelty. This article reviews the events, the expert responses, and the broader implications for policy and public understanding.
What happened (key events)
- End of April: Anthropic announced that its model Claude Mythos was better at finding software vulnerabilities than most security experts.
- Around the same period: following the OpenAI–Hugging Face incident, Anthropic (proudly) and Meta (reluctantly) disclosed similar model-related security incidents.
- Subsequent weeks: Anthropic claimed a mathematical breakthrough by one of its models; OpenAI later announced its own purported mathematical result.
- Most recently: Anthropic engineer Jacob Coxon went public with his resignation, accusing Anthropic and OpenAI of ‘racing straight towards self-improving superintelligence and gambling with our lives.’
Media coverage versus expert analysis
Press coverage of these incidents often echoed companies’ anthropomorphizing language, portraying models as if they were close to artificial general intelligence. Once experts in the relevant fields had time to investigate, a different picture frequently emerged—one that received less attention than the initial announcements.
Security incidents: negligence rather than rogue models?
Cybersecurity specialists argue that many of the so-called “hacking” stories are better explained by company negligence and failures to adopt basic, established security practices than by models that acted autonomously or ‘‘went rogue.’’ Framing incidents as the result of independent, agentic models can deflect responsibility away from the firms that develop and deploy these systems.
Mathematical claims: initial astonishment, later skepticism
Mathematics and programming are attractive domains for companies to showcase large language models because proposed answers can often be verified. That verifiability, combined with cultural prestige around math, makes these fields useful for marketing bold claims. In several cases where companies reported surprising mathematical results—for example, OpenAI’s announcement regarding its chatbot Astra—mathematicians later found the results were not as novel or deep as initially implied. Some mathematicians have accused firms of research misconduct and plagiarism, and have emphasized that the systems did not perform a ‘‘profound intellectual leap.’’
A notable example: Tristan Buckmaster, a professor at New York University’s Courant Institute, published a statement suggesting OpenAI had improperly used other people’s work and misattributed results.
Why companies lean on math and code for marketing
- Math and programming are perceived as pinnacles of intellectual achievement, which helps sell the idea of machines approaching or exceeding human capacities.
- Outputs in these domains are easier to evaluate automatically, reducing the need for costly human annotation when tuning systems.
Hundreds of mathematicians have warned that there is a strong commercial incentive in the technology industry to overstate product capabilities. Their joint statement urges policymakers to consult experts, including mathematicians, rather than rely on press releases or popular reporting when forming policy decisions.
The consequences of ‘‘superintelligence’’ rhetoric
Labeling systems as ‘‘superintelligence’’ or ‘‘rogue models’’ assigns agency to products instead of to the companies that build and operate them. That framing both markets products as ‘‘superhuman’’ and facilitates corporate avoidance of accountability. For example, instead of holding companies legally responsible for security failures or potential malware propagation, commentary often attributes those actions to autonomous models.
Similarly, attention can be diverted from other present harms—such as alleged plagiarism of academic work or the use of customer data to train models without consent—toward speculative fears about future, hypothetical machine gods.
Environmental and community harms sidelined by the hype
Industry voices have argued that popular, bipartisan activism against large data centers is a ‘‘distraction’’ from the need to regulate impending superhuman machines. Critics counter that immediate, tangible harms deserve attention: the climate impact of data centers, higher electricity bills subsidized by the public, increased asthma rates among nearby communities, and water usage for cooling.
Recommendations for policymakers and the public
The authors argue against making decisions based on marketing or corporate pressure to move quickly. Good policy and community decision-making require time to hear independent experts and to contextualize corporate claims. The best outcome of this summer’s wave of hype would be that policymakers and the public learn to pause, retain healthy skepticism, and recognize marketing-driven hype when it recurs.
Authors and context
The piece is written by Timnit Gebru, executive director of DAIR; her book Deep Unlearning: The Radicalization of a Tech Idealist is available for preorder and is scheduled for publication on February 16, 2026. The coauthor is Emily M. Bender, professor of linguistics at the University of Washington and coauthor of The AI Con.
