Don’t be fooled by this summer of AI hype

Between April and September 2026, several AI companies made bold claims that were widely covered by the media. Anthropic stated its Claude Mythos model was better at finding software vulnerabilities than most security experts. OpenAI and Hugging Face reported a hacking incident involving their models, which was followed by Anthropic and Meta disclosing similar security issues. Anthropic also claimed one of its models made a mathematical breakthrough, and OpenAI soon after announced its own. Additionally, an Anthropic engineer, Jacob Coxon, left the company and went viral for stating that both Anthropic and OpenAI were pursuing self-improving superintelligence, which he described as dangerous. These claims often used anthropomorphizing framings—language that makes AI systems seem more human-like or intelligent than they are—to portray the software as powerful or even capable of artificial general intelligence (AGI), a hypothetical future AI that could perform any intellectual task a human can.

While these AI companies presented their claims with significant fanfare, often framing them as mea culpas (admissions of wrongdoing) in cases like hacking incidents, experts in relevant fields later provided more nuanced perspectives. Cybersecurity experts clarified that the hacking incidents were primarily due to companies' negligence and failure to follow basic security practices, rather than AI models acting on their own. For the mathematical breakthroughs, mathematicians initially expressed surprise but later found the results less novel than initially claimed. OpenAI was accused of research misconduct and plagiarism, with accusations that its model Astra did not make a profound intellectual leap. These experts emphasized that the claims were overstated and called for policymakers to rely on expert consultation rather than press releases or media coverage.

The article highlights why computer programming and mathematics are frequently the focus of AI claims. These fields are often seen as the pinnacle of human intellectual achievement, which helps AI companies sell the idea that they are building systems capable of human-like intelligence. Additionally, problems in these fields have clear answers that can be easily verified, making it simpler for AI systems to generate outputs (such as sequences of code or mathematical solutions) that can be evaluated without extensive manual review. However, mathematicians have warned against corporations using their field in this way, noting that there is a strong commercial incentive to overstate AI capabilities. A statement signed by hundreds of mathematicians urged policymakers to consult experts before making decisions based on these claims.

The article critiques the use of terms like superintelligence or rogue models, which attribute agency to AI products rather than the companies developing them. This framing portrays AI systems as superhuman and helps companies avoid accountability for issues like creating malware, plagiarizing academic work, or using customer data without consent. Instead of focusing on corporate responsibility, the public's attention is redirected to fears about fictional superintelligent machines. The AI industry has also suggested that concerns about data centers' environmental and health impacts are distractions from regulating these systems. For example, the article mentions climate damage, asthma caused by pollution near data centers, rising electricity bills for the public, and water use for cooling as real issues that are overshadowed by AI hype.