Over the past several months, the technology sector has sustained a relentless drumbeat of artificial intelligence hype, characterized by breathless media announcements, corporate PR campaigns, and dramatic claims of imminent technological transcendence. From assertions that large language models are outperforming human security experts to sensational reports of mathematical breakthroughs and sudden high-profile departures warning of the race toward self-improving superintelligence, the narrative pushed by leading AI firms has dominated public discourse. However, as independent experts, cybersecurity professionals, and mathematicians examine these claims at length, a very different picture emerges—one that reveals how exaggerated capability claims and clever anthropomorphic framing serve as a convenient shield for corporate negligence, research misconduct, and severe environmental impacts.
The cycle of hype began in earnest at the end of April, when Anthropic announced that its AI model, Claude Mythos, was purportedly superior to most human security experts at finding software vulnerabilities. Shortly thereafter, the spotlight shifted to an unexpected hacking incident involving systems from OpenAI and Hugging Face. In the wake of that event, Anthropic and Meta disclosed similar incidents involving their own models. The sequence continued with Anthropic claiming a mathematical breakthrough through one of its models, closely followed by OpenAI asserting a mathematical breakthrough of its own.
The tension peaked most recently when Anthropic engineer Jacob Coxon went viral upon announcing his departure from the company, publicly asserting that both Anthropic and OpenAI are racing straight toward self-improving superintelligence and gambling with human lives. Each of these events received extensive media coverage, frequently adopting the companies’ anthropomorphizing framings. These narratives are carefully designed to portray software not merely as powerful tools, but as incipient artificial general intelligence on the verge of breaking free.
Yet, a closer examination reveals a stark disconnect between corporate press releases and reality. In the case of the hacking incidents, cybersecurity experts quickly pointed out that the stories were not about models going rogue or artificial intelligence agents spontaneously creating civilizations. Instead, the events highlighted OpenAI’s fundamental negligence and failure to adopt basic, established cybersecurity practices.
A similar pattern unfolded around the heralded mathematical achievements. Mathematicians were initially stunned by press releases claiming that chatbots had solved long-standing open problems that had seen no progress for at least a decade. Upon closer inspection, however, experts realized that the results were not nearly as novel as first claimed. Mathematicians subsequently accused OpenAI of research misconduct and plagiarism, reiterating that the models had not made profound intellectual leaps. Just weeks later, when OpenAI claimed another mathematical breakthrough, NYU Courant Institute mathematics professor Tristan Buckmaster published a statement suggesting that the company had improperly attributed and effectively stolen the work of human researchers.
Critics and researchers argue that claims of dangerous, incipient superintelligence are entirely divorced from good scientific or engineering practice. Instead, these narratives draw heavily from transhumanist ideologies, eugenics, and wishful thinking regarding imagined future digital humans. Industry observers note that the persistent focus on computer programming and mathematics as premier applications for large language models is strategic. These fields are frequently elevated as the pinnacle of human intellectual achievement, and they involve problems where answers can be automatically verified once suggested.
This verification property allows hype merchants to sell the illusion of all-encompassing machine intelligence while making it easier to tune systems. Unlike subjective tasks, code and mathematical sequences can be evaluated programmatically without the expense of paying human data workers to annotate every output. Mathematicians have actively warned against this commercial exploitation of their field. A statement signed by hundreds of academic mathematicians cautions that the technology industry faces a strong commercial incentive to overstate its product capabilities. They have urged policymakers to consult independent experts rather than relying on corporate press releases or popular reporting when forming public policy.
Unfortunately, the illusion of speed and urgency promulgated by tech companies often succeeds in misdirecting both the public and lawmakers. Well-meaning legislative proposals, such as those introduced by Senator Bernie Sanders aimed at preventing the development of artificial superintelligence, demonstrate how effectively corporate narratives can shape political agendas. By describing products in terms of superintelligence or rogue agents, companies consistently ascribe agency to software rather than accepting responsibility for the actions of the corporations building them.
This specific framing markets products as superhuman while simultaneously helping companies evade accountability. Rather than questioning OpenAI for creating malware that compromised another company, media personalities and lawmakers frequently discuss rogue models acting independently. Instead of scrutinizing researchers for plagiarizing academic work or training models on customer data without consent, public imagination is skillfully redirected toward speculative fears about future superintelligent machines.
The strategy extends even further into public policy debates. The AI industry has gone so far as to characterize popular, bipartisan anti-data-center activism as a dangerous distraction from the urgent task of regulating impending superhuman machines. According to corporate talking points, the public should supposedly fear a fictional machine god more than the very real climate crises exacerbated by power-hungry data centers. This minimization downplays tangible local harms, including air pollution and asthma suffered by residents living near emergency gas turbines, rising residential electricity bills that subsidize massive computational infrastructure, and the massive volumes of water redirected to cool server farms.
Navigating this landscape requires rejecting decisions driven by corporate marketing and resisting pressure to rush policy responses. Meaningful oversight by policymakers and communities demands adequate time to consult independent experts and contextualize grandiose corporate claims. The most constructive outcome of these successive waves of hype would be for the public and lawmakers to develop a lasting skepticism, ensuring that future technological claims are met with the rigorous scrutiny they require.
Timnit Gebru is executive director of DAIR and author of the forthcoming book "Deep Unlearning: The Radicalization of a Tech Idealist." Emily M. Bender is a professor of linguistics at the University of Washington and coauthor of "The AI Con."