MetaEra: AI 'Hype Summer' Masks Marketing and Misconduct

ME AI messages: Over the past several months, hype around AI has continued to escalate. At the end of April, Anthropic claimed that its model, Claude Mythos, outperformed most security experts in detecting software vulnerabilities. This was followed by a hacking incident between OpenAI and Hugging Face, after which Anthropic (quite proudly) and Meta (reluctantly) disclosed similar incidents involving their respective models. Soon after, Anthropic announced that one of its models had achieved a mathematical breakthrough; shortly thereafter, OpenAI claimed its own mathematical breakthrough. Most recently, Anthropic engineer Jacob Coxon went viral online, announcing his departure from the company and asserting that Anthropic and OpenAI are “racing headlong toward self-improving superintelligence, betting our lives on it.” Most of these events have been reported by the media with near-fanatical fervor, often repeating corporate anthropomorphic narratives designed to portray their software not merely as powerful, but as on the cusp of becoming “artificial general intelligence.” So what is the truth? Are we witnessing a series of civilization-altering technological breakthroughs—or a marketing campaign? In all these incidents, corporate grandstanding—sometimes disguised as confessional statements in the case of illegal hacks—has been accompanied by intense media coverage. By the time experts in the relevant fields have had time to examine the details, an entirely different story often emerges—though one that receives far less media attention. Regarding these “hacking” incidents, cybersecurity experts have stated that the issue lies less in “model runaway” or “AI agents creating civilizations,” and more in OpenAI’s negligence and failure to adopt basic, well-established security practices. As for the mathematical claims, mathematicians were initially “shocked” by OpenAI’s press release, which claimed its latest chatbot, Astra, had solved problems “that had remained unresolved for at least a decade with virtually no progress.” But they soon realized these results were not as novel as they first appeared. Mathematicians later accused the company of research misconduct and plagiarism, reiterating that Astra had not achieved any “profound intellectual leap.” Just weeks later, OpenAI announced yet another mathematical breakthrough. Just two days ago, Tristan Buckmaster, a mathematics professor at NYU’s Courant Institute, issued a bombshell statement suggesting OpenAI had stolen others’ research and made improper claims of authorship. Claims about the imminent arrival of dangerous superintelligence are not grounded in sound science or engineering practice. Rather, they stem from ideologies such as transhumanism and eugenics, along with wishful fantasies about imagined future digital humans. It is worth considering why computer programming and mathematics have become focal points for applying large language models and related technologies. Not only are they often viewed as the pinnacle of human intellectual achievement, but they also involve problems whose solutions can be verified once proposed. The former quality helps AI promoters sell the idea that “they are building omniscient machines”; the latter makes math and programming problems easier to use for debugging and optimizing systems, since outputs—possible sequences of words or snippets of code—can be directly evaluated without needing to pay humans to do so. We endorse this call and note that the illusion of speed and urgency cultivated by tech companies is also a strategy to mislead policymakers and the public. Unfortunately, this strategy sometimes works—for example, in the case of Senator Bernie Sanders’ bill aimed at halting the development of “artificial superintelligence.” Though well-intentioned, the proposal ultimately misses the mark. Labeling these products as “superintelligence” or “runaway models” effectively attributes agency to the products rather than to the corporations that build them. This narrative simultaneously elevates corporate products as “beyond human” while allowing companies to evade responsibility for their own actions. OpenAI creating malware and hacking another company should have been held accountable; instead, press releases, media outlets, media personalities, and legislators framed it as a “runaway model,” as if the model acted on its own. Researchers should have been questioned about why their companies routinely plagiarize scholars’ work or train models on customer data without consent; but public imagination has been steered elsewhere—toward fears of fictional superintelligent machines yet to arrive. The AI industry even suggests that popular, bipartisan opposition to data centers is a distraction meant to divert attention from regulating the terrifying “superhuman” machines these companies are allegedly about to unleash. According to AI industry rhetoric, we should be more afraid of a fictional machine-god than of climate disasters worsened by data centers, asthma suffered by nearby residents, rising electricity bills subsidized by the public, or water resources diverted solely to cool these facilities. We are clear: decisions must not be made based on marketing hype or under corporate pressure for so-called efficiency. Whether for policymakers or communities, wise decisions require time to hear from independent experts and to place corporate claims within concrete context. The best outcome of this “summer of hype” may be that policymakers and the public learn to take a deep breath, remain skeptical, and recognize such hype for what it is when it next appears. Timnit Gebru is Executive Director of DAIR and author of the forthcoming book *Deep Unlearning: The Radicalization of a Tech Idealist*, available for pre-order and set for publication on February 16. Emily M. Bender is Professor of Linguistics at the University of Washington and co-author of *The AI Con*. (Source: MLion)