Don’t be fooled by this summer of AI hype
Sep 22, 2026, 4:04 AM · MIT Technology Review
Gebru and Bender argue this summer’s AGI theater — hacks, math “breakthroughs,” resignation fireworks — collapses when domain experts get time to look.
Why it matters
MIT Technology Review published Timnit Gebru and Emily M. Bender’s case that recent AI headlines are marketing dressed as civilization-scale breakthroughs. Their list runs from Anthropic’s Claude Mythos vulnerability claims through OpenAI–Hugging Face hacking coverage, dueling math announcements, and Jacob Coxon’s viral exit warning about a race to self-improving superintelligence.
The stake isn’t whether models are useful. It’s whether anthropomorphizing press cycles redirect accountability from companies to “rogue models” — and stampede policymakers into the wrong fights.
From the desk
We’re with the skepticism even when we don’t share every ideological frame. The pattern they describe is real: maximal fanfare on day one, quieter expert corrections later. Cybersecurity pushback that the “hacking” stories were more about neglected basic controls than models founding civilizations. Mathematicians who went from stunned to accusing overclaim and, in Tristan Buckmaster’s statement as cited, improper attribution around OpenAI’s Astra math narrative. Hundreds of mathematicians signing that industry has a commercial incentive to overstate results.
Where we land as a desk: useful AI is real — coding help, search, tooling — and hype laundering helps vendors dodge product liability. Calling malware incidents “rogue models” softens the company that shipped the stack. Talking superintelligence while communities fight data-center siting, power bills, and water use is a political magic trick. Gebru and Bender are right that agency belongs with builders and deployers, not with fictional machine gods.
We don’t need to accept every historical analogy in the essay to accept the operational advice: slow down, consult independent experts, read past the press release. Senator Sanders’s proposed superintelligence legislation gets dinged here as well-meaning and misguided — a reminder that bad ontology makes bad law.
I’m watching whether newsrooms change sourcing habits after this summer, whether math and security communities keep publishing rapid rebuttals, and whether regulators keep getting pulled toward sci-fi risk categories while mundane harms — plagiarism disputes, non-consensual training data, local environmental load — move slower. Pro-tech with eyes open means shipping useful tools and refusing the AGI sales trance.
Context
MIT Technology Review opinion by Timnit Gebru (DAIR) and Emily M. Bender (University of Washington), September 22, 2026. Characterizations of specific incidents reflect the authors’ account and the sources they cite.
Who feels it
- Policymakers
- Pressure to wait for independent technical review before drafting around “superintelligence” frames; consult domain experts, not launch blogs.
- Newsrooms and analysts
- Day-two expert reality checks deserve comparable oxygen to day-one corporate claims.
- AI companies
- Anthropomorphizing incident reports may buy attention and cost credibility — and invite harder liability theories later.
- Researchers in math and security
- Continued public corrections are load-bearing infrastructure against benchmark theater.
What to watch
- Legislative text that still centers ASI versus bills aimed at data centers, transparency, and liability.
- Whether OpenAI/Anthropic math and security claims receive formal third-party write-ups.
- Shifts in how major outlets headline agent incidents — company vs “rogue model.”
- Uptake of the mathematicians’ open letter language in policy hearings.