An interview with computational linguist Emily Bender, who coined the term “stochastic parrot”, on her AI skepticism, co-writing the book The AI Con, and more
The computational linguist on her motivations for taking on Big Tech, the dangers of chatbots — and why AI is just a ‘glorified Magic 8 Ball’ Bluesky: @hypervisible , @tonytassell , and @nataliegreenpeer Mastodon: @emilymbender@dair … , @emilymbender@dair … , @emilymbender@dair … , @emilymbender@dair … , and @emilymbender@dair … X: @egrefen and @nathanbenaich . LinkedIn: Marcus Weldon Bluesky: @hypervisible : Cool profile, although I'm guessing @emilymbender.bsky.social would use something other than “ai sceptic” to describe herself. Tony Tassell / @tonytassell : “AI is, despite the hype, pretty bad at most tasks and even the best systems available today lack anything that could be called intelligence” - AI sceptic Emily Bender in this Lunch with the FT by @georgehammond.bsky.social. She also says LLMs are “born shitty” www.ft.com/content/9029... Natalie Bennett / @nataliegreenpeer : “AI is pretty bad at most tasks and even the best systems available today lack anything that could be called intelligence”. — If an academic said, it can I get away with “born shitty” in the House, I am wondering? — www.ft.com/content/9029... Mastodon: @emilymbender@dair-community.social : I my talk called “ChatGP-Why: When, if ever, is synthetic text safe, appropriate, and desirable?” I outline criteria for what would have to be true of a use case to make synthetic text a good match for it. … @emilymbender@dair-community.social : For example, when people ask me (as George did) if I see positive use cases for “AI”, I always ask them to specify what they mean by that term for the purposes of the question. In this case, we narrowed it down to large language models run as synthetic text extruding machines. @emilymbender@dair-community.social : More importantly, I would not describe chatbots or LLMs as “a fancy wrapper around some spreadsheets.” That remark comes from The AI Con, where @alex and I are describing the range of things now marketed as “AI”. Sometimes, the marketer is talking about an LLM (or synthetic text extruding machine), but other times, it's even more mundane. … @emilymbender@dair-community.social : My remark that LLMs were “born shitty” was in the context of a discussion of Cory Doctorow's notion of “enshittification”. That process involves something that starts of beneficial for consumers (or at least the consumers in focus). LLMs used as synthetic text extruding machines have no legitimate use cases and — … @emilymbender@dair-community.social : Sitting down for a lunch interview is definitely an interesting experience! Leaving the roughly two-hour conversation with George Hammond, I felt like we'd covered good ground but had no idea what of that would make it into the piece. In the end, I think this covers it pretty well, though I (of course) have a couple of quibbles/some context to add. … X: Edward Grefenstette / @egrefen : I say this as a measured sceptic about LLMs necessarily being the path to AGI (but accepting I might be wrong): Emily Bender is a hack and brings nothing to the table academically or otherwise. The @FinancialTimes has done the field no favours by giving her a platform. Nathan Benaich / @nathanbenaich : it's truly nonsensical to say things like this when genai is legit super useful for consumers and businesses - i mean, i get the academic shtick, but it is automating a ton of work already get over it :) [image] LinkedIn: Marcus Weldon : We reached out to Emily Bender for our Newsweek AI Impact Series interviews but were politley declined and I can see why from this FT story. …
Context & Ripple Effects
Bender’s critique of large language models has been a sustained part of the AI debate since her earlier work explaining what LLMs can and cannot do. This interview extends that line of argument through her work on <i>The AI Con</i>, rather than introducing a new technical claim or product development.
The exchange also sits within a recurring disagreement over what AI risks deserve attention: Bender previously argued that a rogue-AI framing can obscure research into present-day harms. The public pushback from AI advocates in the story makes that divide visible again.
First-order effects
- The interview gives Bender’s case against chatbot reliability and claimed utility a prominent mainstream platform, while giving critics such as Edward Grefenstette and Nathan Benaich a fresh occasion to contest it.
- For readers evaluating LLM deployments, the story sharpens the distinction between persuasive conversational output and evidence of dependable task performance—a concern also raised in earlier criticism of ChatGPT’s generated answers.
Second-order effects
- AI vendors and business adopters face greater pressure to substantiate concrete usefulness rather than rely on broad claims about automation or intelligence, especially when public debate centers on failure modes and harms.
- The dispute may further polarize AI coverage between capability-and-productivity narratives and critiques focused on social consequences, making it harder to treat “AI skepticism” as a single position.
Third-order effects
- If this debate persists, the durable fault line in generative AI will be less whether models can produce fluent output than which uses can be validated, governed, and justified to affected people.
- That would favor evaluation and deployment practices tied to bounded workflows over generalized claims about AI systems, though the interview itself does not establish how quickly organizations will adopt that standard.
The trend: Generative AI’s public debate is shifting from headline capability claims toward contested evidence of utility, reliability, and real-world harm.