Earlier than Emily Bender and I’ve checked out a menu, she has dismissed synthetic intelligence chatbots as “plagiarism machines” and “artificial textual content extruders”. Quickly after the meals arrives, the professor of linguistics provides that the vaunted massive language fashions (LLMs) that underpin them are “born shitty”.
Since OpenAI launched its wildly well-liked ChatGPT chatbot in late 2022, AI corporations have sucked in tens of billions of {dollars} in funding by promising scientific breakthroughs, materials abundance and a brand new chapter in human civilisation. AI is already able to doing entry-level jobs and can quickly “uncover new data”, OpenAI chief Sam Altman advised a convention this month.
In line with Bender, we’re being bought a lie: AI won’t fulfil these guarantees, and nor will it kill us all, as others have warned. AI is, regardless of the hype, fairly dangerous at most duties and even the very best methods accessible at the moment lack something that could possibly be known as intelligence, she argues. Current claims that fashions are growing a capability to know the world past the info they’re skilled on are nonsensical. We’re “imagining a thoughts behind the textual content”, she says, however “the understanding is all on our finish”.
Bender, 51, is an knowledgeable in how computer systems mannequin human language. She spent her early tutorial profession in Stanford and Berkeley, two Bay Space establishments which can be the wellsprings of the fashionable AI revolution, and labored at YY Applied sciences, a pure language processing firm. She witnessed the bursting of the dotcom bubble in 2000 first-hand.
Her mission now could be to deflate AI, which she’s going to solely consult with in air quotes and says ought to actually simply be known as automation. “If we wish to get previous this bubble, I feel we want extra folks not falling for it, not believing it, and we want these folks to be in positions of energy,” she says.
In a current e book known as The AI Con, she and her co-author, the sociologist Alex Hanna, take a sledgehammer to AI hype and lift the alarm concerning the know-how’s extra insidious results. She is evident on her motivation. “I feel what it comes right down to is: no person ought to have the facility to impose their view on the world,” she says. Due to the massive sums invested, a tiny cabal of males has the flexibility to form what occurs to massive swaths of society and, she provides, “it actually will get my goat”.
Her thesis is that the whizzy chatbots and image-generation instruments created by OpenAI and rivals Anthropic, Elon Musk’s xAI, Google and Meta are little greater than “stochastic parrots”, a time period that she coined in a 2021 paper. A stochastic parrot, she wrote, is a system “for haphazardly stitching collectively sequences of linguistic kinds it has noticed in its huge coaching knowledge, in response to probabilistic details about how they mix, however with none reference to which means”.
The paper shot her to prominence and triggered a backlash in AI circles. Two of her co-authors, senior members of the moral AI staff at Google, misplaced their jobs on the firm shortly after publication. Bender has additionally confronted criticism from different teachers for what they regard as a heretical stance. “It appears like individuals are mad that I’m undermining what they see because the form of crowning achievement of our discipline,” she says.
The controversy highlighted tensions between these trying to commercialise AI quick and opponents warning of its harms and urging extra accountable growth. Within the 4 years since, the previous group has been ascendant.
We’re assembly in a low-key sushi restaurant in Fremont, Seattle, not removed from the College of Washington the place Bender teaches. We’re virtually the one patrons on a sun-drenched Monday afternoon in Could, and the waiter has bored with asking us what we’d like after half-hour and three makes an attempt. As an alternative we flip to the iPad on the desk, which guarantees to streamline the method.
It achieves the alternative. “I’m going to get a type of,” says Bender: “add to cart. Precise meals could differ from picture. Good, as a result of the picture is gray. That is nice. Yeah. Present me the . . . the place’s the otoro? There we go. Ah, it could possibly be they don’t have it.” We surrender. The waiter returns and confirms they do in actual fact have the otoro, a fatty lower of tuna stomach. Realising I’m British, he lingers to ask which soccer staff I help, presents his commiserations to me on Arsenal ending as runners-up this season and tells me he’s a Tottenham fan. I’m wondering if it’s too late to revert to the iPad.
Menu
Kamakura Japanese Delicacies and Sushi
3520 Fremont Ave N, Seattle, 98103
Otoro nigiri x2 $31.90
Salmon nigiri x2 $8
Agedashi x2 $8
Avocado maki $5.95
Edamame $3.50
Barley tea x2 $5
Whole (together with tax and tip) $82.56
Bender was not all the time destined to take the battle to the world’s greatest corporations. A decade in the past, “I used to be minding my very own enterprise doing grammar engineering,” she says. However after a wave of social actions, together with Black Lives Matter, swept via campus, “I began asking, nicely, the place do I sit? What energy do I’ve and the way can I take advantage of it?” She arrange a category on ethics in language know-how and some years later discovered herself “having simply never-ending arguments on Twitter about why language fashions don’t ‘perceive’, with laptop scientists who didn’t have the primary bit of coaching in linguistics”.
Finally, Altman himself got here to spar. After Bender’s paper got here out, he tweeted “i’m a stochastic parrot, and so r u”. Sarcastically, given Bender’s critique of AI as a regurgitation machine, her phrase is now typically attributed to him.
She sees her function as “with the ability to converse fact to energy primarily based on my tutorial experience”. The reality from her perspective is that the machines are inherently way more restricted than we’ve been led to imagine.
Her critique of the know-how is layered on a extra human concern: that chatbots being lauded as a brand new paradigm in intelligence threaten to speed up social isolation, environmental degradation and job loss. Coaching cutting-edge fashions prices billions of {dollars} and requires monumental quantities of energy and water, in addition to employees within the growing world keen to label distressing pictures or categorise textual content for a pittance. The final word impact of all this work and power will probably be to create chatbots that displace these whose artwork, literature and data are AI’s uncooked knowledge at the moment.
“We’re not making an attempt to alter Sam Altman’s thoughts. We are attempting to be a part of the discourse that’s altering different folks’s minds about Sam Altman and his know-how,” she says.
The desk is now crammed with dishes. The otoro nigiri is comfortable, tender and each bit nearly as good as Bender promised. We now have each ordered agedashi tofu, completely deep-fried so it stays agency in its pool of dashi and soy sauce. Salmon nigiri, avocado maki and tea additionally dot the area between us.
Bender and Hanna had been writing The AI Con in late 2024, which they describe within the e book as the height of the AI growth. However since then the race to dominate the know-how has solely intensified. Main corporations together with OpenAI, Anthropic and Chinese language rival DeepSeek have launched what Google’s AI staff describe as “considering fashions, able to reasoning via their ideas earlier than responding”.
The power to motive would characterize a big milestone on the journey in the direction of AI that might outperform consultants throughout the complete vary of human intelligence, a objective sometimes called synthetic common intelligence, or AGI. Plenty of essentially the most outstanding folks within the discipline — together with Altman, OpenAI’s former chief scientist and co-founder Ilya Sutskever and Elon Musk have claimed that objective is at hand.
Anthropic chief Dario Amodei describes AGI as “an imprecise time period which has gathered a variety of sci-fi baggage and hype”. However by subsequent yr, he argues, we may have instruments which can be “smarter than a Nobel Prize winner throughout most related fields”, “can management current bodily instruments” and “show unsolved mathematical theorems”. In different phrases, with extra knowledge, computing energy and analysis breakthroughs, at the moment’s AI fashions or one thing that intently resembles them may prolong the boundaries of human understanding and cognitive potential.
Bender dismisses the thought, describing the know-how as “a elaborate wrapper round some spreadsheets”. LLMs ingest reams of information and base their responses on the statistical likelihood of sure phrases occurring alongside others. Computing enhancements, an abundance of on-line knowledge and analysis breakthroughs have made that course of far faster, extra refined and extra related. However there isn’t any magic and no emergent thoughts, says Bender.
“In the event you’re going to be taught the patterns of which phrases go collectively for a given language, if it’s not within the coaching knowledge, it’s not going to be within the output system. That’s simply basic,” she says.
In 2020, Bender wrote a paper evaluating LLMs to a hyper-intelligent octopus eavesdropping on human dialog: it would decide up the statistical patterns however would have little hope of understanding which means or intent, or of with the ability to consult with something exterior of what it had heard. She arrives at our lunch at the moment sporting a pair of picket octopus earrings.
There are different sceptics within the discipline, comparable to AI researcher Gary Marcus, who argue the transformational potential of at the moment’s greatest fashions has been massively oversold and that AGI stays a pipe dream. Per week after Bender and I meet, a gaggle of researchers at Apple publish a paper echoing a few of Bender’s critiques. The perfect “reasoning” fashions at the moment “face a whole accuracy collapse past sure complexities”, the authors write — though researchers had been fast to criticise the paper’s methodology and conclusions.
Sceptics are usually drowned out by boosters with larger profiles and deeper pockets. OpenAI is elevating $40bn from buyers led by SoftBank, the Japanese know-how investor, whereas rivals xAI and Anthropic have additionally secured billions of {dollars} within the final yr. OpenAI, Anthropic and xAI are collectively valued at near $500bn at the moment. Earlier than ChatGPT was launched, OpenAI and Anthropic had been valued at a fraction of that and xAI didn’t exist.
“It’s to their profit to have everybody imagine that it’s a considering entity that may be very, very highly effective as an alternative of one thing that’s, you already know, a glorified Magic 8 Ball,” says Bender.
We now have been speaking for an hour and a half, the bowl of edamame beans between us steadily dwindling, and our cups of barley tea have been refilled greater than as soon as. As Bender returns to her fundamental theme, I discover she has quietly constructed an origami hen from her chopstick wrapper. AI’s boosters may be hawking false guarantees, however their actions have actual penalties, she says. “The extra we construct methods round this know-how, the extra we push employees out of sustainable careers and likewise lower off the entry-level positions . . . After which there’s all of the environmental affect,” she says.
Bender is entertaining firm, a Cassandra with a wry grin and twinkling eye. At occasions it feels she is taking part in as much as the function of nemesis to the tech bosses who stay down the Pacific coast in and round San Francisco.
However the place Bender’s bêtes noires in Silicon Valley may gush over the potential of the know-how, she will be able to appear blinkered in one other method. Once I ask her if she sees one constructive use for AI, all she’s going to concede is that it would assist her discover a track.
I ask how she squares her twin claims that chatbots are bullshit turbines and able to devouring massive parts of the labour market. Bender says they are often concurrently “ineffective and detrimental”, and provides the instance of a chatbot that might spin up plausible-looking information articles with none precise reporting — nice for the host of a web site making a living from click-based promoting, much less so for journalists and the truth-seeking public.
She argues forcefully that chatbots are born flawed as a result of they’re skilled on knowledge units riddled with bias. Even one thing as slim as an organization’s insurance policies may comprise prejudices and errors, she says.
Aren’t these actually critiques of society fairly than know-how? Bender counters that know-how constructed on high of the mess of society doesn’t simply replicate its errors however reinforces them, as a result of customers suppose “that is so huge it’s all-encompassing and it may see every thing and so subsequently it has this view from nowhere. I feel it’s all the time essential to recognise that there isn’t any view from nowhere.”
Bender dedicates The AI Con to her two sons, who’re composers, and he or she is very animated describing the deleterious affect of AI on the artistic industries.
She is scathing, too, about AI’s potential to empathise or provide companionship. When a chatbot tells you that you’re heard or that it understands, that is nothing however placebo. “When Mark Zuckerberg means that there’s a requirement for friendships past what we even have and he’s going to fill that demand along with his AI mates, actually that’s mainly tech corporations saying, ‘We’re going to isolate you from one another and be sure that your entire connections are mediated via tech’.”
But employers are deploying the know-how, and discovering worth in it. AI has accelerated the speed at which software program engineers can write code, and greater than 500mn folks commonly use ChatGPT.
AI can also be a cornerstone of nationwide coverage below US President Donald Trump, with superiority within the know-how seen as being important to successful a brand new chilly conflict with China. That has added urgency to the race and drowned out requires extra stringent laws. We focus on the parallels between the hype of at the moment’s AI second and the origins of the sector within the Fifties, when mathematician John McCarthy and laptop scientist Marvin Minsky organised a workshop at Dartmouth Faculty to debate “considering machines”. Within the background throughout that period was an existential competitors with the Soviet Union. This time the Crimson Scare stems from concern that China will develop AGI earlier than the US, and use its mastery of the know-how to undermine its rival.
That is specious, says Bender, and beating China to some degree of superintelligence is a pointless objective, given the nation’s potential to catch up shortly, which was demonstrated by the launch of a ChatGPT rival by DeepSeek earlier this yr. “If OpenAI builds AGI at the moment, they’re constructing it for China in three months.”
Nonetheless, competitors between the 2 powers has created enormous industrial alternatives for US start-ups. On Trump’s first full day of his second time period, he invited Altman to the White Home to unveil Stargate, a $500bn knowledge centre undertaking designed to cement the US’s AI primacy. The undertaking has since expanded overseas, in what these concerned describe as “industrial diplomacy” designed to bolster America’s sphere of affect utilizing the know-how.
If Bender is true that AI is simply automation in a shiny wrapper, this unprecedented outlay of economic and political capital will obtain little greater than the erosion of already fragile professions, social establishments and the atmosphere.
So why, I ask, are so many individuals satisfied it is a extra consequential know-how than the web? Some have a industrial incentive to imagine, others are extra trustworthy however no much less deluded, she says. “The emperor has no garments. However it’s shocking how many individuals wish to be the bare emperor.”
George Hammond is the FT’s enterprise capital correspondent
Discover out about our newest tales first — observe FT Weekend on Instagram, Bluesky and X, and enroll to obtain the FT Weekend publication each Saturday morning
