Facebook VP of AI Jerome Pesenti says OpenAI's GPT-3 can “easily output toxic language that propagates harmful biases”, as some critics point out issues of bias
While GPT-3 has earned ecstatic reviews from many experts for its capabilities, some critics have pointed out clear issues around bias.
AxiosBryan Walsh
Context & Ripple Effects
Pesenti's critique lands mid-arc of GPT-3's reception: months after the model drew ecstatic reviews from AI researchers for its language abilities, a rival lab's VP is the highest-profile voice naming the flip side — that scale amplifies the toxic patterns in training data rather than smoothing them out.
The critique forces OpenAI into a defensive posture at peak hype: it must answer a named competitor's charge about bias while simultaneously managing the access pipeline for a model whose outputs are demonstrably harmful without safeguards.
Second-order effects
Bias auditing becomes a competitive weapon between labs — Pesenti's framing sets up 'capability with caveats' versus 'capability without disclosure' as a public positioning battle, pushing every major lab to publish its own harm evaluations first.
Third-order effects
The pattern holds across the corpus: informal criticism matures into formalized pre-release review (OpenAI's red team) and then into quantified improvement claims (GPT-5's stated 30% political-bias reduction), suggesting bias measurement becomes a standard, comparable product feature rather than an external accusation.
The trend: Frontier labs are absorbing external bias critiques into formal pre-release governance, turning harm evaluation from a rival's talking point into a published, benchmarked part of each model launch.
People get mad when AIs do or say politically incorrect things. What if it's hard to prevent them from drawing such conclusions, and the easiest way to fix this is to teach them to hide what they think? That seems a scary skill to start teaching AIs.
Last week I raised concerns about using #gpt3 in production because it can easily output toxic language that propagates harmful biases. I thought it was a pretty uncontroversial stance but the responses ranged from complete misunderstanding of AI to total irresponsibility. 1/13
gpt3 is surprising and creative but it's also unsafe due to harmful biases. Prompted to write tweets from one word - Jews, black, women, holocaust - it came up with these (https://t.co/...). We need more progress on #ResponsibleAI before putting NLG models in production. https://…
=GPT3()... the spreadsheet function to rule them all. Impressed with how well it pattern matches from a few examples. The same function looked up state populations, peoples' twitter usernames and employers, and did some math. https://twitter.com/...
Hi Jerome! It's great to get feedback from someone with so much experience deploying AI at scale. We share your concern about bias and safety in language models, and it's a big part of why we're starting off with a beta and have safety review before apps can go live. https://twit…
Just taught GPT-3 how to turn legalese into simple plain English. All I gave it were 2 examples 🤯 Might build a term sheet and investment document interpreter out of this 🤓 https://twitter.com/...
We need to make AI developers and researchers responsible for what they create. Claiming “unintended consequences” is what lead to the current distrust in the tech industry. We can't let AI become the poster child of that irresponsibility. We need more #responsibleAI now. 13/13
The Turing Test measures whether Artificial Intelligence has advanced enough to fake Natural Intelligence. The Graham Test measures whether AI has reached the point of being able to mimic Orwell's “Protective Stupidity” and thus avoid blurting out politically impious truths. http…
.@sama That's great you acknowledge bias in language models. I first raised concerns awhile and @an_open_mind followed up. Sadly that was not acknowledged but @jackclarkSF tried to discredit me. Women face this a lot more, our voices never get amplified https://twitter.com/... ht…
It's good to review an app before it goes live. Presumably this filters out sketchy uses of GPT-3. But one cannot review the entire range of outputs a language model like GPT-3 can produce even in constrained circumstances. https://twitter.com/...
Finally by far the most disturbing criticism I got was from @paulg who compared my point to forcing AIs to be politically correct. 11/13 https://twitter.com/...
This is a bizarre anthropomorphic view that makes little sense. AIs are not people but algorithms created by humans making deliberate design choices (eg, model, objective, training data). When AIs make sexist or racist statements, these humans should be responsible for it. 12/13
How GPT3 works. A visual thread. A trained language model generates text. We can optionally pass it some text as input, which influences its output. The output is generated from what the model “learned” during its training period where it scanned vast amounts of text. 1/n https:/…
These are important points and I'm glad to see someone in your position making them. Even better would be to link them to the body of scholarship on real-world harms of biases collected & amplified by ML, including work by Latanya Sweeney, Safiya Noble, Ruha Benjamin, Deb Raji >>…
GPT-3 is big: 175 billion params But for context, big models like this exist, and have for some time (160b ICML15) GPT-3 is important for reasons beyond size And when GPT-X does the things it does with a *smaller* model, it will be a *bigger* breakthrough Smaller is harder
And these are the people funding all the startups and companies that are supposed to save us from “unsafe” AI and the sort. This is the network and the paradigm that the people with money and power in Silicon Valley are contemplating. https://twitter.com/...
Ahh the days when Twitter pleasantly surprises you, wasn't expecting powerful AI folks to advocate strongly for #responsibleAI Eager to see how Facebook AI research puts its 💰 where its mouth is 🤞 https://twitter.com/...
Teaching AIs the same “scary skill” that most Americans have had to acquire in order to live their lives without cancellation in our Diversely Strong Nation would humanize AIs, though — make them more like the cowards that most of us are. https://twitter.com/...
We must completely reject the idea that any one person — or company for that matter — should have the power to determine what is “toxic” or “harmful” for billions of people. It is one step from that to a China-style censorship regime enacted over Facebook. https://twitter.com/...
As tech & VC twitter are exploding in excitement over the human-like capabilities of GPT-3, I'd like to remind everyone that at short time scales, “intelligence” is in the eye of the beholder. Here's of thread of examples where it is easy to fall victim to anthropomorphic bias:
An interesting critique of GPT-3 from Facebook AI. But not all applications are like Facebook, where there's a huge number of racist users that you need to protect from themselves. This is a nonissue in many cases! Like basically any b2b software. https://twitter.com/...