Q&A with Princeton CS professor Arvind Narayanan on why he calls ChatGPT a “bullshit generator”, his worries over its boom, developing his AI taxonomy, and more
Hello, friends, — If you have been reading all the hype about the latest artificial intelligence chatbot … Mastodon: @JeroenJeremy@mastodon.green . Tweets: @random_walker , @juliaangwin , @random_walker , @harrymccracken , @jonathandriddle , @digiphile , and @fuzheado Mastodon: JeroenJeremy / @JeroenJeremy@mastodon.green : Interesting take from the Princeton professor who calls ChatGPT a bullshit generator: — “If you have been reading all the hype about the latest artificial intelligence chatbot, ChatGPT, you might be excused for thinking that the end of the world is nigh.” … Tweets: Arvind Narayanan / @random_walker : In a conversation with @JuliaAngwin about AI hype, I argue that the most worrisome uses of ChatGPT aren't malicious ones but rather everyday people and organizations using it to cut corners due to the pressures they face, like CNET's misguided experiment. https://themarkup.org/... Julia Angwin / @juliaangwin : ChatGPT is great at its job. But its job is to be a bullshit generator — an automated creator of plausible lies. “It is very good at being persuasive, but it's not trained to produce true statements” @random_walker says in my latest newsletter. /1 https://themarkup.org/... Arvind Narayanan / @random_walker : In our AI Snake Oil book blog, @sayashk and I describe ChatGPT as an unparalleled bullshit generator. We've found this to be a useful framework to understand the power, limits, and dangers of large language models. https://aisnakeoil.substack.com/ ... Harry McCracken / @harrymccracken : “It is very good at being persuasive, but it's not trained to produce true statements. It often produces true statements as a side effect of being plausible and persuasive, but that is not the goal.” https://themarkup.org/... Jonathan D. Riddle / @jonathandriddle : A skeptical computer science professor provides some helpful context and criticisms of ChatGPT and the broader turn toward AI. https://twitter.com/... Alex Howard / @digiphile : This list of 18 pitfalls for journalists to avoid in reporting on artificial intelligence is more broadly applicable to emerging technologies in general. This matrix will be useful to anyone who wishes to spread insight instead marketing snake oil: https://aisnakeoil.substack.com/ ... https://twitter.com/... https://twitter.com/... Andrew Lih / @fuzheado : One of my favorite quotes is: “The truth is paywalled, but the lies are free.” With ChatGPT and generative AI, the “lies” created distort the equation even more. Excellent interview about these risks with @random_walker by @JuliaAngwin https://twitter.com/...
Context & Ripple Effects
When [[a:985866|generative ML was framed in December as a step change raising questions about tolerances for error]], most commentary treated fluency as the story. Princeton's Arvind Narayanan — who with Sayash K runs the AI Snake Oil blog — flips that: his 'bullshit generator' label argues ChatGPT's defining trait is producing plausible statements indifferent to truth, and he is building an AI taxonomy to characterize what large language models can actually do versus what the boom assumes.
The interview lands mid-debate rather than starting one: researchers have since documented ChatGPT producing clean, convincing text that repeats conspiracy theories while sometimes debunking falsehoods, and journalists and academics are openly criticizing six months of generative-AI hype. Narayanan's framing gives the skeptical side a vocabulary at exactly the moment OpenAI itself is doing Q&As about fixing problems after its chatbot went viral.
First-order effects
- Journalists and academics covering the boom gain a ready-made critical frame — Narayanan's taxonomy and 'bullshit generator' label give newsrooms language for accuracy failures that earlier coverage described only as hallucinations or errors.
- OpenAI faces a sharpened credibility critique: its own team is already answering questions about training choices and post-virality fixes, and the taxonomy raises the bar for what counts as addressing them.
Second-order effects
- The empirical record strengthens the critics' hand — documented cases of the model laundering misleading narratives turn a philosophical objection into a product-safety argument that pressures OpenAI's safety and policy teams.
- Media outlets face a self-examination problem: with academics cataloguing hype in generative-AI reporting, editors must decide whether to cover LLM claims as capabilities or as unverified outputs requiring checking.
Third-order effects
- If the pattern holds, the industry conversation shifts from benchmarking fluency to building trust infrastructure around synthetic text — verification workflows and accuracy standards become the differentiator, not raw generation quality.
- Academic taxonomies like Narayanan's position universities as the arbiters of what these systems actually do, a counterweight to vendor-defined narratives that regulators and newsrooms may end up citing.
The trend: The generative-AI debate is moving from celebrating output quality to interrogating truthfulness, with academic skeptics supplying the framework that journalists, vendors, and eventually regulators argue within.