The short version
Emily Bender and Alex Hanna's The AI Con argues that AI hype and AI doom are two costumes on one sales pitch. Both scripts need the machine to be smarter than you, and both end with the same instruction to buy early.
Book Analysis
Emily Bender and Alex Hanna's The AI Con argues that AI hype and AI doom are two costumes on one sales pitch. Both scripts need the machine to be smarter than you, and both end with the same instruction to buy early.
Scroll to read
01 / The seating chart
Boosters promise a cure for disease, a replacement for your SDR desk, and a quarter-over-quarter miracle. Doomers promise deception, escape, and an ending. They argue about this constantly, in public, at volume.
Bender and Hanna point at the thing neither side brings up. Both scripts need the same machine to exist. One wants a savior and one wants a monster, and the funding comes from the same conviction.
02 / The shared premise
Each script assumes the machine is smarter than you and arriving either way. That assumption is the product.
Remove it and what remains is a text prediction system with an unusually good press operation. Nobody puts that on a slide.
03 / The swap test
Verb stays. Object stays. Scale stays. You change the mood and file the same purchase order.
The model already reasons better than most of your analysts.
Both versions agree: The model reasons.
04 / Who the future is working for
Bender has made this point in interviews since the book came out: the discourse insists the future is already fixed and the only live question is speed. It is a neat trick. Once arrival is inevitable, asking whether the thing works becomes a scheduling complaint rather than a procurement question, and the only rational move left is to get in early.
Roadmaps turn into prophecy. Fear of missing out does the closing.
05 / Verbs that grow a brain
Every one of those verbs quietly installs a mind inside the box. Once a mind is in there, the meeting shifts from evaluation numbers to intentions, which is exactly where a vendor would like it to go.
The authors keep making a very dry observation about this. We do not call airplanes superhuman flyers. We do not call rulers superhuman measurers. It only happens here.
06 / Meanwhile, in the present tense
The data came from whatever was reachable. Cleanup runs through people paid by the task, invisible in the pitch and load-bearing in the product.
Hanna's version of the jobs question is blunter than either camp's forecast: the technology is not coming for your job so much as it is coming to make your job worse.
07 / The merge
Move before the window shuts. Sign before the price moves. One pitch, two costumes, running more or less continuously since 1956.
What to do with this in a vendor meeting
If nobody will describe the training data, the honest answer is everything that was not nailed down. Ask anyway. Watch who stalls.
A statistic without a citation is decoration. Ask for evaluation on your data, in your conditions, before the contract.
Behind most autonomous systems is a person cleaning up after it. Find out whether that person is on your team, on a contract, or entirely hypothetical.
None of this makes the technology useless. Plenty of it works, on narrow tasks, with evidence you can inspect. What the book takes away is the excuse for not checking.
Related reading
Half the AI You're Sold Can't Work
A buyer's bench test built from Narayanan and Kapoor's AI Snake Oil: five tests that tell working AI from the kind that only works in the demo.
You're Buying Into an Empire. Know Whose.
A structural read of Karen Hao's Empire of AI — what OpenAI's mission drift means for anyone procuring frontier AI.
By Thomas Rogers ·
If something here resonated with a problem you're working on, let's spend 15 minutes on it.
Schedule a Discovery Call