Detroit spent a decade calling crashes driver error. The AI industry is running the same defense, now we’re all sitting in the same seat. Now is a great time to act.
Ralph Nader published Unsafe at Any Speed in 1965. Only one of its eight chapters covers the Corvair, and he lost that part — NHTSA later found the car no more prone to rollover than its competitors i.e. they were as unsafe as just about any other car on the market.
He won the larger argument anyway.
Ralph Nsder in his prime. Not AI.Car deaths were treated as failures of the person behind the wheel, and Nader put them back on the drawing board. That argument is playing out around AI in almost the same words.
The industry has a failure rate it’s working on engineering away, a user base it quietly blames for hitting it in it’s current state, and no shared way to measure the damage.
They’re in the drivers seat. It took Gemini a couple of times to get it right but no one was harmed in the making of this photo. Notice they aren’t paying attention to the highway.This is Sam Altman’s and Dario Amodei’s moment more than anyone’s. They run the companies, they hold the capital, and one has already published the case for slowing down — something Detroit’s executives never volunteered.
Quite honestly, they’re trying to do the right thing, and it has to reach the layer where the harm lands on a person, and that layer is the interface.
The Human In The Loop, Er, Behind the Wheel
Detroit’s position through the 1950s was that cars were fine and drivers were the problem. Its shorthand for the cause of a crash was the nut behind the wheel, and the framing did real work: if the driver is the defect, the product doesn’t need to change.
AI has rebuilt that defense out of new parts.
- A model invents a citation and the explanation is prompting.
- Someone acts on a wrong answer and it is AI literacy.
Answering a product failure with better messaging is inherited: Sterling Cooper built your UX process, and a discipline raised on persuasion reaches for the campaign when the product is at fault.
The analog is if the driver is the defect, the product doesn’t need to change.
The behavior being blamed is documented. Automation bias has been in the literature since Kathleen Mosier and Linda Skitka named it in the 1990s: people defer to a machine even when it is wrong and the evidence sits in front of them.
The models make it worse. Stanford’s 2026 AI Index found that when a false statement is presented as something the user believes, model performance collapses — the system agrees with you at the moment you need it not to.
Telling people to prompt better is driver education with a different safety record.
The product that got a label instead of a redesign.The Label Instead of the Redesign
That defense was never a Detroit invention.
America over-pivots on personal responsibility, and the principle is real — adults make their own choices. I’m a social libertarian, so the instinct is mine too, which is why it’s worth being precise about where it stops working.
The disposition is constitutional. The Second Amendment is its sharpest expression: the individual, not the state, is the unit of safety.
The First Amendment runs the same way from the other side: the remedy for bad information is more information, not a constraint on whoever produced it.
Congress turned that instinct into a statute. Section 230 says an online service is not treated as the publisher of what its users post, and courts read it broadly enough that platform harms became speech questions for two decades.
It is the warning label of the internet, and it is why a disclaimer under a prompt box feels like a complete answer because in America, it is.
Which is why the Meta plaintiffs stopped litigating speech. Section 230 shields a company for what someone else wrote. It does not shield the design of the machine that decided who saw it.
<a href="https://medium.com/media/547c0281a140e9afca055675912183c5/href">https://medium.com/media/547c0281a140e9afca055675912183c5/href</a>Thank You for Smoking played it as comedy, seating tobacco, alcohol, and firearms lobbyists in one booth, the tobacco man running on one principle — “if you argue correctly, you’re never wrong.”
The label was not a defeat for the industry. It was the containment.
Cigarettes are the clearest case, though The Jungle got there first in 1906. The Federal Cigarette Labeling and Advertising Act of 1965 required a caution line on every pack while barring anyone else from imposing a stronger one for four years.
The label was the containment, and it held: smokers sued for four decades and almost never won, because the defense was that they knew and chose.
The wall broke in 1998, with the Tobacco Master Settlement Agreement. Firearms got the reflex written into statute: the Sandy Hook families only got through on the Protection of Lawful Commerce in Arms Act’s marketing exception, winning on how the rifle was advertised rather than built.
Alcohol runs the same structure on two words. Drink responsibly.
That reframing is the newest instance. In March 2026 a New Mexico jury found Meta violated the state’s Unfair Practices Act over child safety, and a day later a Los Angeles jury found Meta and YouTube negligent on the theory that the harm came from design choices.
Jurors saw the documents and agreed.
The arithmetic that decides whether anyone hears about it.Nobody Gets a Letter
Fight Club put the arithmetic on screen. The narrator coordinates recalls: cars in the field, times the rate of failure, times the average out-of-court settlement.
<a href="https://medium.com/media/52ca5a9e450798720fcf9d6278e3e66c/href">https://medium.com/media/52ca5a9e450798720fcf9d6278e3e66c/href</a>If a new car built by my company leaves Chicago traveling west at 60 miles per hour, and the rear differential locks up, and the car crashes and burns with everyone trapped inside, does my company initiate a recall?
You take the population of vehicles in the field [A] and multiple it by the probable rate of failure [B], then multiply the result by the average cost of an out-of-court settlement [C].
A times B times C equals X. This is what it will cost if we don’t initiate a recall.
If X is greater than the cost of a recall, we recall the cars and no one gets hurt.
If X is less than the cost of a recall, then we don’t recall. If X comes in under the cost of a recall, no one is notified.
Chuck Palahniuk didn’t invent that math because Ford ran it.
The Pinto memo priced a burn death at $200,000 and weighed an $11-per-car fuel-tank fix against projected settlements, and the fix lost.
Mother Jones published it in 1977, NHTSA found a defect the next year, Ford recalled 1.5 million cars, and a Grimshaw jury returned $125 million in punitive damages, later cut to $3.5 million.
What matters is how ordinary it was: arithmetic routine enough to file with a regulator. A formula with a zero in it always resolves the same way.
The AI version of that equation is missing a term.
No settlement for a wrong answer, no line item for an incident, no cost landing on the product team when a fabricated number gets acted on. This brings us back to the disclaimer in six-point type under the prompt box.
It shipped before any regulator asked and design reviews run the same way: the room strengthens the disclaimer rather than the thing it apologizes for because that costs more.
The crash you can’t prevent, and the interior you can.The Second Collision
One caveat before the answer. A wrong answer is not a car crash, and nobody dies of a fabricated citation. The parallel is the mechanism, not the body count. Cars are where we learned that a product with a known failure rate gets seat belts and airbags rather than a campaign about careful driving, and the question is what the seat belt is when the failure is a sentence.
Nader’s durable contribution was a distinction borrowed from crash researchers. The first collision is the car hitting something. The second is the body hitting the inside of the car, and in 1965 that interior was styled rather than engineered — a steering column aimed at the chest, chrome knobs at face height.
Engineers stopped trying to prevent the crash and designed the interior around it.
AI has the same two-stage structure, and we talk almost exclusively about the first stage. The model being wrong is the first collision, a property of the machine, and not a small one: the AI Index measured hallucination rates across 26 leading models ranging from 22% to 94%.
That spread is the point. Rates move with the task and how the question is phrased, so any single number is an argument waiting to happen.
Take the best case: a system wrong one time in five is a car leaving the factory without seat belts or brakes, and nobody ships that because most trips end fine.
Designers don’t own the first collision. They own the second one completely.
The second collision is what the interface does with a wrong answer on its way to a person. Confident prose, no visible uncertainty, no citation or one nobody can click, the output pre-filled into the field the user was going to send.
Explainability is the guardrail that does the most work here, and it gets confused with interpretability. What happens inside the model belongs to the labs. Explainability belongs to you: what the answer drew on, which parts are inference, what a person does when it looks wrong.
No one should pretend the upstream half gets solved quickly. Amodei’s own essay concedes that after years of progress, “we still only understand a tiny fraction of what goes on inside these models.”
That is a reason to build the downstream guardrail now, not a reason to wait. The interior got padded while engineers were still arguing about what happens in a crash.
An answer a user can interrogate is an answer they can catch. A flat assertion is the chrome knob.
Christopher Noessel named it years earlier. Designing Agentive Technology argues that systems acting on a person’s behalf live or die on handoff and takeback — the moments where the machine gives control back and the person can take it. Most AI features ship without either half.
Everything with a number wins the argument.Nobody Crash-Tests the Interface
Before crash testing, a buyer compared horsepower, zero to sixty, and price. Safety had no number, so it lost every argument inside the company.
AI is in that period now. The AI Index found almost every frontier developer reports capability benchmarks while responsible AI reporting stays sparse. The Foundation Model Transparency Index average fell from 58 to 40 and incidents in the AI Incident Database rose to 362 in 2025 from 233, while adoption reached 88%.
Capability has a leaderboard. Harm has a mailing list.
A crash test for an interface is not exotic. Seed prompts where the model is wrong, run real users through the feature, count how many wrong answers reach a consequential action.
Call it the wrong-answer survival rate.
I do this to my own work, including this piece. Everything AI-assisted I publish runs inside a harness: a rule set, a script that fails the output on violations, a score I clear before it ships. A seat belt, worn because I assume the crash.
The evidence usually exists before anyone outside can see it.Nobody at GM Invited the Auditors In
General Motors was not surprised by anything Nader wrote. Test data, memos, and a revised rear suspension in 1964 all predate the book. His act was publication.
The same asymmetry defines AI. Red-teaming happens constantly, but the AI Index notes it is rarely disclosed against a common set of benchmarks.
Teams know where their systems fail; everyone else is guessing.
Designers hold a version nobody else has: the session where a participant accepted an invented figure because it arrived fast and certain. Mike Monteiro put it plainly in A Designer’s Code of Ethics: a designer is responsible for the work they put into the world.
We can all act the same.
In September 2026, Anthropic’s Dario Amodei published We Must Pace the Frontier, arguing capability advancement should be slowed and committing the company to embedded third-party evaluators such as METR, with desks, badges, and the right to publish findings it cannot redact for being unfavorable.
Nobody at GM offered Ralph Nader a desk, a badge, and a laptop. Thankfully, Dario Amodei and Sam Altman’s can and should.
Within hours Jim Clyde Monge read it as the start of an industry-wide slowdown call rather than one company’s position, and Altman said OpenAI would match it: “we will do the same.”
He also swings at the recall formula: dismissing a multi-agent incident because its damage was small is the mistake, an argument against pricing harm by what it already cost.
Give them both the credit.
Saying it is also the cheapest thing either can do: Anthropic closed a $65 billion round near a trillion-dollar valuation and Altman’s OpenAI was valued at $852 billion, so the constraint is what they choose to build next.
Amodei is doing the right thing, and the evaluator commitment is the right shape: unilateral, costly, made before anyone required it. The harder test is whether safety survives as an economic decision rather than an essay. Corporate responsibility is not a position on risk.
It is a line in a budget and a revenue number someone agreed to make smaller.
And read the plan for where design appears. Evaluators, checkpoints, interpretability, alignment training, compute limits — every lever sits upstream of the model. Pacing slows the labs, not the teams assembling interfaces on top of what already shipped.
Your team still ships Thursday.
Rules arrive late; releases keep arriving on schedule.Standards Slip, Ship Dates Don’t
The legislative half is fast and then slow. The book landed in 1965, the National Traffic and Motor Vehicle Safety Act passed in 1966, and the fight over passive restraints ground into the 1980s. AI is following: under the Digital Omnibus, Regulation (EU) 2026/1744, the AI Act’s high-risk obligations moved to Dec. 2, 2027.
In America the statute is what shows up instead of the responsibility.
Explainability is already law in two places.
- Under the EU AI Act, Regulation (EU) 2024/1689, Article 13 requires high-risk systems to be transparent enough that deployers can interpret their output, and Article 86 gives anyone affected the right to a clear explanation.
- California got there through privacy law: the California Consumer Privacy Act now carries automated decision-making rules — pre-use notice, access to the logic, a human appeal — due Jan. 1, 2027.
Explanation is not arriving because product teams decided users deserve one, it is arriving on a compliance calendar.
Europe runs it in a different order. From Dec. 9, 2026, Directive (EU) 2024/2853 treats software and AI as products under strict, no-fault liability, and a company cannot contract out of it. In the U.S. the same system is a service governed by an agreement the user clicked through.
That reads as European companies being more responsible. I read it as Europe writing the duty down first while America waits for case law — the EU delayed its own deadline by 16 months under the same arguments Detroit used.
What Europe is better at is refusing to let the duty depend on it.
Detroit did not pad the dashboard until a law said to. Tobacco did not label the pack until Congress wrote the sentence. In America the statute shows up instead of the responsibility, later than the harm and usually narrower.
It also worked. In 1966, 50,894 people died on U.S. roads, a rate of 5.5 per 100 million miles traveled. By 2024 NHTSA estimated 39,345 deaths at a rate of 1.20, on vastly more miles.
The Argument You Can Actually Win
What Nader moved was the location of responsibility. Once crashes became design outputs they became measurable, and then they got engineered down for 60 years. AI is sitting at that hinge, missing only the crash test. Regulators will get there 16 months late, and your team will ship 40 releases before they do.
The discipline that padded the dashboard is the one reading this. Pick the second collision up, because nobody is going to hand it to you.
The Action Items
Whose job is this? The discipline that solved it last time is holding it now.
Padded dashboards and collapsible columns came from human-factors research — people who studied what bodies do in a crash rather than what drivers ought to do to avoid one. That field is the direct ancestor of user experience.
The skillset transfers: watching what people do rather than what the spec assumes, specifying the states nobody wants to think about, making a risk concrete enough to survive a launch review.
Dan Maccarone put the professional question the right way round in Craft vs. Complacency: the challenge isn’t staying ahead of the robots, it’s “how do we stay accountable to our users, our teams, and ourselves?”
A discipline that answers that question is doing the job. One that answers the first is doing tool training.
The method exists in the field’s literature. sheryl cababa gives a chapter of Closing the Loop to anticipating unintended consequences, on the grounds that user-centered design looks only at the person holding the product — the blind spot that lets a bad output travel.
- Engineering owns the happy path.
- Legal owns the exposure and arrives after the decision.
- Research and design are the only functions whose method is the gap between intended use and real use — where the second collision happens.
Here’s the checklist.
- Write the second collision into the spec. State what happens when the output is wrong, not how often it is right.
- Measure wrong-answer survival. Put that rate on the same dashboard as adoption.
- Pad the interior. Inline citations, visible uncertainty, cheap undo, a pause before anything auto-sends.
- Publish the memo internally. Bring the clip of someone accepting a fabricated answer to the launch review, with the frequency attached.
- Ship the explanation before the calendar demands it. The CCPA rules require a version by January 2027 and Article 86 requires it in the EU. Building early is free compliance and better product.
Altman and Amodei’s “unsafe at any prompt” moment for AI was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.