Every few days the same technology gets described two ways. To investors it will cure cancer and rebuild the economy. To regulators and the public, it might end the species by the end of the decade. The same companies say both, sometimes in the same week. And to the bond market, those same companies, still posting billion-dollar losses, are being readied for a credit rating that would call them safe enough to hold your pension. You are not imagining the whiplash, and you are probably tired of it.
If you are also a little afraid, that is not a failure of composure. “This might kill everyone” is a frightening sentence, and fear is a reasonable response to a frightening sentence. Clear thinking is not the opposite of caring, and it is not a trick for making the fear go away. It is a way to keep your footing while you feel. What follows is not an argument that the fear is silly. It is a small set of tools for holding it without being steered by whoever is loudest.
Here is why the tools matter, and not just the facts. Research on fear-based messages is consistent on one point: fear without a sense that you can do something does not produce clarity. It produces avoidance. People tune out. So the aim here is not to frighten you more, and not to talk you down. It is to hand you something to do with what you already feel.
If you take nothing else, take these three questions
You can ask them of a terrifying claim and a reassuring one with equal force. They are the whole tool. Everything after this is only why they work.
- What exactly, and when? Make the claim name a specific thing, a specific person or company, a specific date. Certainty that evaporates the moment you ask for specifics was never in the facts. Try it: a report says AI companies are becoming “investment grade.” Ask what exactly earned the grade, and when. If a company lost twenty billion dollars last year and the rating is being requested by its own bankers ahead of a stock-market debut, “investment grade” is a hope pointed at the future, not a fact dated to the past.
- Is this a measurement or a mood? Ask whether a number was taken once or is still moving, and whether “this is a risk” has quietly become “this will happen.” Try it: a credit rating is supposed to be a measurement, an independent judgment of whether a company can pay its debts. When the rating is being pushed for by the banks that earn fees on the sale, the label is doing the work of a mood while wearing the clothes of a measurement.
- Who is helped if I believe this, and who is helped if I dismiss it? This is the one that guards you from both sides at once. A warning can sell a moat. A shrug can sell business as usual. When a claim flatters the interests of the person making it, that does not make it false. It makes it worth a second look before it moves you. Try it: who is helped if you believe a loss-making company is safe enough for conservative money? The banks underwriting its offering. The chipmaker that guaranteed up to a hundred billion dollars of its leases and is released from that promise the moment a good rating arrives. The label serves them whether or not it serves the retiree whose pension it unlocks.
Screenshot those. Close the page if you want. You are already better defended than most of the argument.
Why they work: the moves they catch
If you want the deeper reading, here it is. The three questions catch a handful of moves that make AI claims, hopeful or terrifying, feel more settled than the evidence beneath them. These are not our invention. Researchers who study risk communication, the misuse of statistics, and the rhetoric of technology hype have named them for years. We simply saw them all at once in a single viral video, and then caught ourselves making one. Four show up again and again.
The relabel. A precise finding gets renamed as a bigger, scarier one, and the bigger name does the persuading. A study about suicide and unemployment gets called “deaths of despair,” a broader and grimmer category it did not measure. The move runs the other way too, dressing danger as safety. A company that credit analysts still place deep in speculative grade gets relabeled “investment grade,” and the new label is the thing that lets pension funds and insurers buy its bonds. SpaceX did this first, winning investment-grade ratings within two weeks of going public and then selling twenty-five billion dollars of bonds; the same banks are now pressing for OpenAI and Anthropic to be rated that way on their own market debuts, while both are still losing money. Question 1 catches it: does the source use the same word the claim uses, and who chose the word?
The stitched-together “them.” Distinct people, companies, and years get sewn into one guilty actor. One firm lobbying for its own interests in one year, and a different firm’s researchers warning about extinction in another, become “the industry,” a single mind doing one coordinated thing. Question 1 again: who, exactly, and when?
The fixed number on a moving thing. A count that is still climbing - views, users, dollars committed - gets frozen into a fact and quoted as though it were final. Question 2 catches it: was this measured once, or is it a rate wearing the costume of a number?
Exposure dressed as catastrophe. A real structural risk gets inflated into a predicted outcome. A sentence like “this debt is being repriced and moved toward people less able to absorb it” can be careful and supportable. “This will collapse the economy” is a forecast, and forecasts carry confidence, not evidence. Question 2 once more: has “risk” quietly become “outcome”?
You will notice none of these requires anyone to lie.
We did it too
One of those four examples is our own mistake. Checking that viral video, we found a view count lower than the one the video quoted, and we wrote that the video’s number was “inflated.” We were wrong. The count was climbing the whole time and passed the video’s figure within days. We had frozen a moving number to fit the story we were telling about someone else freezing a moving number. The reflex does not spare the people trying to catch it, which is the only reason we mention it. A guide that could not admit doing the thing it warns against would be one more voice performing certainty.
This is not lying, which is why it is hard to stop
It would be easier if these moves were deception. Mostly they are not. Calibrated language - “it depends, here is the range” - does not travel. A confident number does. So the reward quietly favors precision the evidence has not earned, and sincere people reach for the reward without deciding to. The danger is not villains. It is drift, under an incentive that pulls everyone the same direction, the skeptics included, and us.
There is a quiet encouragement underneath this. Researchers who study how people trust numbers have found that giving a range instead of a single figure does not cost a communicator credibility: it leaves trust in the source unchanged, and stating the range up front protects that trust for later, when the estimate has to move. Asking for the range is not cynicism. It is what the most trustworthy people in any field already do. When you ask “measurement or mood,” you are not being difficult. You are holding the claim to the standard the careful ones hold themselves to.
What this actually is
None of this tells you what to conclude about artificial intelligence. That is deliberate. This is not our verdict handed to you. It is closer to a vaccine: a small, early exposure to the moves, so that when a real one arrives, and it will, from someone who genuinely believes it, you recognize the shape and keep your feet. It works best before your mind is made up, which is most of us, most of the time.
You do not have to resolve the future to keep your ability to think inside it. That ability is the scarce thing everyone in this argument is spending on your behalf. Keep it for yourself.