There Is No AI Plan Because Nobody Wants to Be Wrong

Bill Gates has spent fifty years impatient for the future to arrive. This morning he published close to six thousand words asking it to slow down.

The essay is unusually candid for him. He discloses his financial ties to the industry and tells readers to weigh that themselves. He says the benefits and the problems are arriving at the same time, which is the part he seems least prepared for. And he offers the sentence that will get quoted everywhere: “There is no plan to ease the entry into the AI era.”

He’s right that there isn’t one. What interests me more is why.

The plan isn’t missing for the reason we might assume

Look at what he actually proposes. National bodies with authority that cuts across agencies, because a labor department understands workforce disruption and a business regulator understands market concentration and neither one is looking at the whole. An international organization drawing on nuclear inspections, aviation rules, and the ozone treaties. A domain of work he calls Human Reserved, set aside for people on purpose. A tax on AI tokens and robots, on the logic that hiring a person incurs payroll tax while buying a machine is a write-off.

Every one of those asks someone to commit to something substantial while the evidence is still incomplete. And we have built a governance culture in which that reads as reckless rather than prudent.

So the absence of a plan is what happens when a whole class of decision has been made conditional on information that arrives after the decision would have mattered. The request for better data before committing is almost always the most reasonable thing anyone says in the room. It is also, functionally, how the delay gets built.

What the ozone treaty got right in 1987

Gates lists the ozone agreements as one model for the institution he wants. It’s the right example to reach for, though the instructive part is the timing rather than the structure.

When Joe Farman and his colleagues published the Halley Bay ozone measurements in 1985, the cause was genuinely unsettled. Researchers were still arguing over whether chlorine was responsible or whether atmospheric circulation was simply moving ozone around, and the aircraft campaigns that would resolve it hadn’t flown yet. Countries signed the Montreal Protocol two years later anyway, agreeing to cut CFC production in half.

Here is the part that rarely gets mentioned. They wrote in a procedure for changing their own timetable as the science came in. Under Article 2, paragraph 9, the parties can tighten the schedule for substances already covered by a decision that binds everyone, without sending the whole treaty back through national ratification. Adding an entirely new substance still requires a full amendment. That split is the whole trick: the slow, expensive process was reserved for genuinely new commitments, and learning faster about commitments already made was made cheap.

They used it six times. A treaty that started at a 50 percent cut ended up at full phase-out, and it got there by adjustment rather than by renegotiation.

Nobody had to predict their way out of that future. They committed to a direction and built the ability to learn and correct into the commitment itself.

I’ve started calling that second half the adjustment provision, because I think it’s the most portable thing in the treaty and almost nobody borrows it. An adjustment provision is a pre-agreed mechanism for changing the terms of a decision as evidence arrives, without reopening the decision itself. It is what makes an early commitment survivable.

You can make a decision reversible on purpose.
You can document what would have to be true for it to hold, and when you’ll go back and check.

The three parts of an adjustment provision

A direction rather than an endpoint. Montreal committed to reducing CFCs, not to a final number that would have to be defended forever. The direction was the thing the evidence supported in 1987; the magnitude was the thing it didn’t. Separating those two is most of the work.
In practice: when you write the decision down, put the direction in the commitment and the magnitude in the schedule. They should live in different sentences, because one of them is going to move.

A named mechanism for changing the schedule. Not a vague willingness to revisit. An actual procedure, agreed in advance, that says who can change what and by what process. The reason Montreal’s worked is that it was cheaper than renegotiating, so people used it instead of avoiding it.
In practice: name the person who can adjust, the threshold that lets them, and the process it takes. If adjusting is more expensive than quietly persisting with a plan you’ve stopped believing in, everyone will quietly persist.

A date and a trigger. The parties adjusted on the basis of scheduled scientific assessments, which meant nobody had to be brave enough to reopen the question. The calendar reopened it.
In practice: pick the date when you’ll look again, and write down the specific finding that would change your mind. The second half is the one people skip, and it’s the one doing the work.

The reason this matters right now is that AI decisions have exactly the shape the provision was designed for. The direction is legible and the magnitude isn’t. Nobody sensible is confident about which roles will be affected how fast, which is precisely why so many organizations are holding still — and holding still is itself a commitment, made without any of these three parts attached.

After all, waiting for certainty is itself a decision.
And it cedes the outcome to whatever is already in motion.

What leaders can do this week

Take one AI decision you’ve deferred and separate its two halves. Write the direction you believe in as one sentence, and the magnitude you’re unsure about as a separate schedule with dates. You will usually find you’re more confident about the direction than the deferral implied.

Attach the adjustment terms before you commit, not after. Try something like: “We’re moving customer triage to an assisted model over three quarters. Reviewed at the end of each quarter by the ops lead, who can extend the timeline without escalation. We revisit the whole approach if handle time doesn’t improve by 15 percent, or if agent attrition rises above the current baseline.” That’s three sentences. It converts an irreversible-feeling decision into a reversible one, and it takes about a minute to write.

Make the reversal cheaper than the persistence. This is the one that actually determines whether any of it works. If changing course requires an executive apology and reopening a budget, nobody will change course. Give someone standing authority to adjust within a defined range, and the plan will get corrected instead of defended.

A short version to keep:

• Name the direction in one sentence; put the magnitude in a dated schedule
• Write down the specific finding that would change your mind
• Set the date you’ll look, before you need it
• Name who can adjust, and within what range
• Check that adjusting is cheaper than persisting
• Log what you assumed, so the next decision inherits it

This is the practical core of deciding, communicating, and acting responsibly under uncertainty, and it’s closer to hand than most of the institutional machinery Gates is calling for. Individual organizations can do it unilaterally, starting now.

The people who don’t get to wait

There’s a reason to hurry that has nothing to do with competitive advantage.

The Stanford Digital Economy Lab found that after generative AI became widely adopted, employment fell measurably among young workers in the occupations most exposed to it while holding steady among their older colleagues. Whatever the eventual size of the effect, someone graduating this spring is already inside it. She doesn’t get to wait for the evidence to firm up before deciding what to do with her twenties.

There’s an important asymmetry to note there Waiting is free for the institutions doing the waiting and expensive for the people the decision lands on. Gates makes essentially this argument himself in his most personal passage, about the caregivers who looked after his father through Alzheimer’s and understood him when he couldn’t say what he needed — the origin of the Human Reserved idea. He’s honest that he can’t yet answer who would decide what gets reserved, or what would stop a company from automating inside the boundary anyway. Those questions get worked out in public or they don’t get worked out.

He isn’t the only one arriving here. Pope Leo XIV’s encyclical on safeguarding the human person in the age of AI got there in May, which is worth noticing mostly as evidence of how many institutions are now trying to answer a question governments have been deferring.

What matters next for AI governance

In the AI policy discussions I’ve taken part in around the UN’s High-level Advisory Body on AI, and in the board rooms where I facilitate strategy conversations about AI and governance, the same pattern shows up at both scales. Broad agreement on direction. Then a request for better information before anyone commits. The information arrives, and by then the decision has changed shape. Montreal’s drafters got around this by agreeing in advance to be wrong on a schedule, which still strikes me as the most practical idea in the treaty.

The moral position here is unglamorous. You will not know enough. Commit anyway, in a direction you can defend, with the correction written in — and be the kind of organization that can be wrong out loud without it costing anyone their standing. That’s not a lower standard than certainty. It’s the only one that survives contact with a situation that keeps moving.

What’s one decision you’ve been holding open, where the thing you’re actually waiting for is permission to be wrong later? You can grant that yourself. It takes three sentences, and it’s the closest thing to a plan any of us are going to get this year.