Scientists and screenwriters alike estimate that they keep somewhere between one and five percent of what AI offers them.
getty
The machine is a tireless crane. In a single session, it hands us more material than any dissertation, annual report, or Hollywood script could ever use. I’m reminded of this every time I talk to scientists and screenwriters, who tell me — independently, almost word for word — that they keep somewhere between one and five percent of what the machine offers them.
That number may sound alarming. It isn’t. Both groups insist the machine’s real value was never the volume of material it generates. It’s the time it saves them gathering that material in the first place.
I tested this idea out loud recently, during a lecture on AI and judgment in Chicago. I built it around a story I’d heard directly from Dr. Jonathan Stokes, a biochemist at McMaster University whose lab uses AI to hunt for new antibiotics. Searching a near-infinite chemical universe, Stokes’s team asked a model to generate candidate molecules. It returned 24,335. Of those, 2,868 were predicted to be viable antibiotics. From that pool, the team selected 58 to test in the lab. Screening the same chemical space by hand — to arrive at that many active molecules — would have taken them about a year.
One year of laboratory time, handed back to a team of scientists in the time it takes to run a query. That is the crane at work, and it deserves to be named plainly, without irony or hedging: This is what artificial intelligence is extraordinarily good at, and no honest argument about its limits should pretend otherwise.
The scale only grows from there. Stokes describes a second project now under review: “We took an AI model, gave it 12,347,622 molecules from an online chemical repository, and asked it to identify new chemicals that could be antibacterial. We brought some of the model’s predictions into the lab and found multiple structurally novel chemicals with promising activity against clinically burdensome bacteria.” The computation took roughly a day. Testing all 12 million by hand, he estimates, would have taken about four years.
The Crane Saves Time, The Humans Apply Judgment
So here is the question I put to the room in Chicago, and the one I want to put to you: What did Stokes’s team do with the year they got back?
They didn’t spend it generating more molecules. They spent it discriminating among the ones they already had — running the survivors through the slower, far more human work of testing, refining, and arguing about which anomalies were worth chasing and which were noise. The machine bought back time. The team spent it on judgment, not on further searching.
This is where the story stops being about science and starts being about every desk in every office that now has a chatbot open in another tab.
Meanwhile in Hollywood
I asked the Hollywood screenwriter Jonathon Stewart — whose career runs through some of the biggest animated franchises of the last decade — what percentage of what an AI generates in a session survives into a script. His answer matched Stokes’s almost exactly: one to five percent. Ask the machine for a scene, and it might get back twenty competent variations. He’ll use one line from one of them, maybe a structural idea buried in a draft he otherwise discards. The other nineteen disappear.
If you’ve never worked inside a profession built on selection — editing, casting, acquisitions, drug discovery, screenwriting — that number can look like waste. It’s the opposite. A magazine doesn’t publish every manuscript that crosses an editor’s desk; a casting director doesn’t hire the first actor through the door. The ratio of considered versus chosen has always been brutal in any field where taste is the actual product. What AI changed isn’t the ratio. It’s that the ratio is now visible, countable, and instant, instead of spreading across months of slush piles and casting calls.
Which brings me to the real danger I see inside executive teams right now, the one nobody is naming: Most leaders are using AI to save time, then spending the saved time generating more instead of choosing more conscientiously. They run the same prompt five ways, produce five strategy decks, five ad concepts, five go-to-market plans, and call the pile “options.” But a pile of options is not a decision, and an afternoon spent admiring abundance is not the same as an afternoon spent earning a verdict.
Stokes’s team did not treat their recovered year as more search time. They treated it as more selection time. That distinction is the entire argument of this essay, and I’d wager it’s about to become one of the defining competitive advantages of the next decade: organizations that reinvest AI-driven time savings into sharper judgment will out-execute those that reinvest the same savings into producing yet more content nobody asked for.
There’s a second thing worth saying about that discarded 95 to 99 percent, because I don’t think it disappears the way we assume. Every molecule Stokes’s team rejected sharpened their eye for the ones that mattered. Every draft Stewart throws away trains his instinct for the line worth keeping. The discard pile isn’t waste. It’s the negative space that gives the chosen few percent its shape. You cannot develop taste by looking only at what you keep. You develop it by looking hard enough at everything you didn’t, until you understand precisely why it fell short.
Stokes puts it more bluntly: “AI models are simply suggestion-generation boxes. Humans — with our domain expertise and lived experience — take those suggestions, most of which aren’t great, and make a better decision than we might have made otherwise.”
What Are You Doing With the Time You Got Back?
This may be the most important thing AI has quietly done to professional expertise: it has made the act of rejection the visible, measurable center of the job, rather than the invisible part that happened before anyone saw a first draft. For decades, judgment happened off the page — the editor’s private slush pile, the scientist’s private dead ends, the years of false starts no one outside the lab ever saw. AI has dragged that hidden work into the light. We can now watch, in real time, exactly how much a professional throws away in order to keep the right thing.
That should change how we evaluate AI adoption inside a company. The wrong question is, “How much did the team generate this quarter?” That number will always climb, it will always look impressive, and it will tell you almost nothing about whether anything good happened. The right question is closer to what Stokes was describing: How much more thoroughly, and how much faster, is your team now choosing? If the answer is “We generate ten times more and ship ten times more,” you’ve built a faster machine for producing the Wall of Gray — competent, forgettable, average material at higher volume. If the answer is “We generate ten times more and ship the same amount, but it’s measurably better, because we spent the recovered time arguing about which one percent deserved to survive,” you’ve built something durable.
The crane lifts. That part is not in dispute, and Stokes would be the first to tell you his lab is faster and better for it. But the crane was never the point of the building. It cleared a year of rubble so the people in the room could spend that time deciding, with real scrutiny, what was worth constructing. Saved time reinvested in more generation is just time wasted at a faster speed. Saved time reinvested in more discerning judgment is the only version of this story that ends with an antibiotic, a finished script, or a strategy that survives contact with reality.
So, I’ll leave you with the question I left that room in Chicago: Your team has almost certainly gotten faster this year because of AI. What, exactly, are you doing with the time you got back?