Let the Machine Ruin the Blank Page

We keep measuring AI by how long it can work after we leave the room. Can it turn a goal into tasks, call the right tools, recover from mistakes and deliver a finished result?

Those are fair engineering questions, but they have narrowed the conversation. Most people do not spend their day waiting for a fully autonomous system. They spend twenty minutes avoiding a document they know they need to open.

Next to the promise of autonomous agents, that looks trivial. In an actual workday, it is where a surprising amount of useful effort disappears.

The blank page is not neutral

Vincent van Gogh knew the problem. In an 1884 letter to his brother Theo, he wrote:

“You don’t know how paralyzing it is, that stare from a blank canvas.”

The untouched surface was not an invitation. It was an accusation. It asked the painter to prove, before making a mark, that the painting deserved to exist.

Painters sometimes deal with this by “killing the white”: putting down a first mark simply to end the tyranny of the pristine surface. The mark does not have to be good. It gives the next mark something to answer.

Anne Lamott made the same case for bad first drafts in Bird by Bird. Twyla Tharp begins her account of creative discipline in The Creative Habit with a mundane ritual: she gets in a taxi to go to the gym. The cab ride is not the creative work, but it gets her past the moment when she might negotiate herself out of doing it.

The shared insight is procedural rather than inspirational. Make the first action too small to deserve a debate.

This is one place where generative AI is genuinely useful. Not as an oracle and not as an employee in a browser tab. It can put an object on the table.

Ask for a rough outline. Ask for three openings you are likely to hate. Ask for the obvious argument so you can see where it fails. The result should be disposable. Its job is to give you something specific to resist.

Once text exists, the mind behaves differently. You notice the missing premise. You bristle at a phrase you would never use. You remember the example that actually proves the point. Disagreement becomes productive.

What happens after the first draft

Ethan Mollick called the coming AI feature inside everyday software “The Button.” Faced with a blank page, people will push it because beginning with something is easier than beginning with nothing.

He also pointed to the trap. The first suggestion anchors what follows. Even when we rewrite every sentence, the initial structure can quietly decide which questions get asked and which never appear. A fluent draft may save ten minutes while making the eventual idea more conventional.

We do not get to borrow Mollick’s case for pressing the button and ignore his warning about what happens next.

The practical answer is to keep the generated material provisional. Ask for competing structures rather than one polished draft. Make the human author state what each version gets wrong before choosing a direction. Verify claims outside the model. If nobody can explain why the final piece says what it says, nobody should put a name on it.

That last standard matters because AI writing is cheap enough to create its own weather. There is more text, more often, aimed at more people who did not ask for it. “AI slop” is an ugly phrase for an ugly experience.

Clay Shirky saw an earlier version of this problem in Cognitive Surplus. He argued that even the “stupidest possible creative act” is still an act of creation. When the cost of participation falls, mediocre work multiplies, but so does experimentation.

Shirky’s point is useful, though it is easy to romanticize. Not every bad AI draft is a brave experiment. Some of it is spam. Some of it is a cheap substitute for attention. Lowering the cost of making things does not increase the value of everything made.

The old bargain was bad too. Good ideas routinely died because the person who had them was not quick with an opening paragraph, comfortable in front of a slide deck or willing to endure a clumsy first attempt. Generative AI lowers that barrier while raising the noise floor.

Cheap output makes judgment visible

Leaders can get further by mapping where capable work gets stuck and what would count as evidence that a person moved it forward intelligently. This is less glamorous than an automation roadmap and usually more useful.

A sensible workflow starts with the person defining the audience, the decision and the constraints in their own words. AI can then produce alternatives. The person has to reject, combine or redirect them, and remain accountable for every factual claim in the result. That is slower than pressing a button and forwarding the output. It is faster than worshipping the blank page.

This is close to our work at Aramis because cybersecurity skill verification has the same underlying problem. A polished response is no longer proof that someone knows what to do. In truth, it never was, but AI makes the mismatch impossible to ignore.

If a candidate can summon a credible answer in seconds, an assessment has to reveal something harder to manufacture: how the person handles ambiguity, tests an assumption, spots a dangerous omission and responds when the tool is confidently wrong.

AI leaves human skill in place and makes our old proxies for it brittle.

The blank page was one of those proxies. We treated the ability to produce a clean first draft from scratch as evidence of thought. Sometimes it was merely evidence that someone was good at first drafts.

Let the machine spoil the surface, then pay attention to what the person does next. That is where the skill becomes visible.

Editorial Source Notes

Vincent van Gogh, letter 464 to Theo van Gogh, October 2, 1884

Anne Lamott, Bird by Bird (1994), “Shitty First Drafts”

Twyla Tharp, The Creative Habit (2003)

Clay Shirky, Cognitive Surplus (2010)

Ethan Mollick, “Setting Time on Fire and the Temptation of The Button” (2023)