AI Removes the Punishment for Iteration

Most arguments about AI and creativity get stuck on the wrong question: Can the machine create?

Yes. Obviously. It can make novel combinations, produce surprising work, and occasionally land on something beautiful. Pretending otherwise does not protect human creativity. It just makes the rest of the argument less honest.

The more useful question is what happens when the cost of trying again collapses.

For most of history, a creative decision came with a tax. Another draft took another night. Another product direction meant another design cycle. Another film sequence needed more people, equipment, and money. Another arrangement meant calling everyone back into a room. Even a small change could travel through four people, three tools, a calendar, and a budget before it became real.

That tax did not just make work expensive. It changed the work.

People accepted the first decent answer because the second answer was costly. They kept a weak section because rebuilding it would create a week of meetings. They stopped exploring because a collaborator had already done them a favor. They shipped “good enough” because their money, attention, political capital, or stamina was gone.

Some of that friction was useful. A deadline can force a decision. A talented collaborator can challenge your instincts. A budget can prevent self-indulgence. But a lot of it was just punishment for wanting to see one more version.

Generative AI removes a large part of that punishment.

Not the work. Not the need for skill. Not the consequences of being wrong. The punishment for trying again.

That distinction predicts both more AI garbage and the exceptional work the same tools can produce. They are not contradictory outcomes. They come from the same change in the economics.

Of course there is more garbage

When producing another page, image, interface, song, analysis, or campaign gets cheap, people produce more of them. Most attempts at anything are mediocre. So a system that allows vastly more attempts will create vastly more mediocre output.

This should not be surprising.

The first draft a model gives you is usually a statistically polished guess at what something like your request should look like. It can be clean, coherent, and completely dead. It has the shape of an answer without a reason to exist. If you publish that because it arrived in twelve seconds and sounds professional, you are not practicing some new form of creativity. You are operating a vending machine.

Companies will use these machines to fill every available surface. Search results, inboxes, social feeds, app stores, sales decks, support centers, and internal knowledge bases will be flooded with material nobody cared enough to inspect. Some of it will be harmless. Some of it will be false. A lot of it will be plausible enough to waste someone else’s time.

That is AI slop. It is real. I hate it too.

But counting the visible garbage and stopping there misses the more interesting side of the equation. The work killed by friction has always been invisible. The book that was never drafted, the product idea that never survived the cost of a prototype, the strange visual direction that never got past a committee, the song that never justified studio time—none of those show up in an inventory of what the old system lost. They just never existed.

AI does not guarantee those things will be good. It gives them a chance to exist.

One 2024 Science Advances experiment by Anil Doshi and Oliver Hauser found this tension in miniature: access to GPT-4 story ideas led evaluators to rate individual short stories as more creative, better written, and more enjoyable, while the AI-assisted stories became more similar to one another. That is one controlled writing task, not a law for all creative work. The gain was also largest for the less creative writers, which is worth sitting with. It is still useful evidence: better individual attempts and less collective variety can arrive together.

I do not know whether the percentage of good work goes up. It may go down. I do think the absolute number of serious attempts goes up, because more people can afford to attempt them and the same person can explore more directions before running out of resources.

The first answer is almost never the point

The useful unit of AI work is not the prompt. It is the loop.

Intent, attempt, inspection, correction, another attempt. Keep what worked. Kill what did not. Change one assumption. Attack the strongest version. Put it in front of reality. Repeat.

This is why “prompt engineering” has always felt too small to me. A clever sentence can improve an answer, but the larger skill is process design. What do you ask the machine to generate? What evidence does it get? How do you compare the outputs? When do you bring in a second model or another person to challenge the work? Which claims need to be checked against a primary source? What has to be tested with a customer, a browser, an instrument, a spreadsheet, or the physical world before you believe it?

I understood the loop most clearly while making music with generative tools. I could try a direction, hear what was wrong, keep one useful choice, and throw the rest away without turning every experiment into a budget decision or a favor. Most versions were disposable. Their job was to make the next decision clearer.

The same loop runs when the thing being iterated is a business process instead of a song. Cheap iteration does not magically reveal the correct workflow. It makes the gap visible. The first proposed automation may misunderstand where the judgment lives, automate a symptom instead of the repeated work, or save clicks while creating a new review burden. Another pass is cheap. Deploying a bad automation across real work is not. Each pass earns evidence the first prompt did not have.

That is the operating model:

1. State what the thing is supposed to do in terms strong enough to reject a polished failure.
2. Generate materially different attempts, not twenty cosmetic variations of the same bad idea.
3. Compare them against the intent, the evidence, and the real constraint.
4. Preserve the strongest parts without letting them become sacred.
5. Let an independent reviewer attack the result.
6. Test the surviving version somewhere the model cannot flatter you.

The last part matters. A model can tell you an onboarding flow is clear. A confused user can prove that it is not. A model can tell you a business case is persuasive. A buyer can ignore it. A model can tell you the code works. The system can fail. Generation is cheap. Reality still charges full price.

Taste is not a magic human aura

People talk about taste as if it is a mystical substance artists secrete onto finished work. That is not useful.

Taste is the ability to recognize the distance between what exists and what should exist. Intent is knowing what the work is for. Judgment is deciding which tradeoff matters now. Accountability is owning what happens after the decision.

Those are different things.

A model can help with all of them. It can compare versions, surface contradictions, predict objections, and find patterns a person missed. Other models can judge the first model. A good system can eliminate an enormous amount of weak material before a human sees it.

But somebody still has to set the objective. Somebody has to decide that the surprising failure is more important than the conventional success. Somebody has to notice that the technically better sentence changes the actual argument. Somebody has to say, “No, that is not what I mean,” even when the substitute sounds smarter. Somebody has to own the result when it reaches a customer, an employee, a court, or the public.

That is where taste stops being vibes. It becomes a control system.

The more output AI can generate, the more valuable that control system becomes. When production is scarce, execution itself carries value. When production is abundant, selection carries more of it. The bottleneck moves from making an acceptable thing to identifying, defending, and finishing the right thing.

This is also why AI can make an experienced operator look dramatically better and an undisciplined one dramatically worse. The first person uses cheap iteration to test assumptions and get closer to the truth. The second uses it to manufacture infinite validation for whatever they already wanted to believe.

Both can generate forty answers. Only one is willing to discard thirty-nine.

The machine can do the work

Sometimes the machine is doing the work. A model writing a usable function, producing an illustration, restructuring an argument, or turning raw material into a coherent first pass is doing work in every practical sense. Calling it a calculator does not change the economics for the person whose task disappeared.

When one person can complete work that used to move through a team, some people gain leverage and other people lose work. The creative upside and the labor damage can be true at the same time. Lower production cost is not an argument that the transition is fair.

Who gets the capacity that automation creates is a different fight, and I have written about it separately.

Cheap iteration still has limits

“AI removes the punishment for iteration” is a useful mental model. It is not a law of nature, and it does not apply evenly.

It applies most strongly where the work can be represented and evaluated digitally. Words, code, images, audio, plans, analyses, interfaces, simulations. It applies less cleanly when each attempt consumes physical materials, risks a human body, moves money, changes a legal right, or alters a production system. You can generate a hundred bridge designs cheaply. Building the wrong bridge is still expensive.

And it does not remove evaluation cost. When possible outputs become infinite, attention becomes scarce. Somebody has to review the work, compare it, verify it, and stop. Without a clear objective, cheap options become decision paralysis: endless variations that confuse motion with progress.

It does not make access equal. A cheap tool lets more people enter. It does not give everyone time, distribution, money, safety, education, or an audience.

And it definitely does not make the output true. A cheap iteration of a false claim is still false. A beautifully written strategy built on invented evidence is worse than a rough one, because it is easier to trust. The closer the machine gets to fluent, the more important verification becomes.

The penalty has moved. It used to sit heavily in production. Increasingly it sits in judgment, verification, and consequences.

Make more attempts. Raise the bar.

The lazy response to generative AI is to accept more because making more is easy.

The serious response is the opposite: reject more because trying again is cheap.

Do not confuse polish with thought. Do not let the model that generated the work be the only place its rules and verification live. Do not hide behind the machine when the result causes harm.

Use the new economics properly. Try the direction that would have been too expensive. Build the ugly prototype and test the objection before somebody else raises it. Throw away the thing that took six minutes with the same seriousness you once had to throw away the thing that took six weeks.

AI will produce an ocean of competent garbage. It will also let specific people with specific intent make things that the old economics filtered out before they were born.

The difference will not be who had access to the generate button.

It will be who knew why they were pressing it, who could tell when the answer was wrong, and who cared enough to try again.