Polish Is No Longer Proof

AI can make the work better without necessarily making us better at the work.

For a long time, we could look at the work and make a reasonable inference about the person who created it.

A thoughtful strategy suggested thoughtful reasoning.
A sharp presentation suggested command of the material.
A strong research synthesis suggested someone had wrestled with the evidence.
A polished portfolio suggested craft.

Not perfectly. But reliably enough. AI is making that inference much less safe.

Shannon McKeen describes a revealing example from his classroom. Students who used AI produced cleaner, more impressive-looking problem structures than many of those who worked through them themselves. But when they later had to explain and defend the work, the pattern changed: some of the messier work was backed by deeper understanding.

AI had improved the artifact without necessarily improving the thinking behind it. That distinction matters far beyond school.

In product development, Stephen Wunker describes a related shift: AI has dramatically increased our ability to generate analyses, prototypes and plausible solutions. Capacity used to impose a useful constraint. Now even weak ideas can arrive looking surprisingly polished.

The bottleneck moves. Producing possibilities becomes easier.

Deciding which possibilities deserve our attention becomes harder. And that raises a question I think we'll have to ask ourselves more often:

When the work looks good, how do we know the thinking behind it is good too?

Polish has become a weaker signal

This creates an interesting problem for many of the systems we've built around knowledge work. We hire using portfolios, writing samples and take-home exercises. We evaluate people through presentations and deliverables. We often infer expertise from the quality of what someone produces. Those artifacts haven't suddenly become meaningless. But they may no longer tell us as much as they once did.

A beautiful presentation can contain reasoning its presenter hasn't fully examined.

A convincing recommendation can include assumptions its author hasn't challenged.

A clean research synthesis can hide contradictions that deserved more attention.

AI doesn't make any of those outcomes inevitable. It simply makes them easier to produce.

Which means the questions around the artifact become more important.

Why did you choose this direction?

What evidence mattered most?

What did you reject?

Where are you least confident?

What would cause you to change your mind?

What happens if an important assumption turns out to be wrong?

The goal isn't to prove that you did the work without AI. I don't think that's a particularly useful standard.

The better question is whether you still understand the work well enough to think beyond what was produced.

Some friction is doing useful work

This also complicates one of AI's biggest promises: removing friction. Usually, that sounds obviously desirable. Why spend three hours doing something AI can help you finish in thirty minutes?

Sometimes that's exactly the right question. But I think we're beginning to lump together two very different kinds of friction.

There is friction that mostly gets in our way:

Formatting.
Transcription.
Searching.
Repetitive transformation.
Administrative work.

Remove as much of that as you reasonably can.

Then there is another kind.

Forming an opinion before seeing someone else's answer.

Struggling with conflicting evidence.

Choosing between two imperfect options.

Trying to explain something, realizing your explanation isn't very good, and trying again.

Sitting with a question before rushing toward a conclusion.

That friction can feel inefficient too, but the inefficiency may be part of what develops the capability.

A recent World Economic Forum piece makes a similar argument about education and work: judgment and critical thinking develop through doing the work, including portions of the work we might increasingly be tempted to automate away.

So perhaps the challenge isn't deciding between AI and no AI.

It's learning:

Which friction should I remove, and which friction should I protect?

The messy edges may be the work

This becomes especially visible in fields where interpretation matters. One recent research paper looked at AI-assisted sensemaking of qualitative UX data. It's a small case study, so I wouldn't treat the results as universal. But the failure mode it highlights is worth paying attention to.

AI can efficiently organize qualitative information while smoothing unusual responses and contradictions into clean categories. Yet those awkward contradictions are often exactly where an experienced researcher becomes curious.

Why did this person behave differently?

Why do what people say and what they do conflict?

What circumstance could make two apparently contradictory responses both make sense?

Why doesn't this observation fit our emerging theme?

Good synthesis creates order. Good judgment also knows when not to tidy something away. That's an important distinction. AI is very good at reducing mess. Human insight sometimes begins by noticing that the mess means something.

Faster is potential value, not value

There is a similar tension at the organizational level. Companies continue to report significant individual productivity gains from AI, while translating those gains into business results has proven less straightforward. That shouldn't be particularly surprising.

If something that took four hours now takes one, AI has created three hours of capacity. It hasn't decided what those three hours are worth.

Do we use them to investigate the problem more deeply?

Talk to a customer?

Test another assumption?

Explore an alternative?

Improve the decision?

Or do we simply create three more things?

Time saved is potential value. What we do with the saved time determines the value.

The same principle applies personally: Suppose AI helps me write a report in 20 minutes that previously took two hours. That's useful, but speed isn't the only outcome worth measuring.

I can also ask:

Do I understand this better?

Did I notice something I wouldn't otherwise have noticed?

Did I make a choice rather than simply accept one?

Can I explain the reasoning?

Can I respond when someone challenges it?

Could I take the thinking somewhere new without asking AI to do that part too?

These aren't arguments against using AI. They're arguments for being more deliberate about what we expect to retain for ourselves. Because increasingly, producing impressive work will be easy. Knowing what deserves to be produced, what should be questioned, what doesn't quite fit, and what to do next?

That still takes practice.

---

Read

AI Made Expertise Cheap. Judgment Is The New Competitive Advantage
A useful challenge to the assumption that polished work reliably signals competence. The interesting question isn't whether AI helped create the answer, but whether the person can still interrogate and defend it.

AI Has Made Judgment The New Product Management Bottleneck
When generating and building become dramatically easier, deciding what deserves to exist becomes more important. Abundance doesn't eliminate judgment. It increases the need for it.

Which Skills Will Help People Most in an AI-Driven Future?
A useful argument for distinguishing efficiency from development. Some capabilities are strengthened through portions of the work we'd most like to skip.

Between Algorithm (AI) and Intuition (Human)
A small but provocative UX research case study about what can disappear when messy qualitative evidence gets organized too neatly.

The AI Productivity Paradox Leaders Need To Solve
A reminder that individual efficiency and organizational value are not the same thing. Saving time creates capacity. Humans still decide what that capacity is for.

---

Prompt differently

One of the easiest ways to preserve your own thinking is to change when you invite AI into it.

Before asking for an answer

I’m going to give you a problem. Do not solve it yet. First ask me what I currently think, what evidence I'm relying on, and what assumptions I'm making. Once I've answered, challenge the weakest part of my reasoning.

After receiving a recommendation

Don't improve this recommendation yet. Act as a skeptical reviewer and ask me questions one at a time that test whether I understand the reasoning, assumptions and trade-offs behind it.

When synthesizing research

Help me organize these observations, but preserve contradictions, outliers and evidence that doesn't fit the dominant themes. Surface those separately before proposing any interpretation.

Before making a decision

Here is the decision I'm leaning toward. Give me the strongest reasonable case against it. Then identify what evidence, if true, should cause me to reconsider my position.

---

Practice

1. Think before you prompt

Pick one meaningful task this week.

Before opening AI, spend ten minutes writing your own rough answer, hypothesis or structure.

It doesn't need to be good.

Then use AI.

Afterward, look at what changed.

Don't just ask, Which version is better?

Ask, Why is it better?

If you can't answer that yourself, keep working.

2. Close the screen

Take something important you created with AI assistance.

Close the document and the AI conversation.

Then explain aloud:

What are you recommending?

Why?

What evidence supports it?

What alternative did you reject?

Where might you be wrong?

Notice where your explanation gets fuzzy.

That's probably where more thinking is required.

3. Preserve one useful struggle

Find one step in your work that AI could make easier.

Then deliberately don't automate it this week.

Not because doing things the hard way is virtuous.

Choose something where the struggle itself exercises a capability you care about: forming a point of view, interpreting evidence, making a trade-off, explaining an idea, or deciding what matters.

See what you notice when efficiency isn't the only goal.

The objective isn't to resist AI. It's to get better at recognizing which parts of the work you still need to practice yourself.