AI Is Making Answers Cheap. That Makes Questions and Judgement More Valuable.

What happens to the value of knowledge work when producing a plausible answer takes seconds?

That question matters more to me than whether a piece of work is labelled “AI-generated”.

I use AI extensively: to research, test ideas, challenge assumptions, find connections, draft and edit. Often its most useful contribution is not an answer I keep. It is an objection I had not considered, a theory I did not know, a different way of framing the problem, or a question that sends me somewhere more useful.

I have come to see AI less as an answer machine and more as an inquiry technology. It lets me explore more before I decide what I think.

And that changes where the value sits.

In much of knowledge work, a lot of the effort used to sit in the middle of the process. You had a question, did the research and analysis, and produced an answer. Research took time. Analysis took time. Writing a decent report took time.

AI is making that middle much cheaper. It can generate a competent first answer, alternatives, summaries, comparisons and drafts in seconds.

But the work on either side has not changed nearly as much.

Before the answer comes the question. What are we really trying to understand? Have we framed the problem properly? What are we assuming? Is this even the question worth answering?

After the answer comes judgement. Does this make sense? What is missing? What do I believe? What should we do?

When answers become abundant, good questions and judgement become more valuable.

The important work may happen before the answer

We often describe the human role in AI as checking what the machine produces: the AI generates; the human judges. That matters, but it starts too late.

Some of the most valuable work happens before there is an answer to judge. A good strategist, researcher, adviser or executive does not simply respond well to the question in front of them. They notice when the question itself is wrong. They see that a symptom is being treated as a cause, an assumption has gone unchallenged, or two things others regard as separate may be connected.

AI can help here too. I can ask it for twenty questions in seconds. But generating questions is not the same as recognising the one that matters.

This is where I get most value from AI. I can follow an idea far enough to discover that it is weak, ask for the strongest objection to something I believe, test whether an idea that feels new is already well established elsewhere, or explore several explanations before deciding which deserves more attention.

Sometimes I keep almost nothing from the exchange. It has still been useful because the inquiry has moved.

That contribution may leave no obvious trace in the final document. The most useful thing AI did could have been to raise an objection I rejected, surface a theory that changed my framing, or expose a question that caused me to rewrite the argument.

That is one reason simple labels around AI use trouble me. They focus on the provenance of the final artefact when much of the value may have been created in the inquiry that came before it.

“AI-generated” covers too much

There is another problem with the category itself.

Using AI to generate an article from a short prompt is not the same as using it to check spelling. Nor is it the same as asking it to tighten a paragraph, challenge an argument, summarise a paper or suggest a different structure. Yet all of these can end up described as “AI-assisted” or “AI-generated”.

There is a longer history here too. Word processors changed how we compose and revise. Spellcheck removed some of the manual work. Search changed how we find information. We did not conclude that a document was somehow less ours because software had helped us produce it.

AI goes further because it can contribute ideas and language, not just mechanics. But that makes the distinction between different kinds of use more important, not less.

As AI becomes built into word processors, search, transcription, presentation software and email, asking whether a document has been touched by AI may soon be about as informative as asking whether it was produced using software.

The better question is whether AI made a material contribution to the substance of the work, and what role human judgement played in accepting, rejecting and shaping that contribution.

For me, the important distinction is between tool involvement and intellectual dependence.

Slop is what happens when production outruns judgement

AI slop is real. But AI did not invent empty management language, generic articles or confident reports that add very little. What it changed was the cost of producing them.

The generation clock has collapsed to seconds. The inquiry and discernment clocks have not.

We can now produce ten reports where we once produced one, generate fifty options instead of five, and publish every day rather than when we have something worth saying. The supply of answers can rise much faster than our capacity to decide which ones deserve attention.

Production has outrun judgement.

Producing something has become so easy that not producing it may now require more judgement than producing it.

That seems a better description of much of what we call AI slop. It is not defined by AI involvement. It is what accumulates when generating material is easier than deciding whether that material should exist.

Abundance has moved the value before

There is a historical pattern here.

Printing made reproduction cheap. The internet made distribution cheap. AI is making the production of plausible answers cheap. Each shift created abundance in one place and moved the scarce resource somewhere else.

Printing did not remove the problem of deciding what was worth reading. It made that problem larger. Over time we built a much richer trust and filtering system around abundance: publishers, editors, reviews, citations, professional institutions and reputations.

AI is likely to require its own trust architecture.

Some of that will be technical. Some will be institutional. But knowing that AI touched an artefact is only one possible signal, and often not the one that tells us what we most need to know.

What about watermarking?

There are good reasons to care about provenance.

If an image is presented as evidence of an event, I want to know whether it is synthetic. If a video appears to show someone saying something they never said, its origin matters. If an assessment is meant to test unaided capability, how the answer was produced is part of the test.

Copyright, consent and attribution raise other legitimate questions too. An illustrator whose work has been used without permission has a very different interest in provenance from mine.

The harder case is when provenance becomes a way to deal with volume. If AI allows hundreds of applications, tenders or reports to be generated cheaply, human judgement cannot inspect every one in depth. A cheap first filter is exactly the kind of thing an overloaded system will reach for.

That makes sense. But a useful filter has to correlate with what you are trying to find.

“AI was involved” often does not.

A specialist who knows a problem deeply and uses AI to draft a strong response quickly may be flagged. Someone with much less understanding who writes every word unaided may pass straight through. The filter points the wrong way in both cases.

Provenance tells us something useful about provenance. The mistake is asking it to do the work of judgement.

For knowledge work, I usually care about different questions. Is the problem well framed? Is the reasoning sound? Is the evidence good? Does this person understand what they are saying? Are they prepared to put their name to it and stand behind it?

A watermark cannot answer those questions.

Reputation is the accountability mechanism

This is where my own position is simple.

If my name is on something, I own it.

I cannot outsource responsibility to the AI I used, any more than I can blame a search engine, an editor or a spreadsheet for a bad conclusion. If I publish something that is wrong because I did not check it, that is my error. If I repeat an argument I do not understand, that is my failure. And if I use AI to produce generic material simply because I can, readers are entitled to think less of my work.

That is what reputation does.

It builds slowly through the quality of what we choose to say, whether our ideas survive challenge, whether our advice proves useful and whether we stand behind it when it matters. Every piece of work adds to that reputation or draws against it.

AI can make it easier to manufacture the appearance of expertise or a body of work. It cannot make either worth trusting simply by producing more.

For me, that is a much stronger form of accountability than claiming a document is “100% human-written”.

What matters is not whether AI touched the work, but whether I own the thinking I put my name to.

What changes for organisations?

Most organisations still approach AI mainly as a productivity technology. How can we write the report faster, produce more analysis, respond to more requests or generate more options?

Those gains are useful, but they may not be where the larger advantage lies.

If producing answers is no longer the bottleneck, accelerating it further simply moves the constraint. More analysis creates more to review. More options create more choices. More experiments create more learning to interpret.

The organisation can become very efficient at producing material that its decision system cannot absorb.

The more interesting use of AI may be as an inquiry technology: helping people explore possibilities, challenge assumptions earlier and ask better questions before they commit time and resources.

That puts a different premium on human capability — not on typing faster, or even on producing more, but on framing, judgement, responsibility and knowing when enough is enough.

If AI makes answers cheap, the advantage will not go to those who produce the most answers. It will go to those who know which questions are worth asking, which answers are worth believing, and what is not worth producing at all.