AI Helps Us Produce More. When More Is Too Much.

Diagram showing 50 pages reduced to 5 pages and then to 1 decision.

AI Helps Us Produce More. Who Is Going to Read It?

What happens when our capacity to produce material grows exponentially, but our capacity to understand meaning remains stubbornly human?

I have been thinking about this during our product-planning work at Pieoneers. AI can help us research a market, compare competitors and explore technical options in a fraction of the time this once required. It is useful work. It can also produce 50 pages before anyone has decided which five pages matter.

Eventually, our team sits down together. Discuss what’s important to the client. The report gets shorter. Not because most of it was wrong, but because a person still has to make a decision.

AI makes producing information easier. It does not automatically make information easier to understand.

The bottleneck has moved

Producing information used to be expensive. Writing a book took years. Research required interviewing people, finding documents, and assembling an argument. Writing software meant writing the code. That effort never guaranteed quality, but it limited how much human engineers could produce.

AI removes that limit. We can generate ten product concepts instead of three, ask a report to cover every possible consideration and produce an implementation in seconds. But someone still has to choose among the concepts, read the report and maintain the code.

Production is getting faster. Human attention is not.

We have seen this before

Search engines made the internet easier to navigate. Then publishers learned what search engines rewarded.

Publishing adapted to the algorithm. Simple questions acquired 2,000-word answers, and the same information appeared on hundreds of sites, rewritten just enough to look new.

Generative AI did not create those incentives. It made it much cheaper to respond to them.

Ed Zitron describes this forcefully in his conversation on The Diary of a CEO. Before AI slop, he says, there was SEO slop: material created to rank rather than to be useful to a person. In his words, generative AI gave us a “slop machine” capable of serving them at far greater scale.

Bad material is easy to dismiss. The harder problem is an abundance of material that is reasonably good. A useful 50-page report still takes time to read. Ten plausible product ideas still require judgment.

Software makes the problem visible

AI can now produce a great deal of code quickly. Yet the amount of code written has never been the same as the value of the software.

Does the code solve the right problem? Can another engineer understand it? What happens when it fails?

Sometimes the best engineering decision is to replace 1,000 generated lines with 200 simpler ones. Sometimes it is to avoid writing the code at all.

A May 2026 NBER working paper followed more than 100,000 GitHub developers using three generations of tools: autocomplete, interactive agents and autonomous agents.

The newer tools produced striking gains at the coding stage. With autonomous agents, commit activity rose by 180%. But the effect became smaller as the work moved toward something a customer could use. The number of projects rose by about 50%, while releases rose by 30%. Across four large app marketplaces, the researchers found more new software but no increase in total usage.

AI can generate code, but software still has to be reviewed, integrated, tested, released and adopted. A 30% increase in releases is meaningful. But it also shows that writing code and shipping valuable software are different measures.

The discipline of one page

One client taught me this lesson before generative AI made it so easy to produce too much. Jerry Kroll, CEO of Jevitty Life Science, always asked us to keep proposals to one page. At first, I treated it as one of his personal quirks. Surely some proposals needed more room.

Olena Tkhorovska, CEO of Pieoneers, meeting with Jerry Kroll, CEO of Jevitty Life Science, whose insistence on one-page proposals became an enduring lesson in prioritization.

But his rule forced us to make choices. What did he actually need to decide? Which details changed that decision, and which ones merely demonstrated how much work we had done?

One page may be too ruthless for every situation. But the discipline has stayed with me: keep what is necessary and be willing to remove the rest.

Reduction is not the same as deleting context. The goal is to preserve what another person needs to understand the decision.

Who chooses what reaches us?

The same problem appears online. When nobody can consume everything being produced, algorithms decide what reaches us. Research into popularity bias in recommender systems shows how recommendations can favour material that is already popular, making lesser-known ideas harder to discover.

As AI expands the pool of available material, those filters may become more influential. The problem is no longer simply finding information. It is deciding what deserves our limited attention.

AI can also help us reduce

AI can compare hundreds of documents and show us where they disagree. It can find duplication in a report, explain an unfamiliar codebase or turn a long meeting into a short list of unresolved decisions. Some of its greatest value may come not from helping us produce more, but from helping us understand what has already been produced.

But when AI compresses 50 pages into five, it decides what to leave out. The strange observation on page 37 may be irrelevant, or it may be the detail that leads someone to a new idea.

Compression always loses something. We need to know when that trade is worthwhile and when to return to the source.

AI is making production abundant. That may make attention, judgment and curiosity more valuable.

A few things worth reading and listening to

These sources do not “prove” the argument above. They are useful because they approach the question from different directions.

Ed Zitron on The Diary of a CEO — A skeptical and provocative conversation about generative AI, its economics and the growing amount of AI-generated “slop.”

Demirer, Musolff and Yang, Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools — A May 2026 NBER working paper using data from more than 100,000 GitHub developers. It finds that newer AI tools sharply increase coding activity, but much of that gain narrows before the software reaches projects, releases and users.

Klimashevskaia et al., A Survey on Popularity Bias in Recommender Systems — An open-access academic survey of how recommendation systems can favor popular items and limit the visibility of less-popular material in the long tail.