Ideas & Philosophy
We Can Make Almost Anything. That’s Becoming a Problem.
AI makes it easier to build, publish and produce. The harder question is whether we are getting better at deciding what deserves to be made.
We can now produce websites, dashboards, campaign variants, landing pages, research summaries and prototype tools at a speed that would have looked absurd a few years ago. The strange part is not that the work exists. It is how little friction stands between an idea and a finished-looking artefact.
That abundance is useful. It is also slightly unnerving. We are becoming extraordinarily good at making things. I am less convinced we are becoming equally good at knowing which of those things matter.
The illusion of productive work
Imagine a marketing team asked to grow organic traffic. With generative tools, they can draft hundreds of SEO landing pages in a week: tidy titles, plausible intros, FAQ blocks, internal links, even a sitemap entry for each one. Publishing velocity goes up. The content calendar looks busy. Someone presents a slide about output.
What may not appear on that slide is whether any of those pages answer a real search problem. Intent still has to be diagnosed. Customer language still has to be heard. Thin variations of the same pitch still confuse crawlers and people alike.
Faster production can hide the absence of a decision.
This is hypothetical, but the pattern is familiar in growth work: activity mistaken for value. A filled backlog feels like progress. A published page feels like proof. Neither guarantees that a customer problem got clearer, or that the business learned anything it could not have guessed on Monday morning.
I have watched organisations get excited about how quickly marketing assets appear, then quieter about whether those assets changed behaviour. Speed is easy to celebrate. Judgement is harder to put on a dashboard.
The productivity numbers are real
Noy and Zhang ran a preregistered experiment on mid-level professional writing tasks. With ChatGPT, average time fell by about forty per cent and independently graded quality rose by about eighteen per cent. Inequality between workers narrowed. The authors also note the limits: short, generic tasks without deep company context or hard factual constraints. Real jobs are messier.
So the productivity story is not empty. People can finish certain kinds of work faster, and sometimes better, with these tools. What the findings do not establish is that faster competent output is the same as better strategy, clearer customer value or more meaningful work. They measure production under defined tasks. They do not tell a team whether the task was worth doing.
When everything looks competent
Competence used to be expensive. That expense acted, imperfectly, as a filter. When the cost of looking finished drops, finished-looking work becomes a weaker signal. Doshi and Hauser found that writers given generative-AI story ideas produced work judged more creative, better written and more enjoyable, especially among less creative writers. Collectively, though, AI-assisted stories were more similar to one another than stories written without AI. Individual upside; narrower variety across the set.
That social dilemma travels well beyond short stories. Marketing teams can generate campaign copy that sounds professional. Product teams can spin up interfaces that look intentional. Analysts can produce decks that survive a glance. If everyone’s output moves toward the same polished centre, competence stops distinguishing much. The scarce thing is not another plausible paragraph. It is a point of view that survives contact with evidence.
Homogenisation is not destiny. It is a tendency. The research does not say originality disappears. It says that without deliberate pressure toward difference, assisted work can cluster. In a market already crowded with near-identical landing pages and interchangeable thought leadership, that clustering is a practical problem, not an aesthetic complaint.
The economics of getting it wrong
When production gets cheap, it is tempting to treat mistakes as cheap too. They are not. Opportunity cost remains. Attention remains scarce. Customer trust remains expensive to repair. A wrong URL architecture, a misleading promise, a product feature nobody asked for: each still consumes calendar time, organisational focus and, often, goodwill you cannot regenerate with a prompt.
Falling unit costs for drafts do not cancel the cost of choosing the wrong problem. If anything, they can increase it, because you can now scale a weak diagnosis. Ten weak pages are an annoyance. A few hundred weak pages are a system: more to maintain, more to redirect, more noise in analytics, more ways for search engines and customers to get a confused picture of what you offer.
In technical growth work, I spend a lot of time on systems that already exist: indexing behaviour, measurement gaps, migration residue, content that was published because someone could publish it. The bill for that residue is rarely itemised as “we made too much.” It shows up as unexplained flatlines, conflicting signals and meetings about why the site will not behave.
Who is doing the thinking?
There is a separate worry that AI softens critical thinking. The careful version of that worry does not claim machines hollow out minds on contact. Lee and colleagues, surveying knowledge workers about generative AI at work, found that higher confidence in AI for a task was associated with less reported critical thinking, while higher confidence in one’s own ability was associated with more. People also often reported that AI reduced the effort of cognitive activities tied to critical thinking. The study is correlational and self-reported. It does not prove AI causes decline. It does suggest a pattern worth taking seriously: when the tool feels trustworthy, oversight can thin out.
The same tools can also sharpen investigation. They can draft counter-arguments, surface alternative framings, summarise dense material and help you pressure-test a hypothesis before you spend a week on it. The difference is not the model. It is whether you are using it to finish faster or to think more clearly about what finishing would mean.
I am not interested in nostalgia for slow work. I am interested in who still owns the question. If the answer arrives before the question has been examined, the organisation has not become more intelligent. It has become more fluent.
What becomes valuable now?
As competent production gets cheaper, value concentrates around work that decides what production is for. That includes noticing the real problem, not the requested deliverable. It includes listening to customers closely enough to hear the gap between what they say and what they do. It includes systems thinking: how a change in content, measurement, indexing or sales motion will behave together, not in isolation.
It also includes experimental discipline. A page is a guess until something measurable happens. Evaluation is not a vibe. Strategic judgement is the unglamorous habit of saying no to plausible work because the evidence is thin, the audience is imaginary or the metric would move for the wrong reason.
None of these abilities are magically human-only, and none are safe from automation forever. Treating them as sacred is a good way to stop improving them. Treating production speed as the main skill is a good way to drown in artefacts. The practical stance is narrower: keep the expensive judgement close to the cheap generator, and refuse to confuse the two.
The question worth keeping
We can make almost anything. The interesting constraint is no longer whether we can ship. It is whether anyone needed the thing we shipped, and whether we would know if they did not.
Before the next batch of pages, features or campaigns leaves the building, three questions still earn their keep. What problem disappears if this works? Who benefits in a way we can observe, not just assert? What evidence would tell us this was worth doing, and what would tell us to stop?
The tools will keep getting better at sounding sure. The work, if we are honest, is staying interested in what the certainty is standing on.
Research and further reading
The studies above are worth reading in full, especially their limits sections. For productivity under real work conditions, start with Brynjolfsson, Li and Raymond and Noy and Zhang. For creative output and collective similarity, see Doshi and Hauser. For how knowledge workers report thinking with generative tools, see Lee et al. The interpretations in this note are mine; the measurements are theirs.
References /
- Brynjolfsson, Li and Raymond, Generative AI at Work (Quarterly Journal of Economics)
- Noy and Zhang, Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence (Science)
- Doshi and Hauser, Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content (Science Advances)
- Lee et al., The Impact of Generative AI on Critical Thinking (CHI Conference on Human Factors in Computing Systems)