AI Doesn't Level the Playing Field — It Tilts It Toward People Who Already Know Things

The most-discussed engineering essay this week argues that LLMs reward expertise rather than replace it. If you run a small team, that inverts most of the AI advice you’ve been getting for two years.

There’s a comforting story about AI tools, and it’s the one that got most small teams to try them in the first place. The story goes: AI flattens the skill curve. Anyone can code now. Anyone can write. Anyone can analyze. Give everyone a seat, and watch the whole team level up together. The junior gets to punch above their weight, the senior gets to move faster, and the difference between the two narrows because the tool does the hard parts for both.

The story is wrong, or at least incomplete in a way that quietly matters. The best evidence for how these tools actually get used in practice — evidence that keeps piling up faster than the marketing narrative can absorb it — points in the opposite direction. LLMs don’t flatten the skill curve; they steepen it. The person who already knows what good looks like gets dramatically more from the same model than the person who doesn’t. The gap between skilled and unskilled users isn’t closing. It’s widening. And that has real implications for how you deploy AI across a small team, which is what most of this piece is about.

The framing crystallized this week in an essay by Sean Goedecke titled “LLMs reward expertise,” which climbed to the top of Hacker News on August 3 with more than a thousand points and a comment section full of practitioners describing the same pattern from their own work. It’s worth reading. But the argument is bigger than one essay; it’s the direction a growing body of practitioner experience has been pointing for a while, and it’s now getting named.

The evidence, briefly

Three lines of evidence are worth naming, because if you’re going to change how you deploy AI on your team, you should know what you’re changing it in response to.

The “LLMs reward expertise” argument. The core claim is mechanical, not mystical. Better prompts come from knowing what good output looks like. If you can specify the answer precisely, the model produces it. If you can’t, the model gives you a plausible-looking answer and you can’t tell the difference. The bottleneck in every real workflow is verification: can you look at what the model produced and know whether it’s right? Experts verify quickly and cheaply. Non-experts verify slowly, expensively, or not at all. That single asymmetry — verification cost — determines how much value a given person extracts from a given tool.

The “cognitive debt” counter-current. In the same week, a piece titled “Prevent cognitive debt by manually retyping LLM-generated code” hit 502 points on Hacker News, arguing that developers who accept AI-generated code without engaging with it are losing their grip on their own codebases. The interesting thing about that piece isn’t the specific prescription — retyping code by hand is more ritual than solution — it’s the phenomenon it names. Practitioners are actively noticing that people who lean on the tool without the underlying skill lose ground over time, not gain it. That’s the exact opposite of the “level the playing field” story. The debt accumulates for non-experts even as the tool gets better.

The productivity distribution isn’t flattening. Across the informal, plural, hard-to-collate reports of who’s actually shipping work with AI, the shape isn’t “everyone got 20% better.” It’s much more skewed: a small group is extracting 3–5× productivity gains on tasks they already knew how to do, and a larger group is producing the same amount of work as before or slightly less, but with more effort spent on cleanup and debugging. If you squint you can see the same shape in almost every company’s internal usage data that has leaked publicly. The tools are working. They’re just working differently for different skill levels.

Why LLMs reward expertise mechanically

The mechanism is worth understanding, because it explains why this pattern isn’t going to reverse when the models get better — it’s going to intensify.

Prompt quality is downstream of output quality. The best prompts describe the desired output precisely: format, tone, constraints, edge cases, what to do when uncertain. That description is only possible if you already know what a good answer looks like. Ask an expert to prompt for legal contract analysis and you get a prompt that names the specific things a good contract review flags. Ask a non-expert and you get “review this contract for issues.” The expert’s prompt produces expert-quality output. The non-expert’s prompt produces confident-sounding text that may or may not be correct, and the non-expert has no way to tell.

Verification is the bottleneck, not generation. Modern models can generate thousands of words of plausible content per second. The rate-limiting step isn’t generation; it’s whether a human can look at that content and know if it’s right. An expert catches subtle errors instantly and moves on. A non-expert either misses the errors or spends longer verifying than the whole task would have taken without AI. Speed of verification is a compounding advantage — the expert gets ten passes through the loop in the time the non-expert gets one.

Better models raise the stakes of being wrong. Older models produced obvious slop. You could tell it was AI. You could tell it was wrong. Current models produce output that reads as authoritative and correct even when it isn’t. This is much worse for non-experts than the older, worse models were — because the false-confidence rate is higher and the verification burden is heavier. The tool got better; the safety margin for the user who can’t verify got smaller.

Taste is downstream of exposure. The judgment about what makes an output “good” — what tone to use, what to include, what to cut, what edge case to handle — is developed through years of seeing many examples and getting corrective feedback on your own. There’s no shortcut. The AI can produce a hundred outputs, but you still need taste to pick the right one, and taste is what expertise mostly is. Anyone can generate options. Only some people can choose well.

Put those four together and the picture is clear: better models are not going to democratize expertise. They are going to make expertise more valuable. The person who used to be the best writer on your team is now the best AI-augmented writer on your team, and the gap between them and the next person just got bigger, because their taste, verification speed, and specification precision all compound on top of the tool.

The small-shop implication

This is where the strategic advice for small businesses turns nearly all the way around from what the AI vendors have been telling you.

The vendor pitch has been: give every one of your people a seat. Democratize the leverage. Watch productivity rise across the board. This makes sense as vendor economics — it’s how you sell more seats — but as a strategy for a small operation with limited attention and limited budget, it’s exactly wrong. It spreads the tool thinnest across the people who benefit from it least, and it under-invests in the people who could turn the tool into an actual competitive advantage.

The strategy that actually fits the evidence: concentrate AI leverage on your most skilled person and invest in raising the skill of your next-most-skilled person.

Concretely, that looks like this. Identify the person on your team whose judgment you trust most on the highest-leverage work — writing that clients see, decisions that shape the operation, code that ships to production, analysis that guides strategy. That person is your AI multiplier target. Give them the best tools, the most training time, the explicit mandate to spend part of their week experimenting with new workflows, and — this matters — the authority to standardize the workflows that work well so the rest of the team benefits from their discoveries.

Then look at your next-most-skilled person and ask a different question: what would it take to raise their taste and their verification skills in the areas where you need more capacity? That’s a training investment, not a tools investment. Send them to a course. Pair them with your top person for a quarter. Have them review the top person’s AI-augmented work before it ships and explain why it’s good — that’s how taste gets built, and it happens surprisingly fast when you make it deliberate.

For everyone else on your team, the AI question is specific and narrow: what one or two tasks in their job can be genuinely improved by a well-designed, well-verified AI workflow that your top person built? Those workflows go into your standard operating procedure with clear inputs, clear outputs, and clear guardrails. Nobody on that layer needs a general-purpose AI subscription. They need to know the three specific places their day gets easier and how to run them safely.

This is the opposite of “democratize” and I’m suggesting it deliberately. Broad, shallow AI adoption across a team of mixed skill levels is expensive, produces mediocre results, and lowers your average quality because the failures of the less-skilled users get shipped alongside the wins of the more-skilled ones. Narrow, deep AI adoption concentrated on the person who can extract the most from it — and formalized as workflows for the rest — produces disproportionate wins with less friction.

A concrete 30-day plan to level up one operator

If the strategy above resonates, here’s the tactical version. Pick your one operator. Then, in one month:

Week 1 — Map the leverage points. Sit with them and list the top ten tasks they do in a normal month. For each, ask: how much would this improve if you had a well-tuned AI collaborator on it? Some tasks are natural fits (drafting, summarizing, analyzing structured data, exploring options). Some aren’t (client-facing meetings, high-stakes judgment calls, work that lives entirely in someone else’s head). You are not trying to AI-ify everything. You are trying to find the two or three tasks per week where the leverage is highest.

Week 2 — Build tight workflows on two of them. Pick the two highest-leverage tasks. For each, work with your operator to build a specific, opinionated workflow: which model, which prompts, which tools, which checks before output leaves their hands. This is not “use ChatGPT for it.” This is “the exact three-step process we use for producing our monthly client report, with the prompts saved and the review checklist attached.” Prompts are versioned. Checks are explicit. This is craft, not casual use.

Week 3 — Run those workflows for real, and log the misses. Ship real work through them. Every time the output needs correction, log what went wrong: was it the prompt, the verification step that missed, the tool choice, the specification? A week of real use produces a punch list of ten to twenty small improvements. Make them.

Week 4 — Standardize and hand off. For each workflow that’s now stable, write it up as a short procedure and identify who else on the team could use it under supervision. Not “give them access to the tool.” Give them the specific procedure your top person built, with the review checklist, and require the checklist be run. This is how a small team turns one skilled operator’s discoveries into a broader lift without giving up quality control.

At the end of thirty days you don’t have “AI adoption across the team.” You have two or three workflows that measurably work, one person who is significantly more capable than they were a month ago, and a scalable pattern for adding the next workflow. That is worth an order of magnitude more than a general-purpose seat for everyone.

What to stop doing

If you’re already on the “everyone gets a seat” path, here are the specific habits worth stopping.

Stop measuring AI adoption as if it were a good thing on its own. Adoption is an input, not an outcome. A team where 100% of people use AI weekly but whose work is not measurably better or faster hasn’t achieved anything except a line item on the software bill. The metric that matters is throughput on specific workflows and the quality of what comes out of them.

Stop shipping unreviewed agent output. This should be obvious and yet the “let the agent do it and post the answer” pattern keeps showing up in small businesses. Someone with taste — usually your top operator, sometimes you — needs eyes on anything AI-generated that leaves your business. Not to rewrite it. To verify it. When verification is treated as optional, the cost of AI-generated errors compounds silently until a client catches one, and then it compounds loudly.

Stop expecting the tool to teach the skill. People do not learn to write well by watching AI write. They do not learn to code well by accepting AI completions. They do not learn to analyze well by reading AI summaries. The tool produces the output; it does not build the underlying capability in the user. If you want your team more capable, you have to invest in their capability directly, on top of any AI use.

Stop treating “which model is best” as your strategic question. The gap between the best model and the second-best model right now is small compared to the gap between two users of the same model with different skill levels. Choosing the right model matters less than concentrating usage on the right people and building the right workflows.

The uncomfortable version, one line

If you want the single sentence version of everything above, here it is: the winning move for a small business in 2026 is to make your best person twice as effective with AI, not to give ten average people mediocre AI seats and hope the average goes up. It won’t. The math doesn’t work that way. The tool is a multiplier, and multipliers reward the highest base rate.

That’s an uncomfortable strategy, because it looks like under-investing in most of your team. It isn’t. It’s investing in your team differently: your top person gets AI leverage, your next person gets training and pairing to catch up, everyone else gets specific procedures that let them benefit from what the top people discovered — with quality checks that the tool doesn’t provide on its own. Every dollar and every hour goes further this way, and the compound effect over a year is not subtle.

The playing field isn’t leveling. The people who already know things are going to pull ahead faster than they used to, whether we like it or not. The interesting question for a small business isn’t whether that’s fair. It’s whether your business is on the ahead side of that gap, or the behind side, in 2027. That’s mostly a choice about who you invest in, and how, this quarter.

Sources

• Sean Goedecke — LLMs reward expertise, published August 3, 2026 (1,034 points on Hacker News). Primary source for the expertise-multiplier framing.
• Ankur Sethi — Prevent cognitive debt by manually retyping LLM-generated code, August 3, 2026 (502 points on Hacker News). Primary source for the cognitive-debt counter-current.
• Hacker News comment threads on both essays, August 3–4, 2026, as evidence of practitioner corroboration of the pattern from independent contexts.
• General observations on productivity-distribution shape are labeled opinion informed by publicly discussed usage data from AI vendors and enterprise deployments across 2025–2026; specific percentage claims are illustrative, not measured.

Freshness blip: primary sources are one to two days old at publication. Strategic implications are intended to hold across the current model generation and, given the mechanical reasoning above, likely intensify with the next.