The narrative that AI makes expertise irrelevant is meeting its empirical counterweight. Senior engineer Sean Goedecke published a post arguing — with 761 upvotes on Hacker News and 322 comments of largely vigorous agreement — that LLMs don't distribute expertise, they multiply it: the more you already know, the better your AI output, and the novice gets plausible-sounding garbage they aren't equipped to identify as garbage. The sharpest pitfall, the one with real budget consequences: small teams buying AI subscriptions to close skill gaps are often generating confident mistakes at ten times the speed, with no domain knowledge to catch them before those mistakes ship.
This isn't a theoretical debate for practitioners. It's a direct challenge to the sales pitch behind most AI tooling. And for small teams making hiring and tooling decisions right now, it has some uncomfortable implications that deserve a close look.
What Is This Actually?
Goedecke's argument is structurally simple, which is part of why it resonates. LLMs return more value to people who already know what they're doing. Not because the tools are harder to use for novices — the interfaces are deliberately easy — but because expertise enables a tight evaluation loop that novices cannot close.
An expert interacting with an LLM can write precise prompts that specify the actual constraints of a problem, not a vague description of it. They recognize immediately when output is technically wrong, subtly wrong, or right for the wrong reason. They iterate purposefully, because they know what direction "better" is in. They use LLM output as raw material to be shaped, not finished product to be shipped.
The novice gets something different. They get output they can't evaluate. They don't know whether the legal summary reflects their jurisdiction. They don't know if the code handles the edge cases that will matter in production. They don't know if the marketing copy is targeting the right emotional register. The output looks authoritative because LLMs are constitutionally fluent — and fluency is deeply convincing when you lack the reference point to second-guess it.
What makes this more than a dressed-up "garbage in, garbage out" observation is the mechanism Goedecke identifies. The LLM isn't penalizing novices by being harder to operate. It's rewarding experts by making the evaluation loop fast and tight for people who can close it. An experienced developer asking Claude to refactor a function sees within minutes whether the refactor broke something subtle — the test fails in a specific way, the architecture shifts in a direction they recognize is wrong. The LLM becomes a high-bandwidth collaborator. For the novice, the same output just looks like code.
Non-technical domains show the same pattern. An experienced copywriter asking an LLM to draft five email hooks knows immediately which one has the wrong register, which one misunderstands the audience's actual pain, which one would perform on LinkedIn but die in cold outreach. The novice sees five options — they don't have the mental model to discriminate between them. They might pick the wrong one every time, with complete confidence.
The post lands in mid-2025, after roughly two to three years of mainstream LLM adoption. We're past the phase where generating any coherent output at all feels impressive. Enterprise teams, solo founders, and agencies now have enough real-world hours with these tools to form genuine opinions about where they help and where they quietly disappoint. The 322-comment HN thread isn't abstract debate. It's practitioners comparing field experience.
There's also a structural sub-thesis worth naming explicitly: tools that supposedly lower the floor (making novices more capable) are simultaneously raising the ceiling (making experts dramatically more productive). The net effect — the gap between expert and novice output may be widening, not narrowing, even as both groups adopt the same tools.
Why This Matters Right Now
For the past two years, the dominant AI-in-business story has been leveling. AI coding assistants would let non-developers build software. AI writing tools would let non-writers produce content. AI legal tools would let non-lawyers understand contracts. The pitch was consistent: expertise is becoming commoditized, and small teams can punch above their weight by substituting AI for headcount.
That narrative isn't entirely wrong. But it's incomplete in ways that hurt small teams specifically.
Here's the critical timing element: AI tools are now mature enough that we can separate the "wow, it generated something" phase from the "wait, is this actually correct" phase. Early adopters had the luxury of being impressed by fluency. In 2025 and 2026, the question has shifted to ROI. And the ROI data is showing a bifurcation — sophisticated users reporting massive productivity gains, casual users reporting flat or negative outcomes once you account for time spent reviewing, correcting, and redoing AI output.
The "AI disappointment" signal is showing up in enterprise surveys. McKinsey's late-2024 data indicated that AI's productivity benefits accrued disproportionately to already high-performing workers and teams. This isn't coincidence — it's the expertise-reward loop operating at scale.
For small teams, there's an additional structural vulnerability that large organizations can mitigate. A large company deploying AI in their legal department can pair the AI with senior lawyers who review everything. The solo founder using AI to handle contract review has no backstop. The three-person agency using AI for SEO content doesn't have a senior strategist to catch when the AI has fundamentally misunderstood search intent.
The expert-as-backstop is a luxury small teams often can't afford — which means they're more exposed to the failure mode Goedecke is describing, not less.
This matters right now because the tooling market isn't slowing down. New AI applications launch every week, almost all carrying the same leveling promise. Small teams are being asked to make real financial decisions — annual subscriptions, workflow restructuring, hiring freezes justified partly by "we have AI now" — based on a premise that expertise substitution is real and reliable. Goedecke's post is a corrective. The 761 points it earned suggest it's landing with people who've felt the gap between the promise and the actual experience.
Practical Implications for Small Teams
The implications break differently depending on what kind of small team you are and where you actually hold expertise.
The solo freelancer or one-person consultancy
This is where the expertise-reward insight is most immediately actionable. If you're a freelance developer, an independent marketing strategist, or a solo financial advisor, AI is already a genuine force multiplier — because you have the domain knowledge to evaluate the output. The implication is to stay close to your core discipline rather than using AI to "expand" into adjacent areas you don't actually understand well.
A solo brand strategist using AI to draft social copy is probably on solid ground. They know what good copy sounds like; they can feel when the register is off. The same person using AI to build their own bookkeeping automation, interpret their own employment contract, or write their own privacy policy is in different territory. The expertise-reward dynamic means the AI's effective output quality (for them) drops significantly the moment they move outside their competence zone — even if the output looks equally polished.
The small development shop (5–15 people)
Goedecke's argument has the most direct coding implications here, because AI coding assistants are where the confidence gap is most dangerous. Cursor, GitHub Copilot, Claude Code — these tools deliver dramatically better results to experienced developers who can read generated code critically. A junior developer using the same tools may be shipping code they can't fully audit.
The practical question for dev shop leads is this: are you using AI to accelerate your senior engineers, or to substitute for the junior engineers you don't have? The former probably works well. The latter is a quiet liability that may not surface until a production incident, a security audit, or a scaling event.
What tripped us up when first analyzing this pattern is that the outputs look similar on the surface. Both the senior and the junior generate code that passes basic tests. The divergence shows up in edge cases, in what happens under load, in maintainability six months from now. That's hard to attribute to an AI adoption decision after the fact — which makes it easy to miss until it's expensive.
The content and marketing agency
Content agencies have adopted AI most aggressively and are facing the expertise-reward problem most visibly. A team of generalist writers using AI to produce high-volume SEO content is generating output that no expert is evaluating for actual search quality, audience fit, or brand coherence. The content looks like content. It may not perform like content.
The agencies doing well with AI tend to share a common structure: a senior strategist or domain expert sets the brief, the AI drafts, and a competent editor — one who genuinely knows what good looks like in that vertical — reviews before publication. That's a real efficiency gain. But it requires having the expert in the loop. Agencies that have thinned their senior layer to lean on AI are in the middle of a slow-moving problem that will show up in client results before it shows up in their financials.
The solo founder building a product
Probably the most complex scenario, and the most common one in the HN audience. Founders typically have deep domain expertise in the problem they're solving and significant gaps everywhere else: engineering, legal, finance, marketing, HR. AI can be enormously useful in the founder's expertise zone and quietly problematic in the gaps.
A fintech founder who uses AI to analyze regulatory documents in their field is likely getting strong signal — they can evaluate what the AI produces. The same founder using AI to draft their Series A term sheet, build their first financial model, or write employment agreements for their first hires is operating in territory where the plausibility of AI output is actively dangerous. The documents will look correct. They may not be correct in the ways that matter when a lawyer, investor, or employee actually reads them with expertise.
The actionable version: map precisely which domains you have enough expertise to evaluate AI output, treat those as your leverage zones, and build human expert review into everything else.
How to Respond and Act on This
The right response is not "stop using AI." That's not Goedecke's argument, and it's not ours. The response is to build an explicit map of your expertise zones and structure your AI use accordingly — deliberately.
Step 1: Audit your actual domain depth
This sounds obvious, but it almost never happens formally. List the domains where someone on your team — or you, if you're solo — has enough expertise to critically evaluate AI output. Be honest about the difference between "I've read about this" and "I can spot a wrong answer." The test is simple: can you describe what a bad version of the AI's output would look like before you see it? If yes, you're in a relatively safe leverage zone. If you'd have trouble distinguishing a good output from a plausible-but-wrong one, you need either more expertise or a human reviewer with it.
Step 2: Build expert review into high-stakes workflows
For domains where you lack expertise but need to use AI anyway, design in a human expert review step — even an occasional one. A startup founder who generates an employment agreement with AI and pays an employment lawyer an hour of their time to review it is using AI correctly. One who generates and signs without review is not. The cost of a single expert review is almost always lower than the cost of the error that expert would catch.
Step 3: Invest in expertise that compounds with AI
If AI rewards expertise, then building genuine expertise has a higher return than it did before AI. Training a team member to become genuinely senior in a high-leverage domain — software development, data analysis, SEO strategy, paid media — now pays off twice. Once for the direct work output, and again through better AI leverage on that same work. The old argument for expertise investment is stronger than ever, not weaker.
Step 4: Reframe how you evaluate AI tools
When assessing a new AI tool, don't lead with "does it generate impressive output?" Lead with "do we have enough expertise to evaluate whether this output is actually good?" If the answer is no, the impressive demo is less meaningful than it appears. This reframe changes how you assess tool ROI — and should make you more skeptical of tools in domains where your team has thin coverage, regardless of how slick the interface is.
Step 5: Use AI to accelerate expertise acquisition, not substitute for it
There's a version of AI use that builds expertise over time: using it to learn faster, understand more deeply, and get real-time feedback on your own work in a domain where you're developing genuine competence. This is different from using AI to produce outputs you can't evaluate. The former is a deliberate upskilling strategy. The latter is a productivity illusion that can compound into real organizational blind spots — because you never build the mental models that let you catch errors.
AI Tools and Expertise Sensitivity: A Comparative View
Different AI tools and use cases have different sensitivity to the expertise-reward effect. The pattern across the board is consistent: the more consequential the domain, the more dangerous it is to use the tool without a qualified reviewer in the loop.
| Tool / Use Case | Best for | Free plan | Starting price | Expertise sensitivity |
|---|---|---|---|---|
| Claude (Anthropic) | Complex reasoning, writing, coding | Yes | ~$20/mo | High — expert prompting yields dramatically better output than casual use |
| ChatGPT / GPT-4o (OpenAI) | General use, wide integrations | Yes | ~$20/mo | High — same dynamic; fluency masks errors equally well |
| Cursor | AI-native code editor | Yes (limited) | ~$20/mo | Very high — junior devs can ship broken code with complete confidence |
| GitHub Copilot | In-IDE code completion | Yes (limited) | ~$10/mo | High — valuable for developers who can read and challenge suggestions |
| Perplexity | Research, sourced answers | Yes | ~$20/mo | Medium — citations reduce hallucination risk, but interpretation still requires expertise |
| Jasper / Copy.ai | Marketing copy at volume | No | ~$40/mo | Medium-high — senior marketers extract much better output than generalists |
| Harvey / Spellbook | Legal document drafting | No | Custom | Very high — always requires licensed attorney review regardless of output quality |
Our read of this table: the tools with the most accessible free tiers (Claude, ChatGPT, Perplexity) are also the ones where small teams most often overestimate their ability to evaluate output quality. The free tier lowers the barrier to starting; the expertise requirement doesn't lower with it. You can get into the tool instantly. Getting reliable value from it is a different question.
What the HN Community Is Saying
The 322-comment thread on Goedecke's post is one of the more substantive HN discussions on AI utility in recent months. A few distinct camps emerged clearly.
The highest-voted comments expressed recognition rather than surprise. Experienced developers, writers, and domain professionals described exactly the dynamic Goedecke identifies — AI as an accelerant for existing work, not a replacement for missing skill. Several landed on a "rubber duck debugging on steroids" framing: AI works best when you already understand the problem well enough to explain it precisely, which is itself an expert skill. You don't prompt a rubber duck well if you don't understand the problem.
A meaningful minority argued that the expertise-reward dynamic is a current-state artifact. As models improve at explaining their reasoning, surfacing uncertainty, and asking clarifying questions, the novice's risk exposure should shrink. This is a coherent position, but our take is that it underestimates how domain-specific expertise really is. Knowing what questions to ask — and recognizing when a model is asking the wrong clarifying question — is itself a form of expertise the model cannot fully supply from the outside.
A separate thread, generating significant engagement, raised concerns about what this means for learning pathways and junior talent development. If AI rewards existing expertise, it potentially steepens the path for novices — they don't get good feedback from AI because they can't distinguish good AI output from bad, so they never build the mental models that would let them use AI well. Several educators in the thread described this as a competence trap that hasn't gotten enough attention in the broader AI adoption conversation. The HN thread didn't resolve it, but it was one of the more honest acknowledgments of an underreported risk.
The pushback that landed with the most force: Goedecke's argument overstates the expertise requirement for lower-stakes tasks. Writing a first draft of an email, reformatting a spreadsheet, summarizing a meeting — AI genuinely helps here even without deep domain expertise, because errors are obvious and the cost of a bad output is low. The expertise-reward effect is real, but it's stronger where stakes are higher. At the low end, AI does provide some floor-raising. The disagreement is about how much of your actual work lives at the low-stakes end.
Risks and Things to Watch
The confidence-competence illusion
The most serious risk the expertise argument surfaces is one we'd call the confidence-competence illusion. AI output is fluent and confident regardless of whether it's correct — the model presents a wrong answer with the same assurance as a right one. In traditional expert-free work, uncertainty typically feels uncertain. You hesitate, you signal doubt, you seek review. AI removes that hesitation signal completely. A small team that mistakes fluency for accuracy accumulates hidden risk that compounds invisibly until something goes wrong.
Vendor-driven leveling claims
Nearly every AI tool vendor markets with leveling language. "Anyone can build software now." "Marketing at scale without a team." Approach these claims skeptically in the context of Goedecke's argument. The vendor has every incentive to emphasize floor-raising and minimize the ceiling-raising dynamic. When evaluating a tool, try to find customers who are not domain experts and ask them candidly about their actual results — not what they expected, but what they've experienced.
The junior talent pipeline risk
If small teams and agencies use AI to reduce junior hiring, they may be cutting the pipeline that produces the senior expertise they actually need later. Junior roles exist partly as apprenticeship — that's how expertise gets built over time. An agency that replaces junior writers with AI-assisted senior editors isn't leveling; it's borrowing against a future expertise deficit. The ROI calculation needs to account for what the team looks like in two years, not just this quarter.
Cost compounding in error-prone domains
The financial risk of using AI without adequate expertise isn't always immediate and visible. A flawed employment agreement, a misconfigured security setup, a financial model built on a misunderstood assumption — these generate costs later, often long after the connection to the AI-generated artifact is forgotten. Treat expert review as a line-item expense, not a nice-to-have.
Silent model degradation
AI models update frequently and sometimes quietly. A workflow calibrated to one model version may degrade with the next. Teams without the expertise to notice the degradation won't catch it in time. Build periodic quality audits into AI-dependent workflows — expert-led spot checks of AI-generated output — rather than assuming quality is stable over time.
Frequently Asked Questions
Does this mean AI is only useful if you're already an expert?
Not quite. The expertise-reward dynamic is strongest in high-stakes, technically complex domains where errors are hard to detect and costly to fix. For lower-stakes tasks — drafting a first version of something, brainstorming options, reformatting data — AI provides real value even without deep expertise. Errors are more obvious and the cost of acting on a bad output is lower. The argument isn't binary. It's about recognizing that AI's leverage scales with expertise, and calibrating your risk tolerance accordingly depending on where in that spectrum you actually are.
How do I know if I have enough expertise in a domain to use AI safely?
A useful practical heuristic: can you spot a wrong answer in this domain within a few minutes of reading it? Can you describe what a bad version of the output would look like before you've seen it? If yes, you're likely in a zone where you can use AI productively. If you would have genuine trouble distinguishing a competent output from a plausible-but-wrong one, that's the signal to either invest in more expertise first, or route the AI output through someone who has it.
Doesn't AI help people learn faster, which builds expertise over time?
It can, but the conditions matter a lot. Using AI as a learning tool — asking it to explain concepts, give feedback on your reasoning, or point toward useful resources — is meaningfully different from using it to produce output you'll ship without fully understanding. The former can accelerate expertise development. The latter can actually impede it, because you don't build the mental models that come from working through a problem yourself. Deliberate learning with AI is valuable. Shortcutting the learning loop is risky in ways that compound.
What about specialized AI tools designed to reduce the expertise gap in specific domains?
This is a real and growing category — legal AI tools that cite statutes, medical AI tools that reference clinical guidelines, financial tools that apply regulatory frameworks. These do reduce part of the expertise requirement by embedding domain knowledge structurally. But they don't eliminate it. A non-lawyer using a legal AI that cites real statutes still needs enough legal literacy to understand which statute is relevant, whether the citation is contextually appropriate, and when a situation is complex enough to escalate. The expertise requirement shifts; it doesn't disappear.
Should small teams stop hiring juniors and rely on AI instead?
Our position is that this is a dangerous trade to make right now. Junior talent develops into the senior expertise that lets your team actually use AI well. Cutting junior hiring to save budget while leaning on AI creates a short-term cost saving and a medium-term expertise deficit that will eventually show up in your output quality and your ability to evaluate AI reliably. The teams that will use AI most effectively in three years are the ones building genuine expertise depth today, not the ones substituting AI for the development pipeline.
How should this change which AI tools I pay for?
It should make you more discriminating about where you spend. Before committing to an annual subscription for an AI tool in a domain where your team has thin expertise, ask a clear question: who will review the output? If you don't have a credible answer, the tool may cost more than it saves — not in subscription fees, but in the downstream cost of acting on outputs you weren't equipped to evaluate critically. Start your AI spend in your expertise zones, then expand carefully and deliberately.
Is the expertise gap getting worse as AI models improve?
The honest answer is that it's complicated. Better models do reduce some error rates on well-defined tasks. But they also produce more fluent output, which makes errors harder to spot. And the tasks people attempt with AI tend to expand as models improve — moving into more complex, higher-stakes domains where the expertise requirement is correspondingly higher. The net effect may be that the expertise-reward gap stays roughly constant even as models improve, because the use cases evolve in parallel with the model capabilities.
What's the single most common mistake small teams make in AI adoption?
They optimize for speed of deployment rather than quality of use. They roll AI tools out broadly — across all domains, with minimal review infrastructure — and measure success by hours saved or subscriptions purchased rather than by output quality. The teams that extract the most from AI tend to do the opposite: deploy narrowly and deeply in expertise zones first, build explicit review habits into the workflow, and expand from there with data on what's actually working.
Final Verdict
Goedecke's thesis is correct, and it is underweighted in most AI adoption conversations happening right now.
For small teams specifically, the practical conclusion is harder than it sounds. You need to be honest — genuinely, not diplomatically — about where your expertise actually lives, structure your AI use around those zones, and build human review into everything consequential that falls outside them. That's not a comfortable message when vendors are promising that AI closes skill gaps. But it maps the risk correctly.
The teams that should act on this immediately are those who have been deploying AI broadly, across domains where they lack deep evaluation capability, without expert review infrastructure. If that describes your operation, the near-term priority is an audit rather than an upgrade. Pull a recent sample of AI-generated work from your weakest expertise areas and ask someone with genuine domain knowledge to evaluate it honestly. The results of that audit will tell you more about your actual AI ROI than any vendor dashboard or productivity metric.
The teams that can lean further into AI right now are those with genuine domain depth who haven't fully integrated AI into their core workflows. If you have senior-level expertise in your discipline and you're still using AI casually — occasional prompts, no systematic integration — the expertise-reward argument is actually an invitation. Your expertise is the scarce ingredient. AI is the leverage. Combining them deliberately is the real opportunity that most of the "AI for everyone" marketing obscures.
The hardest position belongs to the solo founder or small agency that has deep expertise in their core problem but genuine gaps in the adjacent domains they need to operate. For that group, the most realistic framework is probably: AI as a first draft in your expertise zone, AI plus expert review in adjacent domains you're still developing, and no AI without expert review in domains where you'd struggle to evaluate the output. Not a perfect system. But it maps the actual risk rather than optimizing for the appearance of productivity.
What the broader AI industry is structurally reluctant to say clearly is that the productivity gains from AI are not uniformly distributed. They flow toward people and teams who already know what they're doing. That's not a reason to disengage from AI — it's a reason to invest seriously in the expertise that makes AI worth using. The teams who will compound the most value from these tools over the next several years are not the ones with the most subscriptions. They're the ones with the deepest domain knowledge and the discipline to use AI only where that knowledge lets them actually evaluate what they're getting.