Kristian Polso Says He Will Not Use LLMs for Professional Coding

On November 8, 2026, Finnish engineer Kristian Polso published Why I will not be using LLMs in my coding and said he will stop using large language models and coding agents in paid work. The essay reached the Hacker News front page the same morning with more than 70 points and a long comment thread. Polso is not a novice critic: he describes roughly ten years shipping production machine learning, including computer-vision and air-quality models. His post is an opt-out manifesto at a moment when Stack Overflow’s 2026 Developer Survey shows AI agents everywhere but thin trust for high-stakes decisions.

TL;DR — the essay in one table

Why the story cleared the bar

This is not a generic “AI is bad” rant. Polso names specific failure modes builders already measure: review load, trust gaps, and dependency on vendors. The timing matters because employers increasingly treat agent adoption as default while surveys show developers rarely trust models on important calls.

Polso’s essay at polso.info lists the dealbreakers verbatim; our summary tracks his headings and examples. Automated fetch sometimes returns HTTP 418, so corroborate wording on his site when you quote him in policies or posts.

Dealbreaker one — cognition and “brainfog”

Polso reports subjective disorientation after long LLM sessions and points readers to essays on cognitive debt from leaning on assistants for writing. Whether you accept the neuroscience framing or not, teams already describe a verification hangover: less time typing, more time checking.

Stack Overflow’s 2026 headline question — “How are you doing?” — lands in a year when 45.1% of respondents say they are complacent at work and 32.6% unhappy. Among the complacent, burnout or tech fatigue is the top reason (21%). That does not prove LLMs caused the fatigue, but it shows Polso’s felt experience is not isolated.

For managers, the actionable read is to watch calendar load after agent sprints, not just story points closed.

Dealbreaker two — quality theater in diffs

Coding documentation metaphor: gear linked to an open booklet, illustrating verbose agent output that looks documented but may lack substance

Polso’s sharpest technical example is review waste: an agent “refactored” toward greenfield code when a schema tweak would suffice. Hacker News echoed that pattern. User sgt noted that Claude-style output with exhaustive comments and tests looks expert-grade until you read the logic.

That is the same economic transfer our verification tax guide documents: generators save keystrokes; reviewers pay in attention. Microsoft’s proof-driven development kit coverage exists because line review no longer scales when agents dump thousands of lines.

Teams should score review minutes per story point, not tokens consumed.

Dealbreaker three — maintainability and vendor dependence

Polso warns that if agents write most code, eventually no human knows why the system behaves as it does — and vendors effectively own your architecture choices. That is stronger than most corporate policies admit.

The 0rwx framework gives language for this risk: EXECUTE-level work is output you cannot realistically audit line by line. Polso is arguing for staying in ZERO or tight WRITE modes professionally.

Low-quality agent contributions on a conveyor beside one green gem, illustrating maintainer slop filtering

The Linux Foundation and OpenSSF report Navigating Open Source Software Development in the Generative AI Era finds roughly one in four incoming pull requests classified as low-quality AI contributions while still reporting double-digit time savings on some tasks. Polso’s maintainability fear is the maintainer-side mirror of that slop rate.

Dealbreaker four — liking to write code

Polso says coding is how he thinks and that reading others’ code reveals how they think. Several HN replies called that a hobbyist luxury. User iLoveOncall countered with a FAANG anecdote: heavy AI users on the team consume more tokens, while the author raises more pull requests with less model use.

Both can be true. Craft preference is not disproved by aggregate productivity stats. Lead engineers who opt out may still be your best auditors if you let them.

Dealbreaker five — environment

Polso ties tool use to data-center expansion and cites Finland’s halt on preparatory clearing for Google-linked sites after environmental complaints. explainx.ai has covered data-center siting fights separately. Environmental harm is a reason some individuals boycott products; it is not a universal argument against every local or small-model workflow.

If environment is your line, document which workload runs where — cloud frontier calls versus on-device inference — instead of treating all coding assistance as identical.

Dealbreaker six — politics, compliance, and the CLOUD Act

As an EU citizen, Polso objects to routing professional artifacts through US-headquartered services when the CLOUD Act can reach data stored abroad. That matters for regulated clients even when vendors market “EU data residency.”

Our Claude connector and OAuth bug-bounty post shows how opaque routing already creates security surprises. Compliance-sensitive teams should map data paths and subprocessors the way Polso wants, whether or not they ban chat tools entirely.

What the 2026 surveys add

The Stack Overflow 2026 survey is the cleanest second source for this story:

• 30,903 respondents across 169 countries.
• 69.9% used an AI agent while problem-solving; 82.7% still start with online search.
• On trust: 48.0% trust AI when verification is easy; 6.6% trust it for important decisions too.
• 52.1% say source attribution is very important when evaluating generated code.

Polso’s position is the extreme tail of that distribution: if you will not trust important decisions to models, why automate the craft entirely?

SD Times and other outlets have reported similar adoption-trust gaps in smaller samples during 2026. The direction aligns even when exact percentages differ.

Polso also points readers to Esther Weidauer’s essay Why I Resist Generative “AI” for a fuller critique of treating chat tools as neutral “just tools.” He agrees with parts of that framing while keeping his own production-ML background in view. That citation matters because his opt-out is not anti-statistics or anti-automation wholesale. It targets delegating professional authorship of code to opaque services whose incentives are not his employer’s or his clients’.

For team leads, the survey gap suggests a policy template: mandate assistance where it helps exploration; forbid unreviewed merges on critical paths. That matches how should developers stop reviewing AI-generated code balances trust scores from prior Stack Overflow waves with the reality that usage outran confidence years ago.

Finally, Polso’s Finland environmental example is a reminder that vendor choice is infrastructure choice. Developers who boycott cloud APIs for energy reasons may still run local inference on existing hardware. The ethical line is not always “no models” but “which bill of materials and jurisdiction you fund.”

What developers are saying on Hacker News

Clustered speech bubbles around one highlighted reply, illustrating a substantive Hacker News comment thread

The thread on Polso’s essay is substantive enough to treat as field research:

• Career pressure. el_psy_congroo wrote that passion projects might tolerate a no-LLM rule but employment often will not.
• Spectrum, not binary. rowanseymour argued for strict review on production paths and permissive agent use on throwaway tooling.
• Privileged take? RockRobotRock challenged whether opting out is realistic for most workers.
• Counterexample veterans. XCSme said they stopped typing code for six months yet deliver faster — the opposite lived experience.
• Heroin metaphor debate. incrudible rejected Polso’s drug analogy while agreeing tools can be brutishly fast.

None of these comments verify Polso’s environmental or legal claims; they show why the essay resonated: identity, review economics, and employer mandates.

How this fits explainx.ai’s labor coverage

Polso’s post is the individual version of structural stories we already track. Ricardo Pallás’s Indeed essay separates macro posting cycles from AI hype. Kevin Boone’s trust-in-AI-generated web content piece asks what readers should believe when machines author pages. Proton’s engineering manifesto reaction shows teams arguing that review queues, not keystrokes, became the bottleneck.

Polso adds a personal policy: zero professional delegation.

Practical guidance for teams

Do not punish auditors. If your strongest engineer reads every diff and uses few tokens, that is an asset while agents flood review queues.

Require disclosure, not uniform adoption. Labels like 0rwx make opt-in and opt-out legible in retros.

Measure slop rate. Track rejected or heavily rewritten agent PRs weekly. Our human slop and legacy code post shows why unreviewed merges have security tails.

Offer narrow exceptions. Polso admits agents help when he is stuck framing a question. Many teams allow READ-level planning without EXECUTE-level merges.

Separate environment policies from editor policies. On-prem or regional inference may satisfy compliance without banning all assistance.

Antonin Carette’s relay-race metaphor for human–agent coding argues against both total refusal and sofa-only delegation—adjacent to Polso’s maintainability concerns.

Related reading

• The AI verification tax
• 0rwx permission levels for AI work
• Indeed data and AI job fear
• Should developers stop reviewing AI-generated code?
• Human slop and legacy vulnerabilities
• Proton’s programming-is-dead debate

Survey percentages and Hacker News scores are accurate as of November 8, 2026. Cite Polso’s post for verbatim dealbreaker wording.