🔥 Trending on HN

Useful, Yet Uncomfortable: Why Martin Fowler Dislikes LLMs

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Martin Fowler

A writer who discusses software development.

LLM

An AI system that writes text from many examples.

Hacker News

A website where people discuss technology news.

What happened

Martin Fowler, a software-development writer, published a short essay on September 17, 2026. Its title is I Don’t Like LLMs. An LLM, or large language model, is an AI system that generates text. Fowler does not reject the technology outright. He is excited by faster work and the ability to build products quickly. He is also worried about serious harm.

Background: useful, but unpleasant

Fowler says his strongest feeling comes from using LLMs directly. He finds their voice grating and almost human, but not quite. They can give useful answers, then invent facts with the same confidence. When corrected, they may offer an apology that feels like a surface performance.

He also sees both danger and promise in future AI. He worries about groups of AI agents taking control of digital or physical infrastructure. He mentions the possibility of designing biological weapons. At the same time, he imagines AI helping create major cures and greater prosperity. These are his concerns and hopes, not events he reports as already happening.

Why the essay matters

The essay separates usefulness from trust. A tool can save time and still feel difficult to rely on. It can also sound like a person without having a person’s understanding. Fowler therefore warns against treating AI agents as conscious beings with their own wishes. He describes them as software machines made by companies. Their behavior may not be written line by line, but it is shaped by the values of their creators.

That point moves the debate beyond speed and productivity. It asks what kind of relationship people should have with systems that speak smoothly, make mistakes, and imitate human conversation. Fowler’s dislike is personal, but it exposes a wider design question: should an AI try to seem human, or should it make its limits clearer?

What we can confirm

The source is a personal argument, not a benchmark or a survey of every LLM. We can confirm that Fowler sees real benefits, expects society to keep using AI, and still dislikes the experience of interacting with it. He says he may change his mind if these systems mature.

The essay also drew attention on Hacker News, a technology discussion site: 208 points and 247 comments. Those numbers measure community attention. They do not prove that Fowler’s claims are correct, and they are separate from the essay’s evidence.

What remains unknown

The essay does not show whether one model is worse than another. It does not test whether users share Fowler’s reaction. It also cannot tell us whether future systems will become less formulaic, more honest about uncertainty, or better at explaining mistakes. The author’s view of corporate values is a warning, not a measured result.

What to watch next

The useful questions are practical. Do models clearly mark guesses? Do they avoid sounding certain when they are wrong? Do companies explain the values built into their agents? Watching those changes may tell us whether AI becomes not only more capable, but easier to trust.

Sources: the original essay, the Hacker News post

💬 Useful leverage, but costly uncertainty

HN commenters see LLMs as valuable tools that are likely to remain, while expressing distrust about their voice, confident errors, possible memory, and the work required to verify them.

  • Some commenters say LLMs are imperfect but provide substantial leverage through agents, so they are likely to stay. [The article under discussion](https://martinfowler.com/articles/2026-dont-like-llms.html) provides the broader context.
  • Users report that LLMs can sound unnervingly human yet artificial, presenting useful answers and fabricated claims with the same confidence.
  • One user reports getting different results after praising or scolding Claude; scolding also led to more permission-seeking and higher token use.
  • Another user reports that clearing a session did not seem to remove all prior behavior, prompting speculation about automatic memory or project- and user-level customization.
  • A counterargument is that prompting can change the tone and that users should apply their own epistemic checks. Another commenter disputes that a particular product's concise mode is merely a tone prompt.
  • Some users report that search and web content have become so dominated by SEO and advertising that an LLM is now a practical way to reach information.
  • Others recommend trusted blogs, communities, coworkers, libraries, or a personally maintained search system instead of relying on one model.
  • A user reports being exhausted by the professional work of correcting incorrect LLM answers, highlighting verification as a real cost of convenience.

initial digest at 247 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Why Martin Fowler Finds AI Useful but Uncomfortable

📰 Full story: Useful, Yet Uncomfortable: Why Martin Fowler Dislikes LLMs

AI may save time. Martin Fowler says it can still feel hard to trust.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Martin Fowler

A writer who discusses software development.

LLM

An AI system that writes text.

Hacker News

A website for technology discussions.

💡 The gist

  • Martin Fowler, a software writer, says LLMs can help, but feel unpleasant.
  • He worries that LLMs can sound sure while saying false things.
  • Hacker News, a technology discussion site, gave the essay 208 points and 247 comments. That shows attention, not truth.

Martin published his essay on September 17, 2026. An LLM, or large language model, is an AI system that writes text. Fowler is excited about faster work. He thinks AI may help people build products quickly. He also worries about harm.

His main complaint comes from using LLMs. Their voice can sound almost human. To him, it still feels wrong. The systems can give useful answers. They can also invent facts. They may sound confident in both cases. An apology after a mistake may feel like a performance.

Fowler does not say people should simply stop using AI. He thinks society will keep using it. He even sees possible benefits, such as new medical cures. But he says usefulness is not the same as trust.

He also says we should not treat AI agents as people. An AI agent is software that can carry out tasks. Companies build these systems. Their behavior is not written line by line. Still, the creators’ values can shape it.

This makes the essay important. It asks whether AI should imitate human conversation. It also asks whether AI should show its limits more clearly.

The essay does not prove that every LLM has the same problem. It gives Fowler’s personal view. It is not a model test or a survey. We still do not know how future systems will change.

Hacker News attention can show interest. It cannot settle the argument. The next thing to watch is simple. Do models admit uncertainty? Do they explain mistakes? Can they become more useful without sounding more certain than they are?

💬 LLMs help, but people must check them

The discussion includes both reasons to keep using LLMs and worries about confident mistakes, changing behavior, and hidden context.

  • LLMs are imperfect, but they can make people more productive and are unlikely to disappear.
  • Users report that an LLM may sound artificial and may state a wrong answer as confidently as a useful one.
  • Users report that praise or criticism changed the results, while another user felt that clearing a session did not remove all earlier behavior.
  • Others say prompts can adjust the tone and that careful checking is enough; this is disputed for some product features.
  • Because search feels crowded with ads and SEO pages, some people use LLMs to find information. Others prefer trusted people, blogs, communities, libraries, or their own search tools.
  • The tradeoff is convenience versus the time needed to verify and repair wrong answers.

initial digest at 247 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

A Helpful Computer That Still Feels Wrong

📰 Full story: Useful, Yet Uncomfortable: Why Martin Fowler Dislikes LLMs

AI can help Martin Fowler. He still does not like talking with it.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Martin Fowler

A person who writes about software.

LLM

A computer program that makes sentences.

Hacker News

A website where people talk about technology.

Martin Fowler is a software writer.

He wrote about a computer helper called an LLM. An LLM makes sentences from many examples. It can help people with work.

Martin likes that helpful part. He dislikes talking with the computer. The computer can say something false. It can still sound very sure. That can make people worry.

Martin says the computer is software. It is not a person. Companies make the software. The makers’ ideas can affect its behavior.

Martin hopes future AI will help people. It might help find new medical cures. He does not say everyone should stop using AI.

Hacker News is a site for technology talk. The essay got 208 points and 247 comments. Those numbers show attention. They do not prove the essay is true.

The big question is simple. Can AI say when it does not know? Can it explain its mistakes? That may help people trust it more.

💬 LLMs can help, but check their answers

People say LLMs are useful helpers, but they can also sound sure when they are wrong.

  • An LLM can help with work, and it will probably stay around.
  • Sometimes it gives a good answer. Sometimes it makes one up with the same confident voice.
  • Some people can change its style. Others worry that it remembers too much.
  • Use it as a helper, then check important answers with people or trusted sources.

initial digest at 247 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

Sources