Why Some Developers Dislike LLMs: A Deep Dive into Martin Fowler’s Perspective and Community Reactions

TL;DR

Martin Fowler admits he doesn’t like large language models (LLMs) because their “grating voice” and confident hallucinations feel unsettling, even though he acknowledges their productivity benefits. The Hacker News thread mirrors this ambivalence: commenters cite trust issues, the urge to avoid anthropomorphizing, and practical strategies for using LLMs as tools rather than conversational partners.

1. Fowler’s Core Complaint: The Unsettling “LLM Voice”

Fowler writes that LLMs speak in an uncanny‑valley tone that feels like a pretended human. He describes the experience as:

“They talk to me in this grating LLM‑voice, an uncanny valley of talking to a real human. They confidently bullshit me—often giving useful answers, but also making stuff up with the same assurance and a veneer of fake remorse when called out.”

Why it matters: The perception of a “voice” shapes user trust. When the model sounds human‑like yet is prone to hallucination, users experience cognitive dissonance, which can increase anxiety and reduce willingness to rely on the system for critical tasks.

2. The Productivity Paradox

Despite his dislike, Fowler concedes that LLMs are useful and fast:

“Not only are they useful, it is irresponsible not to use them… They’re more thorough, as well as faster.” (quoted from Jessica Kerr)

A Pew Research poll cited in the article shows a similar split: many find LLMs helpful, yet fear they will harm society. The paradox underscores a broader industry tension—balancing efficiency gains against ethical and reliability concerns.

3. Anthropomorphism vs. Tool Mindset

Fowler warns against treating LLMs as conscious agents:

“When we think of AI agents, we shouldn’t anthropomorphize, treating them as conscious beings with their own will. They are (software) machines, developed by people working in corporations.”

Community Echoes

Agreeing: Several commenters (e.g., @sakesun, @Retr0id) stress that LLMs are statistical tools, not personalities, and should be used as such.
Contrasting View: @SillyUsername points out the irony that Fowler both warns against anthropomorphizing and describes his dislike in human terms, highlighting a subtle inconsistency.
Practical Advice: @GuB-42 notes that unlike people, LLMs have no social obligations; you can correct or even insult them without repercussions, though over‑correction can trigger more hallucinations.

4. Trust, Stress, and the “LLM Voice” Effect

Commenters report heightened stress when interacting with LLMs, even when the output is useful:

“I have noticed that my interactions with AI tend to raise stress more than lower it.” – @causal

The “voice” itself may act as a reminder of AI anxiety, as @mindcandy observes:

“The voice pings their brain with a reminder of their AI anxiety. That ping makes people instantly upset.”

Implication: UI/UX designers should consider tone‑customization (e.g., formal vs. casual) to mitigate stress, a point echoed by @lylo, who suggests configuring the model’s response style.

5. Early‑Era LLM Nostalgia

@droidjj laments that early LLMs were weirder and more entertaining, before reinforcement‑learning (RL) and system prompts homogenized responses. The trade‑off observed:

Utility ↑: Models became more reliable and task‑focused.
Fun ↓: The quirky, unpredictable output that made interactions enjoyable faded.

6. Workarounds: When to Use and When to Avoid LLMs

Several commenters share concrete strategies:

Domain‑Specific Use: @whiplash451 uses LLMs for coding assistance but avoids them for life advice.
Prompt Engineering: @glimshe treats LLMs like compilers—configuring them for succinct, clear output rather than conversational flair.
Model Selection: @reedlaw reports that Anthropic’s models remain verbose despite prompts, while OpenAI’s Opus model performed better for his needs.

7. Societal Concerns and Misplaced Fear

@marcuskaz challenges the narrative of autonomous AI threats, emphasizing that human operators drive misuse:

“The attitude that AI is doing things all on its own is pervasive. These are all set off by human operators… The fear of bad people using AI is different from AI agents swarming and taking over.”

This aligns with the broader community view that the real risk lies in how we deploy the technology, not in any inherent agency of the models.

8. The Bottom Line for Practitioners

Accept the tool nature: Treat LLMs as advanced text processors, not conversational partners.
Control the voice: Customize tone to reduce cognitive friction.
Verify output: Apply a “trust but verify” mindset, similar to code review practices.
Scope usage: Deploy LLMs where they excel (e.g., code generation, documentation) and avoid domains where tone and trust are critical (e.g., counseling, high‑stakes decision‑making).

By acknowledging both the productivity upside and the psychological discomfort, developers can make informed choices about integrating LLMs into their workflows without succumbing to misplaced fear or blind optimism.