You’ve probably tried it. You asked an LLM to draft a sensitive email to a difficult stakeholder, or to summarize a politically charged meeting. It gave you something that looked perfect—sterile, overly structured, and completely devoid of human context. You spent forty minutes editing it so you didn’t sound like a sociopath, and realized you could have written it yourself in ten.
We’ve been sold a “magic assistant,” but what we actually got was an eager intern with no memory, no stakes, and a terrifying tendency to hallucinate facts with absolute confidence.
A recent Hacker News thread asked the question that tech CEOs are dodging: Are any of you actually using LLMs for non-coding work voluntarily? The answers were a collective sigh of relief. Everyone is realizing the emperor has no clothes. We can all agree LLMs are phenomenal at writing code. But code is a very small fraction of what the rest of us call \”work.\”
Here’s the twist nobody in Silicon Valley wants to admit: LLMs aren’t good at coding because they are smart. They are good at coding because code is a cage. Code is a closed, formal system with clear rules, strict syntax, and an unforgiving compiler. If the code works, it works. If it doesn’t, it fails immediately.
Code is a closed system with an unforgiving compiler. Human work is an open system with an unforgiving HR department.
Most non-coding work—strategic analysis, drafting emails, managing interpersonal dynamics—is open-ended and context-dependent. It requires institutional memory, political nuance, and human judgment. When an LLM writes code, a bug is just a bug. When an LLM drafts a strategic memo, a \”hallucination\” is a liability. It’s plausible, dangerous nonsense that can destroy trust, misallocate resources, or trigger a PR disaster. The very features that make LLMs powerful for code—determinism and verifiability—become massive liabilities in ambiguous, socially nuanced work.
The tech industry wants you to believe the bottleneck is the model’s capability. Just wait for the next version, they say. It’ll be smarter. But they are wrong. The real bottleneck is organizational. Non-coding work requires accountability. Who takes the blame when the AI summarizes a meeting and completely misses the unspoken tension between two executives? Who owns the fallout when an AI-generated financial analysis confidently cites a fabricated statistic?
You cannot prompt-engineer your way out of an accountability deficit.
Until workflows are redesigned around the model’s limitations rather than its strengths, blanket AI rollouts will just create undue chaos. The real value of LLMs right now isn’t in replacing human judgment; it’s in narrow, structured tasks. Research synthesis. Template generation. Boilerplate drafting. Tasks where the output is a starting point, not a final product sent to the client.
Stop trying to force a tool built for deterministic systems to navigate the messy, ambiguous reality of human relationships. The \”universal work assistant\” is a marketing fantasy. The moment we accept that is the moment we can actually start getting real work done.
FAQ
Q: If LLMs are so flawed for general work, why is every tech company marketing them as universal assistants?
A: Because 'a really good code completion tool and brainstorming partner' doesn't justify a trillion-dollar valuation. They need the market to believe AI is a universal productivity engine, even if the current reality is narrow and highly contextual.
Q: What's the practical implication for my team right now?
A: Stop trying to automate final outputs like client emails or strategic memos. Restrict LLM usage to narrow, verifiable tasks: research synthesis, generating rough templates, or summarizing raw transcripts. Always have a human verify the context.
Q: Isn't this just a temporary problem? Won't future models handle context better?
A: Models will improve, but the core bottleneck isn't capability—it's accountability. Organizations cannot outsource judgment, institutional memory, and blame. The liability of a 'hallucination' in human dynamics will always be too high to fully remove the human from the loop.