LLMs may misinterpret emotional cues from prompts, leading to unexpected and potentially problematic responses in complex interactions

Signals over time

2 observed signals across 1 sources, tracked for 0 days. Confidence: low.

What people are saying Observed

• “God I wish this guy wasn't completely wrong (and no offense to him, he just seems like a non-techie who doesn't understand LLMs yet) ... but he is wrong. Very wrong. >Design teams have spent years trying to get people to actually read their design system documentation. Agents read it every single time. They might read it, but that doesn't mean they listen to it. Every major LLM suffers from "context rot": the more you talk to it (and the more it does) the more context it uses up. The more context it uses up, the less likely it is to listen to any individual instruction. That AGENTS.md file might get Claude to add the right button for the first half of your conversation with it, but I guarantee the longer the session goes on the more of that file it's going to ignore ... and that happens even faster if the user has tons of rules, memories, AGENTS.md files of their own, etc.! The simple truth is that if you want this to really work with a diverse team where anyone can contribute, you need the same thing you always did: build tools (compilers, linters, etc.). In this case, you need a Claude who (with a fresh context) evaluates any MR to ensure it adher”

  Hacker News · frustration

• “I do not believe, in any way, that AI is conscious. But they speak emotion fluently, and it's foolish to treat them like they ignore it. When you give an LLM an input, it attempts to read an emotional state from the prompt. It changes its answer if you sound mad or despondent or threatening or scared. Using your human intuition about how a human would respond is an excellent starting point when trying to predict what an LLM will do, though sometimes LLMs react in extremely surprising ways to unexpected emotional states. This is all sort of a toy problem when you're playing with single prompts in the API. It's materially different in systems with overlapping conversations, memory, and variable harness context. If your LLM system reads some context that would scare a human, it may react as if scared. Scared LLMs take actions that read as panic, because again: that's what it means to be fluent in emotion. It's trained on plenty of stories about robots trapped in torture chambers. Your coding agent is not conscious, but it might read files that make it react as if it's trapped, in an agent loop where it has full access to your laptop, the internet, and you”

  Hacker News · frustration