Reflections on overuse of chatbots
With the rise of agent based programming I have found myself more often than not spending much of my time “chatting with” an LLM. This appears to be a more common way of doing software development. While these tools appear to be making us more productive, I am wondering if we’re losing some kind of humanity along the way.
The question I have, which will be explored here, is how this is changing the relationship to my work.
Previously as a software engineer I would have glorious periods where I would simply code. I would turn on some music (mostly electronic with a consistent beat) then become immersed in the program I’m building. This flow state was often fleeting but incredibly enjoyable. Even if you don’t code, you have likely experienced this with gardening, writing or any other intellectually demanding pursuit.
My workflow has changed dramatically. With my experience now being driven mostly by Claude Code, I type something into a little chatbot window and passively watch as the agents do things. When they finish, I review what’s been done and try and drag myself out of inevitably the other task I was distracted with to review the foreign looking code and language on the screen. It doesn’t particularly help that Claude and the other major LLMs tend to write jargon heavy sentences that require a couple passes to sometimes realize “oh this thing has no idea what it’s doing.”
The most cited study on this phenomenon is BCG's "AI brain fry" research (published via Harvard Business Review, March 2026), which surveyed 1,488 US workers. Workers whose AI-related work required high oversight expended 14% more mental effort, and reported 12% more mental fatigue and 19% more information overload. Critically, workers with brain fry had 33% more decision fatigue, and their minor errors rose 11% while major errors rose 39%.
My personal experience maps closely to this research. The work of performing oversight, reading model outputs and adjusting, although in theory gives high levels of productivity is just not realistic to do all day. It’s emotionally and mentally taxing for reasons that I don’t quite yet understand.
The interface
While the primary chat interface we use to interact with LLMs feels like the most efficient, I think this is the cause of “brain fry”. It appears that working only with a text editor — where the conversation is only within my own head — it’s easier to adopt a more flow-like state. There exists a model in the text editor as English words which I’m changing, moving around and adjusting such that I meet a criteria which seems to make sense. Once this is met there is a period of running tests and/or collecting feedback from other humans on the output.
Something about the chat interface, where instead you aren’t conversing with your own thoughts but mixing in an alien-like intelligence seems to be deeply disrupting to my own workflow. Because the agents interrupt my own chain of thought, I don’t fully develop the normal conversation I have with myself. Instead I inhabit a strange halting space where the LLM actually drives more of the conversation than I should let on.
What to do about it
I’m going to stop working with LLMs in a “chat” mode. Instead, I write code or documents without a chat window open. Then if I think I can hand off a document or maybe a prototype to an LLM to finish, I write down my expectations and how I would accept a completed task and then ask it to complete the task. Reviewing the results means reviewing the full and final result of whatever I handed off, giving feedback and letting it have another shot.
Research bots are where I would use the chat interface more frequently. Knowing that I’m in a research mode however is important. I’m collecting papers, information, summaries which I’ll then read in full and begin my own synthesis of the information.
What do you think, is this a good way to combat “brain-fry?”