The LLM Flow State: Why AI Makes Work Addictive
In this week's edition: The LLM Flow State: Why AI Makes Work Addictive | News Roundup
I used to pester my wife about when dinner would be ready. Now, when it's her turn to cook, she has to shout me up from my office. I'm not sure that's progress.
Something strange has happened to my relationship with work. I'll sit down to "quickly check something" with Claude or ChatGPT, and two hours later I'm still there, deep in some rabbit hole I didn't know existed when I started.
I mentioned this to Gavin last week. He had the same experience. We both described it the same way: when you're in a groove with an LLM, making big leaps, iterating fast, you don't want to stop. There's a momentum to it. A pull.
So I did what any reasonable person would do: I asked my AI assistant to research whether AI assistants are addictive.
The Science (Simplified, Because I Had to Google Half the Words)
Turns out there's a concept in psychology called "flow state." Coined by a Hungarian psychologist whose name I won't attempt to spell, it describes that feeling of being completely absorbed in a task. Athletes call it being "in the zone." Artists have described getting so lost in their work they forget to eat.
If you've been reading this newsletter for a while, you might remember me banging on about focus time and deep work this time last year. I went through a proper productivity obsession phase. Blocking out calendars, batching tasks, trying to protect uninterrupted hours. The irony is that I've now stumbled into more flow states by accident, using AI tools, than I ever managed to engineer deliberately.
The neuroscience, as far as I can understand it, goes something like this:
When you're distracted and scattered, checking your phone every few minutes, your brain gets little spikes of dopamine. Quick hits that fade fast. It's the neurological equivalent of junk food.
When you're in flow, the dopamine release is steady. Your brain rewards you continuously for staying engaged. The prefrontal cortex, which apparently houses your inner critic, dials down. (Scientists call this "transient hypofrontality," which sounds like something from Star Trek but is apparently a real thing.) Your self-doubt goes quiet. Ideas connect more easily.
Other brain chemicals get involved too. Norepinephrine sharpens your focus. Something called anandamide, nicknamed the "bliss molecule," helps you link distant ideas together. Endorphins make problem-solving feel genuinely enjoyable rather than like pulling teeth.
Time distorts. What feels like thirty minutes is actually two hours. Your wife is shouting that dinner's ready and you haven't noticed.
Why LLMs Are Flow Machines
Flow requires three things: immediate feedback, a challenge that matches your skill level, and a sense of progress.
LLMs deliver all three. You prompt, it responds, you iterate. There's no waiting for a colleague to get back to you. No context-switching to check email. No meetings interrupting your train of thought. Just a continuous loop of input, output, refinement.
The "big leaps" matter too. When you're building something with AI, you're not grinding through incremental progress. You're jumping from a rough idea to a working draft in minutes. Each jump is a little reward. Your brain likes rewards.
A research paper from 2025 actually argued that people aren't becoming "AIholic" in any clinical sense. There's no evidence of the negative consequences you'd associate with addiction: no impaired control, no psychological distress, no functional impairment. What people are experiencing is closer to what painters and writers have always described. Getting absorbed. Losing yourself in the work.
The Downsides (Because There Are Always Downsides)
Here's where I have to be honest: I'm not sure all of this is good.
The same focus that makes me forget about dinner also makes me forget about exercise. I've skipped lunch because I was "nearly finished" with something. The hours I'm spending in flow with AI are hours I'm not spending doing other things. Some of those things, like going for a walk or having a proper conversation with an actual human, are probably more important than whatever I'm building. I drink too much Guinness to be skipping workouts, but here we are.
There's also a quality question. Flow feels productive. But feeling productive and being productive aren't the same thing. I've had sessions where I've generated pages of content that, in the cold light of the next morning, turned out to be waffle. The momentum of the work carried me past the point where I should have stopped and asked "is this actually any good?"
And then there's the impatience problem. When you're used to an AI responding instantly, iterating in real time, and never making you wait, working with actual humans becomes frustrating. I've noticed myself getting annoyed when colleagues don't reply within an hour. When a meeting could have been a prompt. When someone needs to "think about it" instead of just giving me an answer. That's not a good look. The tool is making me worse at collaborating with people who aren't the tool.
There's also a dependency question. If I'm training my brain to expect this kind of rapid feedback loop, what happens when I have to do work the old way? Writing a proposal without AI assistance now feels painfully slow. Reading a long document without being able to ask questions feels inefficient. I'm not sure I'm becoming a better worker. I might just be becoming a different kind of worker, one who needs the tool to function.
What I'm Doing About It
Honestly? Not much.
I could tell you I've implemented time-boxing, or that I deliberately schedule AI-free focus time, or that I've developed a mindful practice around tool usage. But that would be nonsense. The truth is I'm still figuring this out.
What I have done is notice it. Which sounds pathetic as a solution, but it's a start. I'm aware that my patience for human-speed collaboration has taken a hit. I'm aware that some of my "productive" sessions produce nothing useful. I'm aware that my wife has to shout louder than she used to.
Awareness isn't action. But you can't fix something you haven't noticed. So for now, I'm in the "noticing" phase. If I come up with anything more useful, I'll let you know.
The Honest Summary
Using LLMs can trigger a genuine flow state. The science backs this up. It's not addiction in any clinical sense, but it is a powerful pull towards a particular kind of work.
Whether that's a good thing depends on what you're doing with it. If you're using AI flow to build things that matter, ship work faster, and solve problems you couldn't solve before, brilliant. If you're using it to avoid harder tasks, generate busywork that feels productive, or neglect the parts of life that don't fit neatly into a chat interface, less brilliant.
I'm probably doing a bit of both. The trick, I think, is noticing which one you're doing at any given moment.
On the bright side, if I'm not going for walks but I'm also not hearing when dinner's ready, maybe it'll net out.
News Roundup
Barclays ties AI directly to 12% profit jump
UK bank Barclays reported £9.1B in pre-tax profit for 2025, up 12% from £8.1B. The interesting part: they explicitly linked AI to cost reductions, raising their 2028 return on tangible equity target from 12% to 14%+. This is what it looks like when a regulated business integrates AI into financial planning rather than keeping it in separate innovation labs. Read more
Insurance sector: only 7% have scaled AI beyond pilots
Deep dive from AI News on why insurance is struggling to move AI to production. Only 7% of insurers have scaled AI across their organisations, despite the industry absorbing over £100B in annual losses for six consecutive years. The success stories stand out: one major insurer deployed 80 models, cut complex-case liability assessment by 23 days, improved routing accuracy by 30%, and reduced complaints by 65%. The common thread? They fixed their data architecture first. Read more
Best multimodal models still can't crack 50% on visual recognition
New WorldVQA benchmark tested whether vision models actually recognise what they see. Asked for specific details (exact species, product names) not generic labels. Best performer: Gemini 3 Pro at 47.4%. Models are confidently wrong, high certainty even when making errors. If you're building AI that needs to identify specific things, test carefully. The confidence-without-accuracy problem is real. Read more
MIT: That viral AI graph everyone misreads
Deep dive on the METR "time horizon" graph that supposedly shows AI capabilities advancing exponentially. The common misreading: people think the y-axis shows how long AI can work independently. It actually shows how long it takes humans to complete tasks the AI can do. Very different. Error bars are huge. Almost entirely based on coding tasks. Even METR researchers say "You should absolutely not tie your life to this graph." Read more