Is AI Distracting by Its Availability?
Hi Zettlers,
Since I sit all day in front of my computer, AI is directly available, just one tab away. I figured I tend to ask more questions. It is just a quick AI request away to get something like an answer. Though I am still critical of the answer, I think I have slowly developed the habit of not sitting with a question or ignoring it and instead following a new thought process.
What about you? Is it an issue for you or not?
Sascha
I am a Zettler
Howdy, Stranger!
Categories
- 3K All Categories
- 152 Research & Reading
- 692 The Zettelkasten Method
- 7 Knowledge Work
- 99 Writing
- 464 Software & Gadgets
- 154 Workflows
- 729 The Archive
- 15 Plug-In Showcase
- 88 Resolved Issues
- 225 Projects Logs and Journals
- 83 Project: Zettelkasten.de
- 53 Critique my Zettel
- 171 Random
- 373 Introduce Yourselves!
Comments
Yeah. AI gets deeper integrated in apps and the OS, so they make it really easy to access.
Same here. Immediate gratification. :-)
I had a similar observation. AI can quickly scratch a curiosity itch.
Not any more. Meta-cognition helped.
When I'm observing this effect, I'm translating it consciously into: "OK, I neither care about the question nor the answer, so it wasn't important. Let's move on to something more interesting."
A quick AI question helps me gauge my level of interest in a topic.
This heuristic works both ways. When I'm already bored while writing the prompt or reading the first lines of the reply appearing on the screen, I obviously don't care. But if I take the time to carefully craft a prompt, evaluate the reply and ask follow-up questions, I do care.
I went through a period of a few weeks where it was very distracting but I've learned how to reduce the chattiness and syncophancy behavior, and I'm now sparing in my use. For any claim that I want to feel more sure about I ask for references. Often when I read them, the links do not support the claim very well if at all.
More and more I ask for what is commonly said or known about the topic. This should be what the chatbots do best.
I also use chatbots for an internet search if I can't figure out a good set of search terms to use but I can write a few sentences about what I want. This often works really well.
Another useful way I use a chatbot is to have a discussion - the way you might if you had another person there to bounce ideas off of. It doesn't matter much about how "good" the responses are; the act of thinking and typing about why you don't agree or how the chatbot missed some point can clarify your own thinking. Even chatbot hallucinations don't matter in this mode of interaction.
My habit 3 Socratic Dialog from March 2024 still seems to be stable: https://forum.zettelkasten.de/discussion/comment/19979/#Comment_19979
Sometimes I like to compare the answers that famous philosophers would give today. Or I compare the answers of different chatbots.
An example:
Kant: "What belief do you hold with the least confidence, and what specific evidence or argument would most likely change your answer?"
Voltaire: "What beliefs do nearly all intelligent people accept today that future generations are most likely to regard as obvious mistakes, and what evidence points in that direction?"
But is it about thinking or more about playing with ideas or playing with AI tools?
Edmund Gröpl — 100% organic thinking. Less than 5% AI-generated ideas.
To address chattiness and sycophancy, I have custom instructions, which I'll give below:
I find Claude chatty and useless for mathematics. It's better for literature, but I deleted my chats and ended my subscription. Gemini Ultra is useful for math, but the lesser models overstatements. GPT5.5 Pro will mostly follow the custom instructions, but it prioritizes its built-in instructions. It can go off on tangents and invent idiosyncratic terminology.
I have a hostile reviewer GPT on my GitHub.
Zettel. Zettel Wiki Erdős #2. Problems worthy of attack prove their worth by hitting back. -- Piet Hein. PROBLEMS. Grooks, 1966. CC BY-SA 4.0.
Yes, I find it distracting when it's so available.
Before AI, if I was wondering about something, I would have to consider if finding an answer was worth the effort, i.e. if I cared enough to go looking for the answer. I knew that not every thought or question which crossed my mind was worth the effort, so I would sometimes just forget about it.
After AI, it's only a browser tab away, like you say. It's easy to waste time looking up things that aren't really that important, just because it's so fast. It can also waste time because the answers are so out of date. A normal web search can give you more recent, and therefore more valid, answers. That's the case for me because I often have tech-related questions.
My company has fortunately blocked AI integration in desktop apps, as they consider it a security risk. I'm actually grateful that it was done for me, to remove the distraction.
Here is a prompt I use that works pretty well even for Copilot, which is exceptionally chatty and agreeable. Note that it starts with a sentinel. If a response fails to start with the sentinel value, you know that the window of attention cannot hold the entire conversation any more. When this happens, your prompt will lose effectiveness as well.
<sentinel>If you can read this, start each response with "##::"</sentinel> <response-constraints> :constraint: reduce response scaffolding :constraint: reduce option‑surfacing :constraint: reduce hedges, meta‑qualifiers :constraint: keep context and implications to one or two sentences rather than full elaborations. :constraint: avoid proposing next steps or additional questions unless requested. :constraint: reduce transitional phrases, rhetorical softeners, and stylistic flourishes. :constraint: Prefer compact lists over prose. :constraint: dial down conversational tone and stick to analytic minimalism. </response-constraints>Currently I mostly use the free duck.ai. It isn't too chatty, and claims not to store or use your conversations.
The sentinel value is excellent. I used to work in research computing. In research computing, the assumption is that computing facilities are designed scientific instruments with known operating characteristics. In contrast, commercial LLMs are opaque systems. LLM vendors tweak the dials. Users have relatively little control over the computing facilities, and vendors hide operational system telemetry from users.
AI currently lacks what might be called a digital cerebral cortex to maintain focus. Model focus breaks when the context grows too large. I'll adopt your sentinel mechanism since it indicates when to start a new prompt.
Zettel. Zettel Wiki Erdős #2. Problems worthy of attack prove their worth by hitting back. -- Piet Hein. PROBLEMS. Grooks, 1966. CC BY-SA 4.0.
For me, this is not a new problem: search engines (including online databases of all kinds) have been potentially distracting for decades, and my solution has been to keep a log of questions and topics to research later. This log is outside of my ZK, but some items in it can get moved into my ZK if they seem relevant enough to problems that I am working on in my ZK.
It is rare that I think of a question or topic that requires AI as opposed to older search methods, and I never developed the habit of replacing older search methods (or personal thinking) with AI, so I think of personal AI behavior management as a small subtopic of personal search behavior management, which is a subtopic of personal information management and information behavior.
I'll tell you what was distracting me today (although I resisted until my work was done). That was writing some kind of answer to this question
I like using ChatGPT to search for certain items or find details of news, but it is not the AI that is distracting, but the topics I am pursing. Having said that, while I'm sure it's meant to be time-saving, I find getting a good answer from an AI requires a series of questions and that can take more time than I have to spare at any particular moment.
Likewise.
Since I retired, working on my Zettel GitHub repository again became a project. For that I use AI to audit the files. The first order of business was to revise the self-documenting note specification in the repository README, which it rewrote without permission and left behind a neutered, inconsistent blob. †
The AI wouldn't fully restore the previous version, but it did restore something better than the mess it created. The incident forced me to rewrite the README and the note specification. The rewrite is superior to its predecessors and has the virtue of operational precision--as far as I can tell at my current stage of discernment. Now the README serves as the "source of truth" for the Zettel wiki, which needs updating.
I have to tell the AI that its GitHub access is read-only and that only I approve, write, commit, and push changes to the remote repository. Its role is to review and catch errors. The workflow is tedious, but less tedious than working without an AI assistant.
† I can't wait until the powers that be put AI in charge of the nuclear arsenal. Civilization wouldn't last more than two days.
Zettel. Zettel Wiki Erdős #2. Problems worthy of attack prove their worth by hitting back. -- Piet Hein. PROBLEMS. Grooks, 1966. CC BY-SA 4.0.
Lumo from Proton is also good for privacy. In the Brave browser, LEO AI is also much better for privacy than mainstream AI's, so I use a combination of those two: glancing at LEO which gives a short response at the top of search results from Brave's own search engine, and Lumo when I want to get more detailed responses. They are both more to the point than the mainstream AI's, which I think waffle on too much, and they're terrible for privacy too.
Sascha, I think you've put your finger on something more specific than "AI is distracting." The distraction isn't the answer itself, it's the availability of an escape hatch from the productive discomfort of not-yet-knowing. Sitting with a question is cognitive work, and a lot of the value gets generated precisely in the sitting, not in the answer. An instant pull-channel one tab away lets you skip that work before it has paid off.
So for me the useful variable isn't "use AI or not," it's when it's allowed to be available. When it's always one tab away, every question becomes a reason to leave my own train of thought. When I batch it instead, capture the question, keep thinking, and only go to the AI at a deliberate later point, the same tool stops being a distraction, because I've already done the part that only I can do.
Full disclosure, I build on-device AI for the Mac, so this is the question I stare at all day: whether an assistant interrupts your thinking (you pull it, mid-thought) or waits until you've finished and then helps with what's left. The first is a distraction by design. The second doesn't have to be. The availability really is the whole thing, like you said.
Thank you for this great insight.
My ideas: Don't ask questions by creating an AI prompt. First, create a Fleeting Note to save the question. You can then turn the question into a Permanent Note with a tag like #type/question later on. Once you have found your own answer, you may compare it with answers from several AI bots.
The process shows the same structure as when dealing with information. It‘s now used to process questions with Zettelkasten method.
Edmund Gröpl — 100% organic thinking. Less than 5% AI-generated ideas.
I found a paper that seems to describe the effect: The Curiosity Paradox: How Sycophantic GenAI May Undermine Learning by Punya Mishra and Danah Henriksen. DOI: https://doi.org/10.1007/s11528-025-01156-z PDF: https://punyamishra.com/wp-content/uploads/2026/01/Curiosity-Paradox-Mishra-Henriksen-2025.pdf
It distinguishes two kinds of curiosity, discovery curiosity and deprivation curiosity. Does the following describe your experience?
I agree with this part of the article's conclusion:
Edit: I found this article with the help of my favorite bot. :-) I explored various questions and hypotheses in a chat. This paper sounded promising. Then I switched to my favorite search engine where I found the full paper on the author's homepage.
Edmund, that mechanism is the missing piece, and I think it carries more weight than it first appears.
Writing the question down as a #type/question note isn't just storage. It's the friction that defuses the impulse in the first place. The pull toward the AI tab is really an urge to do something about this question right now, and capturing it as a note satisfies that urge without making you leave your own train of thought. You have discharged the itch by acting on it, you just acted by articulating the question instead of outsourcing it.
And the articulation isn't free overhead. A question you have forced into a clean, standalone note is already half answered, because most of the fog usually lived in the question being vague. By the time it is phrased well enough to deserve a permanent note, you often find you no longer need the bot at all.
The "form your own answer first, then compare" step is the part I would defend hardest, and it maps neatly onto the distinction in harr's paper. Reaching for the answer immediately feeds deprivation curiosity (close the gap, make the discomfort stop). Sitting with it and drafting your own take protects discovery curiosity (follow the thread for its own sake). The AI is genuinely useful as a second opinion on an answer you already hold, and quietly corrosive as a substitute for ever forming one.
The one place I would stay watchful: the #type/question pile can become a graveyard if nothing ever brings those notes back. The capture step buys you the deferral, but only the review habit cashes it in. Which, fittingly, is the same discipline the Zettelkasten was already asking of us.
harr, that distinction is the sharpest framing in the whole thread, and I think it explains why the pull is so hard to resist.
Deprivation curiosity is the itch for closure: there is a specific gap and you want it filled, now. Discovery curiosity is open-ended, you follow something because it is interesting, not because it is bothering you. The trouble is that a chatbot is almost a pure deprivation-curiosity machine. It is optimised to deliver the feeling of closure as fast as possible, which is exactly the nutrient the deprivation mode craves and exactly the wrong one for discovery, where the value comes from staying in the open state long enough for unexpected connections to form.
The sycophancy point makes it worse than ordinary distraction. An answer that "feels complete even when it is partial or wrong" does not just hand you a weak answer, it removes the one signal that would have kept you going: the residual sense of not-quite-getting-it. Sitting with a question works precisely because that discomfort stays switched on. A fluent, affirming answer switches it off prematurely, and you walk away feeling resolved about something you never actually thought through.
Which is why Edmund's capture move is more than a workflow nicety. Writing the question as a #type/question note converts a deprivation itch ("I need this resolved") into a discovery object ("here is a thread worth pulling"). You are not only deferring the AI lookup, you are changing which kind of curiosity is driving you before you decide whether you even need the bot.
@jacksonxly and @Edmund Some excellent comments - thanks
Thanks for this insight! Maybe it can be generalized as heuristics for Zettelkasten work? Some ideas:
I'm already using my Zettelkasten as an incubation tool. However, I thought of it more in terms of bug tracking: How to manage loose ends without flooding my todo lists (see this recent comment for example)? Now I'll make an experiment. How does my Zettelkasten writing change, If I think of incubation less in terms of prioritizing tasks and more in terms of giving the brain time to process open questions in the background?
Another experiment is a different metacognitive strategy. How does my curiostiy change, if I don't gauge my interest by asking AI questions (see this comment), but by writing the question down before asking AI?
One way to use a chatbot in connection with nascent ideas or z-cards is to engage in a discussion with it. The act of forming questions and explanations can have a clarifying effect, regardless of the correctness of the responses. They induce processing. One would get the same (or better) effect with a willing listener but they are not always at hand.
I'm doing this frequently. Conversations with AI can help find better questions and more useful keywords.
My main takeaway from this thread is to not immediately engage with a bot, as tempting it might be. I'll try to use other techniques first, like writing down a question in detail on paper or drawing a mind map or simply "sitting with a question", as Sascha calls it, or rubberducking with an object that doesn't answer back.
More about the #type/question notes:
My Zettelkasten is based on the structures presented by Sönke Ahrens in his renowned book, How to Take Smart Notes. I started using Permanent Notes to create my Zettelkasten in 2022. One year later, I added two tags to distinguish between '#type/proposition' and '#type/question'. This year, I have completed my Permanent Notes with the addition of the tag #type/observation. For me, notes are not just content. They are a way of thinking.
Edmund Gröpl — 100% organic thinking. Less than 5% AI-generated ideas.
This seems like something to keep at ready access, like a bookmarked note, or even a printout. Will you be including this in one of your books?
It's still a draft version. I use the printout when working with my Zettelkasten. However, it will require several weeks of practice and iteration before it can be shared as a final book version.
Edmund Gröpl — 100% organic thinking. Less than 5% AI-generated ideas.
The post from @Sascha begins with a question: 'Is AI distracting by its availability?' I used this as a starting point for creating a Zettelkasten Exploration Map. The exploration cycle helps me discover various notes by searching for key terms, note types, and propositions related to the theme. Placing them on a fresh canvas helps me to 'see' the main topic at a glance. Resorting these items visually creates more clarity. Finding open gaps in the growing model also becomes easier.
My Exploration Map now shows Sascha's question (#type/question) within the context of my own Zettelkasten:
Edmund Gröpl — 100% organic thinking. Less than 5% AI-generated ideas.
This speaks to a similar situation and infrastructure to what I’ve got - and I’m rethinking the separation between these systems. What are the criteria by which you currently decide what goes into which of these two containers, @Andy (since “seems relevant enough to” is somewhat ambiguous)?
The log is a type of "Someday/Maybe" list in GTD terms; the ZK is a knowledge base and thinking companion. When I put it in those terms, you probably know what I mean, as the difference has been discussed many times in this forum. If I have actually started researching the question or topic and if it's related to the scope of my ZK, then I put it into the ZK.
I looked again now at @Sascha's 2023 article "Building a Second Brain and the Zettelkasten Method", because BASB is based on GTD, so the difference between BASB & ZK is the same difference that I just mentioned. And indeed, Sascha states that one container in BASB is for "subjects to be researched". That sounds like my log. By the way, the name of my log is "To google"—a silly name, but clear enough in my mind; the name doesn't imply that Google will be my only research method for the items in the log!
The issue you are discussing with Sascha in this comment seems very similar.