# AI in Personal Knowledge Management AI is having a moment in the personal knowledge management (PKM) world--especially in the Obsidian community. Every other week there’s a new plugin, workflow, or hot take promising to _supercharge_ your vault: auto-summarize everything, auto-tag everything, auto-connect everything, maybe even auto-think for you while it’s at it. I do not get the appeal. I mean, I get that people want to learn more, write faster, and build a high-quality vault, but AI does not improve any of these from what I have seen. I don't mean this to sound anti-AI and anti-progress. Much like John Henry vs the steam-powered drill, I think it's futile to be completely anti-AI at this point. It is a useful tool if you understand what it's good at. Hint: it's not good at the majority of products it's advertised in currently. AI has nearly become synonymous with automation when we've already had working forms of automation for decades. ## Ingesting Articles: Garbage In, Polished Garbage Out One of the most common AI workflows is dumping articles, PDFs, or web pages into a model and asking it to summarize or extract “key insights.” The problem is that _ingestion is already the work_. When you offload this step to AI, you skip the friction where learning actually happens: deciding what matters, what’s new, what contradicts existing knowledge, and what’s worth keeping. AI summaries often sound clean and confident, but they flatten nuance, miss context, and quietly introduce inaccuracies that are hard to detect unless you already understand the material--at which point, you didn’t need the summary. You end up with notes that look useful but don’t connect to anything you actually _know_. ## Writing Notes ≠ Learning Notes Using AI to write your notes for you is even worse. Notes are not a storage medium; they’re a byproduct of thinking. When AI writes the note, it performs the synthesis _instead of you_, and you’re left with text that you recognize but don’t internalize. This creates a dangerous illusion of understanding: the vault grows, but your mental model doesn’t. If your notes feel impressive but you can’t explain them without rereading, you have not learned anything. You have just wasted your time. This is similar to all the effort Obsidian content creators put into making [[habits article |impressive, over-engineered dashboards that are ultimately counterproductive]] to actually getting things done. ## Creative Writing I admit this is one use case of AI that hits more of a moral objection for me than a performance one. Unlike the other use cases where AI is counterproductive, it is actually deceptively good at certain forms of art. People already cannot distinguish between the [prose of famous authors and AI](https://completeaitraining.com/news/readers-cant-tell-ai-fiction-from-humanand-they-prefer-the/). My moral objection to it is the "creative" part. AI is incapable of being creative. It has a vast knowledge base to draw from and ultimately copy, arguably plagiarize. There is no intention or soul behind it, and using it without disclosing it is essentially lying about what you produced. There is an implicit social contract that what we create actually came from a human, and this erodes that creative trust. For me there's another element about creative writing being a hobby. I, and many others, do it because I enjoy it. Why would I offload the fun parts to a language model? This is a skill I want to get better at, and the only way to do that is to do it myself. That doesn't mean I can't use tools to catch errors and potentially do some editing, but I want to develop my voice, not a regurgitated amalgamation of others' voices. There is a lot more I could say on this topic, but numerous people with a more vested interest and better communication skills have already chimed in. One of my favorites being [Brandon Sanderson](https://www.youtube.com/watch?v=mb3uK-_QkOo). ## AI Should Not Know Your Vault Better Than You Do One of the selling points of Obsidian is its linking capability and other capabilities to connect notes together. There are numerous plugins attempting to make AI manage this for you. Sure, there are some cases where connections might be missed, but I think if this is a big enough issue to require tools beyond what default Obsidian provides, then there are other issues in your methodology. My vault is over 6,000 notes at this point, which is on the high end relative to the average user, and with [[Navigation article | consistent file naming]] and frontmatter templates, I have had no problem connecting related files together. Another thing is that AI is great at pattern matching, but terrible at understanding _why_ a connection matters. It will happily link notes that share vocabulary but not meaning or suggest relationships that feel insightful until you actually follow them. Worse, once you start trusting these suggestions, the vault slowly shifts away from _your_ mental model and toward the model’s statistical one. ## Infrastructure and Organization: Where AI Is Actively Risky Using AI to reorganize folders, rename notes, restructure metadata, or “clean up” a vault is where things get genuinely dangerous. AI does not understand your constraints, your exceptions, or the historical reasons your vault looks the way it does. It will confidently make large-scale changes that are internally inconsistent but semantically wrong. For heavily repeated tasks under known rules, scripts are _vastly_ more reliable. They’re deterministic, testable, and reversible. Under the same set of conditions, they perform the same *every time*. I have tested the same tasks with AI and the one constant was that AI would always miss files and make mistakes. If anything, AI’s role here should be limited to **helping you write the script to perform the task**, not executing the change itself. The same applies to mass edits of content, metadata, or structure. If you’re changing hundreds of notes, you want guarantees: - clear inputs - predictable outputs - the same behavior for the same set of conditions *every time* - no handholding AI provides none of that. A simple Bash, Python, or JavaScript script will outperform it every time and won’t hallucinate halfway through your vault. This is the only use case for AI in personal knowledge management I have seen so far: script writing and tool building. It should not be performing actions on your vault.