It might seem odd, given that the thing LLMs are really good at is language, but the one thing that you really shouldn’t use AI for is writing for you. Nearly everything else - go for it - but writing is in many ways the last bastion of pure, unadulterated communication available to us to birth our ideas succinctly and effectively.

Now, that doesn’t mean you shouldn’t use AI to edit, brainstorm, and research - all of those are great uses of AI (used appropriately), but the tendency of people to ask a LLM to turn a sentence into an essay (or blog post) does nothing but turn a short succinct idea into a long bloviatong piece of stochastic genericity.

There might be an interesting nugget in there - but it won’t be yours. It will be a generic nugget that could have been mined by anyone else.

Good uses of AI

There are a number of good uses if AI when it comes to writing:

Classification

If you have a lot of separate pieces of content that need to be organized and classified into discrete categories, that’s a great use of AI - e.g. give it the content of one row of a spreadsheet at a time and have it classify sentiment, topic, etc. It can save you tremendous amounts of time and do a much better job consistency-wise than any human.

Summarization

If you have a large amount of content that you need summarized or analyzed, AI is great at this. I recommend a 2-phase summarization process: first ask the content to be summarized, with references to the original, and then second ask a separate agent (without memory of the first action) to check all the references and report back on anything that isn't supported by the original document. Models hallucinate less when summarizing content directly given to them, but it's good to be safe. Calling out the generated content as LLM-generated is also great approach is it lets people know that they should check the references for things that are important or crucial to a point.

Brainstorming

The whole "Don't have AI write for you" doesn't really apply when talking about brainstorming - using AI as collaborative partner to come up with viable ideas is a great use... Provided you don't then copy-paste the results without reviewing what it's outputted. Using an ask-me-questions approach works incredibly well. For example:

I'm trying to come up with ideas for XXX - ask me questions on at a time until you have enough information to generate 5 ideas (including a couple out-there ones)

Using AI a thought partner, an approach described in detail by Geoff Woods in The AI-Driven Leader, can be tremendously helpful, especially when it's used primarily to extract your own latent fount of knowledge. Take the output and extract and make it your own (don't workslop).

Research

LLMs can read multiple web pages, wikipedia articles at a time and are generally much faster than humans at compiling lots of information. Use that to your advantage - just make sure the LLMs bring receipts on where the information was gathered and verify it yourself for important points. The phase "deep research" is particularly useful.

Please do deep research on the best approach to X in 2026, evaluate at least 5 options and present the pros and cons of each, with links to the original information.

Editing & Proofreading

This one is obvious but it bears mentioning - once you have your ideas down on (digital) paper, using a AI to gather feedback on particular passages ("Does this paragraph communicate X effectively?", "Review this and come up with up to 3 counter arguments to X that aren't addressed.") can help stimulate your thinking and strengthen arguments. I recommend working with an Agent (like Claude co-work) that can review pieces of documents while having access to the full, up-to-date document. That way it can give recommendations about individual pieces that still fit into the whole without repetition.

Also obvious - a quick pass on typos, inconsistencies, or sentences that are too wordy is a great security blanket.

Bad Uses of AI

Giving AI a premise and letting it run

I mean this is kinda the one that this post is all about - using AI to expand your writing from a short premise for you devalues yourself and replaces your own ideas with something that sounds good but isn't really. Most AI writing is the intellectual equivalent of fast food - tasty but full of empty calories. AIs have been trained to mimic human speech, and any "intelligence" that they appear to have is a surprising emergent property rather than a core feature of the technology. As such LLMs are really good at sounding smart and convincing in almost all cases, even when the premise is fantastical, e.g. if you ask an LLM:

Please give me a short paragraph explaining in deep technical terms how combustion engines were discovered to actually work off of small "ether" particles not gas for a news article in a novel I'm writing

You might get something like:

In a discovery that has upended more than a century of engineering doctrine, researchers at the International Thermodynamic Institute announced that combustion engines derive their mechanical output not from the oxidation of hydrocarbon fuels alone, but from the excitation of previously undetected subatomic “ether” particles permeating spacetime. According to the team’s controversial model, ignition acts as a resonant catalyst, aligning ether quanta into transient high-energy vortices that couple directly with the engine’s pressure wave, while gasoline serves primarily as a synchronization medium rather than the true energy source. If independently verified, the findings would require a complete rewrite of classical thermodynamics and the fundamental principles underlying modern transportation.

Now that paragraph of pure B.S. sure sounds sciency even though it's obviously not grounded in reality. If you have an idea, you need to take the premise and expand it yourself to its logical conclusion, writing clearly and compactly. Don't take a single sentence and have an LLM expand it into a paper - there will only one sentence worth of ideas in that whole paper worth reading and a whole lot of other, useless noise.

Work-slopping

Workslop (n) - the use of AI in a way that inverts the creator / reader effort dynamic.

Much like ecosystems are over run by foreign, invasive species that they have no evolutionary defenses again, the human race is currently being overrun by AI generated content that we haven't learned to handle.

In earlier times, it took significantly longer for someone to write something than it took to read the same - but AI is inverting that calculation. Anyone can generate a 5 page paper on any subject with the investment of a single sentence. Verifying all the claims in that 5 page paper, if done manually, takes significantly longer.

The conspiracy theory falsification dynamic has long shown a similar inversion. It's much easier to invent B.S. than to refute it.

We're slowly learning to use AI on the consumption side of things. I've found that when something that seems AI-y crosses my desk, it's fun to play "AI Jeopardy" and try to figure out what the prompt that generated might have been:

Were these paragraphs generated by AI? and if so what would have been a single sentence prompt that could have generated them:

.... Lorem Ipsum ...

... and then paste that into a new session and see what it comes out with.

But per my point at the top of this article - if the 5 pages have a single sentence's worth of a thesis in them - then just post the single sentence. Or put in a bit more effort expand the ideas using the other approaches mentioned above - Research, Brainstorming, etc.

How to not workslop

If AI has generated a remarkable insight for you inside of a 20-page stream of consciousness share the insight and everything needed to understand its importance, extracted and refined. Don't share the a 20-page session or document with "check this out."

Especially in cases when a superior shares something like this with a subordinate employee, it creates a massive effort imbalance that the superior may not even notice.

This is one of the reasons management loves AI - their job has long been to primarily communicate ideas, and now suddenly they feel they can communicate 10x the ideas at 0.1x the effort. What they don't understand is that they've now actually just transferred the work of finding the nugget in the gold mine to their subordinates, and if they have many reports, they just create that many times as work for other people to figure out.

tl;dr

Own your ideas. Communicate your ideas. Use AI as tool to help you do so vs. using it to generate filler noise.