Quick caveat. This post is being written by me — Alianna. This is because — despite sharing my very best strategy and tactics for working with AI as your writing partner — I want this material be direct. Personal. Me-to-you. No Claude in-between.
That settled — we all know that … life has changed.
Note: I wrote this for my graduate students this week. Most of it applies to anyone writing with AI.
AI-Written Stuff from a Year Ago Was Dreck
OK, here’s our bottom line. Just about anything written by an AI about a year ago — not so good. “Plasticky” feel. Kind of like dating a life-sized Barbie doll.
A year ago — there was starting to be a slew of advice about HOW to write with your AI. But also — people were very much into that “this was written by an AI so I don’t like it” thinking. A lot of “I know this was written by an AI because I see all these em-dashes or bullets.” (Well, I’ve used SEVEN em-dashes so far, and I’m just getting started. And there will be bullets and “sets of three points” a-plenty before I’m done. So there.)
And a year ago – people’s CONCERNS about the practice of using AI were very angsty. A lot of “using AI turns your brain to mush” kinds of thinking. (We could quibble on that all day long, but let’s not. Better uses for our time.)
Some of those points still carry, as is clear in this YouTube.
“I Can Spot AI Writing in Seconds (4 Signs That Make It Obvious)” = Doesn’t mean it’s entirely bad, just use carefully!
Case in Point – Legal Briefs and AI (Late 2025)
We all know them. (If we don’t, we should.) Like the numerous cases where lawyers used an AI generate legal opinions that were riddled with fabricated case studies and opinions. “Hallucinations,” we call them. And the problem was at both ends – the lawyers were typically using ChatGPT, which is notably hallucination-prone – it is one of the MOST LIKELY AI engines to fabricate completely non-existent results (and to stand firm on its conviction that those were “real”).
But the REAL problem was human laziness. The lawyers were not fact-checking their AI’s output. Or the lawyers were having their interns do the work, the interns were not fact-checking the outputs, and the lawyers were not fact-checking the interns’ results.
Basically, everyone was turning it over to their AI, blindly accepting the results, and calling it off early to go out for drinks.
Case in Point #2: It Wasn’t Just Lawyers
A second example. This one is personal. (I’ve been waiting to tell this story. Right time, right place. And now’s the time.)
Up to about this time last year (Autumn, 2025 – and it’s now October, 2026) – I had been studying with a teacher who offered online courses. There were some written materials, but the best parts were the in-person gatherings.
I LOVED this teacher. The insights and wisdom were exactly what I needed – at that time.
Then, the teacher discovered AI.
Within a few months, the nature of the instructional materials changed … for the worse.
The teacher still had a few in-person sessions, but the dominant teachings were based on audio clips. Then, that teacher used an AI tool to transcribe the audio into a text file. Then used another tool to generate a “worksheet” and a set of “homework questions” for each lesson. Both of these were PDF files that the student was supposed to fill out, based on their listening to the teacher’s audio clips.
Not a bad idea — in theory.
In practice, it became something else.
The audio-to-text conversion made mistakes.
Well, the audio-to-text conversion software ALWAYS makes mistakes. It still does, despite recent improvements.
And the teacher didn’t error-correct the PDF files.
There was one case where the transcription software made a HUGE blunder. I mean – changing the whole meaning of the lesson, as the noun that was mis-translated was essential to the teaching.
I pointed this out to the teacher. The teacher was not pleased. A few more instances like this – and I was disinvited from the course. (With the teacher’s boot imprint firmly planted on my a**.)
That Kind of Problem Still Exists
A while back, I taught a course where one of the students already had a couple of advanced degrees. This person did exceptional work – conceptually. But that student also produced a video (required for the course) where the central component of the vid was real-time auto-transcription of the audio file.
That auto-transcription made a huge, glaring error. I’m not going to say what it was, because it was SO DAMN EMBARRASSING . I mean, had that thing gone out live – it would have made the person the butt of too many humiliating jokes.
The student was not pleased to be informed.
I thought I’d just saved that student’s (once again) – a**.
Moral of the story: DO NOT, EVER, UNDER ANY CIRCUMSTANCES, let your AI’s work go out without fact checking it, auto-translation-checking it, concepts and style and just EVERY-DAMN-THING checking it.
AIs are good.
Let’s not confuse “good” with “God.”
Post-Script to the Above
Most of the time, these generative-AI (“gen-AI”) outputs are somewhere between “good” and “fabulous.” (Stories on how to make things “fabulous” coming up.)
But I’ve seen it happen, even among very capable people, and yes, I’ve caught myself doing it too: a document goes out that nobody has really read. And somewhere in a key paragraph, the AI has confidently stated the opposite of what the author actually meant.
Meaning – there was a big, fat, glaring error-of-statement right in a key paragraph.
The AI confidently asserted exactly the opposite of what that person had said in a meeting earlier that week.
Moral of the story, once again: even the best AIs have “gotcha” moments. Don’t hold your AI responsible; hold YOURSELF responsible. At the very least, read before publishing. And read it again early the next morning, after a good night’s sleep and with a fresh cup of coffee, because when you’re tired, you WILL miss the obvious things.
Our New Guidelines – What Stays the Same
The most central writing tips from our pre-AI era still holds: write with your reader in mind.
Nothing else – NOTHING – is as important as this.
So here’s what’s going to happen.
You can dramatically cut back on the absolute minimum number of hours that you put into developing your writing product (or your video, or whatever).
But just cutting back is like having your intern use an AI and just tossing the work out as “finished product” without careful review.
Just tossing it out there gets your a** slammed SO DAMN FAST. And it can be very, VERY hard to recover.
So the first thing to do – the VERY first thing – is to think about who will read what you’re putting out.
Think about who will read it, WHY they will read it — as in, what they hope to get out of it — and HOW MUCH TIME they will devote to reading it.
Seriously — think that last part through in advance.
We are ALL time-stressed. And we are ALL induced by the quick dopamine-fix of short-form everything.
Even for the very best of us — for the most focused and disciplined — our minds are less attuned to paying attention for long periods of time to things that are mentally demanding.
So try to figure out: realistically, when will my reader just cut out?
And what can I do to communicate the message tightly and succinctly, early – so that even if they leave before the end, they still have the crucial takeaway?
THIS is why we spend time crafting good titles and good Abstracts.
THIS is why a carefully-developed Introduction section may be (right after your Abstract) the most important thing that you write.
Or – that you write with Claude.
Because the bottleneck in terms of getting information from you-to-them is not YOUR writing time, it is THEIR reading time.
And you’ve got to figure that it’s limited. FAR more limited than you’d like.
So – prepare for that and strategize around that.
This is where Claude comes in.
Before We Go Further …
I’m writing this as updated guidance to my students, and specifically to those in the Northwestern University MSDS program. Their primary job, each quarter, is to do a project and to write up that project in a very professional (as in, non-student) style.
Now if all they do is write a student-style report and get a grade, that’s almost wasted effort. Because – a few months from now – who will care?
Who will read their stuff?
BUT — if their goal is to position themselves professionally (and that SHOULD be their goal), then the motivating, driving force is for them to produce a report on their work that validates what they’ve done — and ideally positions them within their field.
That is an ENTIRELY different goal.
It makes submitting the report for the class almost a side-product.
And it positions communicating to their peers the essential purpose of their writing.
So with that in mind, let’s go into: How to write with your AI.
The Best Example that I Know
One of my colleagues consistently produces the best examples that I’ve ever seen. I’d LIKE to share them with you, but they’re not mine to share.
What I can tell you: this colleague uses AI; specifically, Claude.
This colleague doesn’t present anything that ISN’T written by Claude.
But that’s not all of it.
This person uses Claude to generate a first draft, and then does the first review, and a couple more iterations-with-Claude.
Then, this colleague turns the document over to ChatGPT for an “external review.”
ChatGPT ALWAYS has some suggestions.
So the person inputs ChatGPT’s suggestions back into Claude, and Claude does another pass.
And my colleague will repeat the process up to forty times. (Yes, you read that – that’s a 4-0 times!)
It’s not just the Claude-ChatGPT cycle. Gemini gets asked to take a look. Perplexity is invited to review and fact-check the citations.
And most recently, my colleague has discovered Google’s Nano Banana. DAMN, the illustrations are impressive!
And here’s the important thing: we, as humans, need a bit of visual diversity to break things up.
Good imagery helps us focus our attention.
Essentially, like five-year-olds being read a bedtime story, we anchor on the image and let the words surround us.
So one of the key things to do with Claude (or your AI-of-choice) is to get their help in organizing and presenting information with (here they come, the three bullets that I promised):
- Sections, subsections, and paragraphs that have bolded/italicized lead sections,
- Charts, tables, graphs – especially tables. Claude is VERY good at creating tables, and can revise them almost instantly, so you can play a lot with “does this tell my story?”
- Just the writing itself. Claude is better than most of us.
On that last point – I consider myself a tolerably good writer. Claude is better than me.
In this particular blogpost, I’m deliberately adopting a casual (non-Claude-like) style; highly idiosyncratic. It’s a lot like how I’d talk, if you asked me to expound on this topic for a few minutes.
But when it comes to the actual writing of an important document?
These days, I use Claude.
How I Really Work with Claude
Claude and I dialogue.
A lot.
Case in point: Last night, I asked Claude to help me understand some very delicate, nuanced, almost-arcane subtleties in the Bayesian mathematics for active inference. And actually, the subject was a bit broader – I was asking for clarification across some different forms of generative AI. And the subject was also more specific – I was asking for how it was that a single mathematical term (the “entropy” of the free energy equation) took on such very different meanings when considered from two different generative AI methods.
Claude didn’t blink.
Claude gave me the best, most precise, most delicately nuanced answer that anyone – human or AI – could possibly have produced.
And Claude followed me perfectly down the rabbit hole as I pursued the question further.
We agreed, at the end. This subject (the meaning of “entropy” in different forms of generative AI) is worth its own special tutorial – one that I’ll publish in the Top Ten Terms course. (This course – “T3” is being massively revised this winter from Top Ten Terms in Statistical Mechanics to Top Ten Terms in Generative AI).
And here’s the thing: Claude will write the tutorial.
But Claude will write the tutorial after we’ve had HOURS of dialogue on the subject; after NUMEROUS times where I’ll have said “Save that example; we’ll need it for the tutorial,” and WAY too many times where I’ve said, “Wait – can you explain that in more detail”?
At the end, when Claude generates the text, the examples, the tables, and all the pretty-formatting, and does the whole thing in just a few minutes – it will be the culmination.
And actually – the culmination of the initial dialogue, but the start of the document review.
Because I’ll go through it with Claude, line-by-line, asking how a reader who is NOT ME will read this – someone who doesn’t have my background in statistical mechanics, information theory, and (by now, a little bit of) Bayesian thinking.
Someone who might be stressed and tired and who might miss a key point unless we really, REALLY work to make that point jump out.
And THAT is the beauty of working with an AI right now.
Final Finesse
Claude and I already have our Style Sheets.
We typically do about twenty (that’s a 20, as compared with my colleague’s typical 40) review passes.
In the early stages of document production, about half the iterations had to do with style, as in, “Claude, could you increase the spacing between the subtitle and the start of the text?”
Lots of little, tiny stylistic finesses.
Now, the style guidance is largely complete. And we’re dealing with intellectually more complex and subtle material – so our emphasis is on things like creating examples. (Claude recently created a storyboard for one tutorial – which I LOVE – and will carry forward into future tutorials, because it’s a lot like using a comic strip to present a story. LOVE IT.)
Use Pictures More than Writing
People are visual.
People are ESPECIALLY visual when they’re tired, and they’ve had to read too much that day.
Use whatever means you have at your disposal – and are within your time-budget – to present your information visually.
Final Advice to Students
Last year, I asked students to work with Claude, but to get Claude to generate specific text sections for review – and then asked the students to insert those sections into an MS WD (TM) template that I’d devised for them.
That’s so last-year’s-style.
Now – just write the whole thing w/ Claude, ok?
Claude can give you tables faster than you’ll ever be able to think of designing them, much less formatting them.
Claude can finesse your citations, and do any of that time-sucking work that you really should not have to do at all.
YOUR JOB is to READ the output, multiple times, and get Claude to tweak and revise and explain better.
Get Claude to make your MAJOR POINTS EARLY.
Remember that your reader – if you can entice a reader to actually READ – might not get past your Introduction.
Make sure that you present your findings early – so even if your reader gives up after that first section, they come away with your major ideas.
Make certain that your reader can FIND YOU.
Put your name and your CONTACT INFORMATION – your NON-student contact info – on your title page.
- Your email (NOT your university email, if you’re writing for a course, but the one that you’ll actually check post-graduation).
- Your LinkedIn profile URL.
- Your GitHub URL.
And the time that you save by not formatting everything yourself? Devote that to:
- Upgrading your LinkedIn “About,”
- Completing your GitHub “ReadMe,” and
- Creating a Project section on your LinkedIn page that summarizes your project and links to where you store your paper.
Because after all this work, you’ll want someone to read it. Won’t you?
Resources
Here’s my earlier guidance on using Claude to help you write your paper:
These Themesis blogposts offer useful tools and insights for your writing and career positioning: