When I began writing a weekly blog in May of 2005, I expected that artificial intelligence would someday become one of the subjects I would write about. I didn’t expect that AI would eventually become my editor.
Writing each weekly post takes many hours. Before I write a single sentence, I spend considerable time thinking through the topic I want to explore. What do I want to say about the topic? Then comes the drafting, revising, rewriting, and editing. I’ve continued this routine for over two decades because blogging has become much more than a way to communicate ideas. It’s how I explore new subjects, deepen my understanding of emerging technologies, and keep learning.
From 2012 to 2020, edited versions of my posts were republished in the Wall Street Journal CIO Journal. I often preferred those edited versions to my original drafts, which was hardly surprising given the skill of the Journal’s editors. After that relationship ended, I missed having an experienced editor who could improve the clarity and flow of my writing while preserving my voice.
Several months ago, I wondered whether an AI chatbot could play a similar editorial role.
So I logged on to ChatGPT and wrote something very simple: “I would appreciate your help in editing the draft of this blog while preserving my voice.”
I pasted my draft into the chat.
Within seconds, ChatGPT returned a lightly edited version, explaining that it had focused on improving clarity, consistency, and grammar while preserving my analytical arc and writing style.
I was genuinely surprised by the quality of the edits.
I accepted many of the suggestions, rejected others, and remained fully responsible for every idea, every argument, and every final decision. The chatbot wasn’t writing my blog. It wasn’t generating ideas on my behalf. It was doing what a good editor does — helping me express my own ideas more clearly.
I’ve now been using ChatGPT in that role for several months.
What has surprised me almost as much as the quality of its editing is the way I interact with it. I don’t use it the way I use my email, calendar, web browser, search engine, or Wikipedia. Those are tools that I operate.
Instead, I interact with ChatGPT much as I would with a trusted human editor.
After finishing a draft, I generally begin with a request such as, “I would appreciate your help in editing the draft of this blog,” followed by its title.
The chatbot typically replies: “I’d be happy to help. Please paste the draft of the blog or upload it as a document, and I’ll edit it in the same style we’ve been using for your recent blogs.”
Over time, I refined the instructions I gave it. Rather than rewriting my drafts, I asked it to preserve my voice and conversational style, improve the clarity and flow of my writing, tighten repetitive passages, and point out places where the narrative could be made more compelling.
Within seconds of uploading a draft, I receive thoughtful suggestions: alternative titles, stronger openings, improved transitions, and more effective conclusions. After a few additional exchanges, ChatGPT produces a revised version of the blog. I then carefully review every suggestion, deciding which edits to accept, which to reject, and how I ultimately want the essay to read.
This process naturally led me to ask a broader question.
Why have I chosen to interact with ChatGPT as though I were working with a human editor rather than simply operating another software application?
A recent New York Times guest essay by author and entrepreneur Keith Ferrazzi, “We Interviewed A.I. Agents. They Nailed Your Co-Workers’ Worst Habits” helped me understand why.
Ferrazzi argues that one of the biggest obstacles to AI adoption has relatively little to do with the technology itself, which continues to improve rapidly. “The real sources of friction are humans,” he writes. “Helping employees get the most out of A.I. tools requires them to learn how to be better collaborators and better managers. The same lessons that make us better at working with people can make us better at using A.I.”
His company, Ferrazzi Greenlight, has been studying how human-AI interactions affect the quality of AI-assisted work. Their research found that AI agents perform best when treated as collaborative peers and worst when they are micromanaged or given vague instructions.
Humans have long relied on colleagues to infer unstated assumptions and fill in missing context. AI cannot do that nearly as well.
Ferrazzi’s advice therefore sounds remarkably familiar to anyone who manages people: “Brief your A.I. agent the way you would brief a new hire. Tell it not just what you want done but why it matters, what constraints are in play, what success or quality looks like and what pitfalls to avoid.”
Managing AI well, he concludes, looks surprisingly similar to managing people well.
My own experience has been very much the same.
Working with large language models has also made me appreciate something that, as a technologist, I had not previously given sufficient attention: the central role of language in human interactions.
Several years ago I read a 2020 paper by linguistics professors Emily Bender and Alexander Koller that explained the difference between four important concepts.
- Form is the observable expression of language, whether written, spoken, or signed.
- Communicative intent is the purpose a speaker hopes to achieve.
- Meaning is the relationship between the form of language and the communicative intent it is intended to convey.
- Understanding is the listener’s ability to correctly interpret that intended meaning.
Those distinctions changed the way I think about my interactions with an LLM in what’s become known as prompt engineering.
To me, prompt engineering is fundamentally a language problem rather than a programming problem. The goal is to communicate my intent as clearly as possible. The better I explain what I want — and why I want it — the better the chatbot understands the task.
That’s one reason I naturally found myself interacting with ChatGPT much as I would interact with a human editor.
Despite understanding, at least in principle, how large language models work, I’ve continued to be surprised by the quality of their editorial suggestions.
I understand that an LLM generates language by predicting statistically likely sequences of words rather than by reasoning from personal experience. But I found myself wondering how a system based on statistical prediction could consistently improve the clarity, organization, and flow of my writing — activities like connecting dots and sense-making that seemed to require something closer to human judgment.
I had been discussing these questions with my friend and colleague Allan Tate, so we asked ChatGPT directly how it does it.
Its answer was both thoughtful and illuminating.
“The short answer is that I do not connect dots the way humans do. Humans connect dots through lived experience. A person accumulates memories, emotions, successes, failures, relationships, stories, and observations over decades. When faced with a new situation, those experiences influence what feels familiar, surprising, important, or suspicious.”
“I have no life experiences.”
“Instead, I have been trained on a vast amount of human-generated text containing countless examples of people connecting dots. Books, articles, essays, research papers, debates, biographies, histories, and conversations all contain patterns of reasoning. During training, I learn statistical relationships among concepts, ideas, events, and arguments.”
“When presented with a new situation, I search for patterns that resemble patterns I have previously encountered.”
I found that explanation very helpful.
Perhaps the most surprising lesson from my experience has not been that AI can edit prose remarkably well. It has been that the best way to work with AI resembles the best way to work with thoughtful people: communicate clearly, explain your objectives, provide context, define what success looks like, and remain responsible for the final judgment.
Large language models are certainly powerful technologies. But they are also introducing us to a fundamentally new form of human-computer interaction — one based less on commands and menus than on conversation, collaboration, and language itself.
We’re only beginning to understand what that means.
As for me, I know that my AI editor is not human. But after months of working together, I remain genuinely grateful for the thoughtful editorial assistance it has provided. It hasn’t replaced the hard work of writing. Like any good editor, it has simply helped me become a better writer.