What Did I Delegate? AI, Writing, and the Economics of Assistance (*)
(*) NOTE: This article talks about the use of AI and omits the environmental cost. It is in no way intended to dismiss the very real issues that surround the physical and planetary costs.
This article is part of the “It All Makes Sense Now” series:
A series about patterns, connections, hindsight, and those moments when scattered pieces suddenly form a bigger picture.
There is a great deal of discussion at the moment about whether writers should use AI.
I think that may be the wrong question.
Writers have always used tools and other people to help them write. Spellcheck checks our spelling. Grammar checkers identify awkward constructions. Editors reorganize manuscripts, question arguments, suggest cuts and sometimes propose entirely new wording. Research assistants find sources. Librarians help us locate information. Colleagues read drafts and say, “I don’t think you’ve actually made the point you think you’ve made.” Friends listen while we talk through an idea until we finally figure out what we were trying to say.
None of this normally causes us to conclude that the writer is no longer the writer.
Generative AI complicates things because it can potentially do all of those things—and it can also write the entire piece.
BUT: those are not the same use.
How I use AI
I use AI extensively in my writing, and it would be dishonest to pretend otherwise.
But I use it primarily as an organizer, filter, editor, sounding board, and research assistant, rather than as a generator of ideas.
A great deal of my writing begins in conversation. I notice something. I start pulling on it. I connect it to something else. I argue with myself—or, these days, with an AI. I reject interpretations. I refine distinctions. Somewhere in several thousand words of conversation, an article starts to emerge.
At that point, I may ask the AI to gather together the things I have said and organize them into a coherent draft.
The AI may generate the sentences in that draft, but that doesn’t mean it generated the thinking behind them.
Other times, I write the first draft myself and use AI much more conventionally: What’s unclear here? Where have I repeated myself? Is the structure working? What am I assuming that I haven’t actually established?
And sometimes it notices something I didn’t.
That happens with human editors, too.
Recently, while working on a short piece about silence and trust, AI suggested a sentence to capture a distinction I was making:
“This silence does not demand anything from me.”
That wasn’t quite right.
I changed it to:
“This silence does not demand payment from me.”
That one word mattered. I wasn’t merely talking about a silence that asked nothing. I was contrasting safe silence with weaponized silence in which ending the silence carries a price: accommodation, compliance, apology, attention, making oneself smaller.
The AI helped me see the place where a sentence was needed.
I decided what the sentence meant.
That is fairly representative of how I use AI.
AI can indeed write the whole thing.
Of course it can.
I could type:
Write me a moving 1,000-word essay about silence at a Quaker Meeting, connecting it to trauma, trust and faith.
and receive something quite plausible in seconds.
That would be a very different process.
This is why the question “Was AI used?” is not especially useful.
Spellcheck is a form of machine assistance. Grammarly is a more sophisticated form of machine assistance. Developmental editing is assistance. So is help in finding research to support an idea. Asking an AI to organize a long conversation into an article is assistance.
Asking it to decide what you think and write the article for you is something else.
There isn’t a nice clean line separating these activities. There is a continuum of delegation.
With that in mind, perhaps a more useful question is:
What did the writer delegate?
Did I delegate spelling? Sentence construction? Organization? Research? Criticism? The argument? The ideas themselves? The judgment about which ideas matter? The entire act of composition?
Those distinctions tell us far more about authorship than the simple fact that an AI was somewhere in the process. It is more like distinguishing between an editor that helps us tighten our writing, and a ghost writer who writes FOR us, and we simply attach our name to the work. That continuum has existed for about as long as we have written – and perhaps even before that when our stories were largely oral.
And the question of attribution and of full disclosure have been concerns for the entire time we have been storytelling in any form.
The difference now is not how much help we got from other humans, but how much help we got from AI. Yes, AI makes it a lot easier to “cheat” at writing, but the cheating is not the new part.
AND, the use of AI not only makes things easier, but can also make things more accessible, which brings us to…
The Economic Elephant in the Room
If I had unlimited funds, I would be quite happy to employ human editors and research assistants. I would prefer that.
I don’t.
Neither do most writers, and that matters.
Successful writers, academics, journalists and other professionals have long had unequal access to intellectual support. Some people can pay a developmental editor to spend hours working through a manuscript. Some work in institutions that provide research assistants, librarians, graduate students, editorial staff, administrative support and subscriptions to expensive databases. Some can hire someone to check hundreds of references or track down an obscure source. Some can pay for successful writers to review their drafts or teach them how to hone their skills.
Others do all of it themselves—or go without.
We talk a great deal about the digital divide: unequal access to computers, networks and information technologies.
AI may expose another divide that has been there all along: the assistance divide.
People with money or institutional support can buy other people’s time and expertise.
People without those resources generally cannot.
AI does not eliminate that inequality. Paid systems still exist. Human experts remain valuable, and in many circumstances they are substantially better than AI. AI makes mistakes. It can fabricate information, flatten distinctive prose, miss nuance and produce wonderfully confident nonsense. Using it well requires knowledge and judgment of its own.
But it can make some forms of assistance available to people who could never afford to purchase the human equivalent. That’s important.
A writer who cannot afford a developmental editor can still ask for an analysis of a manuscript’s structure. A student without a research assistant can get help constructing search terms and locating relevant literature, while still having time for other classes, and maybe even their personal lives. Someone struggling to organize a complicated argument can talk it through for two hours without worrying about whether they can afford another two hours.
A person with a disability, limited energy, or difficulty turning thoughts into conventional prose may be able to delegate some of the mechanical and organizational work while retaining the thinking that matters to them. Someone struggling with emotional overload can express their messy ideas, and have them turned into something that is clear, and logical without sounding like they are falling apart.
That doesn’t make AI and human assistance interchangeable.
It does mean that some kinds of intellectual scaffolding that were once disproportionately available to people with money, employment, institutional affiliation, or unusually patient friends have suddenly become much cheaper.
I think that’s worth talking about.
Assistance still requires judgment.
There is an important catch.
Delegating work doesn’t necessarily mean delegating responsibility.
If I employ a research assistant who gives me a bad source and I publish an unsupported claim, the claim still appears under my name.
The same applies when my research assistant happens to be silicon.
I routinely ask AI to investigate claims I want to make. That does not mean I regard whatever it produces as evidence. I want sources. I want to know whether those sources actually support the claim. If the evidence contradicts what I thought, then my claim needs to change.
Similarly, when AI reorganizes my writing, I remain responsible for deciding whether its organization represents what I meant. Sometimes it doesn’t. Sometimes its version is prettier than mine and less accurate. Sometimes it smooths away precisely the odd little thing that made the piece sound like me. Sometimes it confidently explains what I “really mean,” and I get to say, “No. That’s not what I mean at all.”
That ability to say no is important.
Perhaps one measure of authorship is not simply who generated the words, but who has both the authority and the capacity to recognize when they are wrong.
Show your work
I’ve been thinking that one useful way to make my own practice transparent would be occasionally to publish the original material alongside the finished work.
When an article begins with a draft I’ve written myself, I could make that draft available.
Readers could see what changed. They could see what I wrote, what the AI helped reorganize, what was removed, what was added, and what survived intact.
For pieces that emerge from long conversations, that’s harder—nobody needs 18,000 words of me wandering around an idea before discovering what I think—but I can still describe the process.
That seems more informative to me than attaching a label saying AI USED.
It shows what the AI actually did.
So yes, I use AI
Quite a lot.
I use it to organize, filter, search, challenge, summarize, restructure and word-wrangle. I use it to help me find research and then expect to be able to inspect the evidence. I use it to notice connections and problems. I sometimes use sentences it proposes. I often change them. I frequently reject them.
And sometimes, after several thousand words of discussion, I ask it to take everything I’ve said and turn it into something another human being might actually want to read.
What I try not to delegate is the part that matters most to me:
- What do I actually think?
- What am I trying to say?
- Is this true?
- Is this fair?
- Is this mine?
Those decisions remain my responsibility, and mine alone.
So when I encounter the increasingly common question, “Did you use AI to write this?”, my answer is yes—but I don’t think that’s a particularly interesting question.
The more interesting one is:
What did you delegate to it?
That is a question I am quite happy to answer.
In the end, it still comes down to the individual author’s integrity: what they delegate, what they claim as their own, and how honestly they represent the process when that distinction matters.
? From the Sausage Factory
If I’m going to argue that “Did you use AI?” is less useful than “What did you delegate to it?”, it seems only fair to show you what I mean.
This article didn’t begin with a prompt asking AI to write an article about AI. It began with these thoughts, expressed over the course of a conversation:
Is the difference between Word’s spellcheck, Grammarly and generative AI really one of kind, or largely one of degree?
I mostly use AI to craft articles and emails that result from things I have already said. I use it as an organizer and filter much more than as a generator. I also use it to find research relevant to claims I want to make.
If I had the money, I would happily hire human editors and research assistants. Could disciplined use of AI help bridge not just a digital divide, but an economic divide in access to assistance?
Maybe the useful question isn’t “Did you use AI?” but “What did you delegate to it?” And perhaps I should occasionally make my original drafts available so readers can see for themselves what changed.
What happened in between: AI helped me separate different kinds of assistance, and our back-and-forth produced the idea of looking at AI use in terms of delegated compositional agency. When I described my own process, another distinction emerged: generating prose is not necessarily the same as generating the intellectual content of that prose. I introduced the economic-access argument; AI helped develop it into the idea of an “assistance divide.”
AI then organized those ideas into the article you just read, proposed wording and examples, and gave me things to accept, reject, modify, or argue with. And THEN, I proofed it, made some additional changes and added a few things.
So yes: AI generated a substantial amount of the prose.
And now you can also see what actually went into this particular sausage.
