One day, a million robots fell from the sky, offering themselves up as the solution to a problem. I didn’t ask for these robots. Neither did you, probably. And I didn’t see a problem that needed solving, either. Still, once they were here, many people stopped to wonder how to use them “well” – or what “well” might even mean. The robots, after all, were everywhere.

Now, just a short while later, it is impossible not to use AI any longer.  I can’t resolve the ethical dilemmas – the impact on people, on the workplace, and the environment. But I can be up front about how I’m using it. 

So this document is my first effort to be transparent about my own evolving relationship to this new technology. It does not define institutional ethics or rules, or the rules of my office and my team. It is purely personal – not in the sense that it stays private, since much of what follows concerns my public and administrative work, but in the sense that no one but me is bound by it. I do not expect my staff or colleagues to adopt these terms, and I would not presume to require it of them. This is a record of the conditions I have set for myself, conditions I have applied to a job that is, admittedly, not private at all. A job that is often collaborative, too.

The pleasure principle

For any writer – for anyone, I mean, who makes their living through prose, or whose career advances because of what they write – there is no pleasure in the use of AI for composition. None. There is no rush when you make the right connection, no exaltation when you discover the perfect turn of phrase, and no joy when you execute the kind of prosaic maneuver that is so perfectly familiar, so uniquely you, that it becomes a love letter to yourself and to your readers.  

For that reason, any piece of writing that reflects my scholarly persona as an historian with an academic appointment, or as a writer with a name and a reputation, and that might be used in any assessment of my scholarship for tenure and promotion, that is offered to readers as my personal voice as a scholar or a sometime memoirist, and that is published under my own name should never be written with the assistance of AI. Except very rarely and diagnostically.

[After posting, another friend writes to ask me what this means. I respond: “Claude, as an example, is a really great spell-checker and proofreader and it flags half-finished thoughts as a part of the effort. It catches compositional errors that Word doesn’t. And it does a great job of assessing the alignment of evidence with argument. I want to limit my use of it when it comes to my own work to the final, finishing stages of preparation for submission or review].

From the heart and mind

I am also an administrator with a title and a public-facing role. A friend once told me that taking an administrative role meant trading voice for agency, and that is generally true – although policy and process are also reflections of who I am and what I bring to this job, so “voice” is inevitably a part of the job. But pleasure, in this role, derives from accomplishment, from collective advancement and shared problem-solving. My voice, in this context, is one tool in the toolbox for a team, and it is shaped by considerations that go beyond my scholarly accomplishments. And pleasure, such as it is, attaches to outcomes that benefit everyone.

In this administrative role, I write many things, not all of them scholarship – even if my scholarship lies behind them. When and where I use AI for assistance, it is almost exclusively in this domain. How I use it matters.

One way to think about the subject of AI use is to imagine that anything that expresses who and what I am – anything that is me, speaking and feeling and thinking – needs to come from me. This limits AI use to assistance and not creation, though I am here considering those things specifically through a humanist lens. “Assistance” and “creation,” I recognize, might mean something different for social scientists and scientists.

As a humanist, if I write a letter or an email or an article or a book, it should come from me. Even a simple email that reads: “OK.” 

For my administrative work, “come from me” should mean that the first draft and the final draft should come from my finger tips – that is, I should be able to explain my thought process clearly for each choice from start to finish. Because I am a writer, and because I type in a hunt-and-peck style, things that come “from my fingertips” come slowly, intentionally, and with constant focus. To me, “from my finger tips” implies not merely a mechanical act, but a cognitive one, as I repeatedly mistype things, routinely revise what I see appear on the page, and habitually read things aloud to make sure that it sounds “right.” Things I write are “mine” because of this process. And what is “mine” has stylistic quirks and miscues – and a limited vocabulary, fixed and derived from the bounds of my life – that belong to me.

The notion of a “first draft” needs a little elaboration, too. When I was younger, I learned to outline with a pencil and paper, often with a thesis paragraph written in longhand. Later, I would type out these same things carefully, because even small changes required “white out” or correction paper. When word processing arrived, my initial efforts to define an idea got a lot messier – in a good way. Sentences could be scrambled easily, and outlines became dynamic. I share all of this to be clear that “first drafts” have changed three times in my lifetime, but the basic idea remains unchanged: a general sense of themes and ambitions and argument, a clear sense of a frame, of what fell into the project – and what fell out. The frame is the first step in the project, the first draft; the sentences “from my fingertips” are what make it mine.

If I use AI for this body of administrative work, at each step of revision and refinement between those first and final drafts I am personally assessing any suggestions, ranging from simple word replacement to the simple reorganization of sentences and paragraphs and the wholesale revamping of sections. At this point of intermediate assessment – in the interstices between the beginning and the end – I am being intentional, thinking about any turn, slowing things down to reflect. And at the close, I am reading each and every word, often aloud, to ensure that the submission of any final draft can be connected in a narrative fashion to the roughest first draft, that it reflects my ways of thinking and seeing and knowing something, and that each step in that narrative has me as the subject in a sentence. 

Because I believe all of this, I am responsible for every word, every sentiment, every idea. And I should never, ever deflect personal responsibility on an agent or bot. If I use an AI in any way, I should disclose it early and often, explaining its use, and perhaps even by sharing this piece of writing.

Another set of eyes

To use AI as another set of eyes in those intermediate steps is to enlist it as a reader, as an iterative partner. It must not be the only reader. And other readers should be human.

In this iterative partnership, AI should be used diagnostically and prescriptively. Diagnostically, it can be used to assess structure and organization, to diagnose weakness and the misalignment of evidence and argument. It can be used to pinpoint a lack of clarity, or to highlight a missed step. It can be asked to assess a document or an image or a policy position. Prescriptively, it can offer a remedy, suggest a fix, offer a place where a particular piece of evidence might be moved or spotlighted. But, as would be true after a visit to the doctor, the actual remedy is on me. Fixes and movements and reorganizations need to come from me. I must be the actor in the sentence. I must understand the “why?” of every tweak. Offers of re-writing must be refused. The changes need to come from my finger tips – again, from me, with my belabored typing, my self-doubt, my limited vocabulary, my quirky style and love of small details made big.

Unlike a visit to the doctor, though, my use of AI as an iterative reader also means that I need to get a second and third opinion (or more) from humans, that I need – once more – to have other, human-created diagnoses and prescriptions in mind while I am revising. The danger in soliciting feedback from an AI agent is that the gap between a first draft and a final draft might be shaped only by an instrument of machine learning. A broader, diverse set of readers at each step – good practice anyhow – is a solution. 

So is time. Another key element here is the need to ensure that there is time to assess, to reflect, to worry over word choice and style and tone. The problem with any reliance on AI is that it is faster than a human reader, that it never tires or needs sleep or eats, and that the pace of iteration, of diagnosis, and of prescription might move more swiftly when there is an agent involved. There needs to be a limit to its use. And there needs to be a cadence that is human.

“This is an impossible ethical standard,” a sympathetic friend cautions, and “it will collapse under the weight of efficiency.” She is right. I am tilting at windmills. As I lay down these words on the page, I am doubtful that any of this – the emphasis on human readers, the desire for a human pace of writing and revision – will survive my generation. These principles, I know, are the terms of my surrender to a future that is already here. I will be knocked off my horse soon enough.

Another set of hands

There are only a few items that might be directly produced by AI, a few examples of when it might offer itself up as an extra set of hands. In all of these cases, it needs to be assistive, not acting independently. All of these examples are about what we might call “work product,” jointly-created material that is not published under my name and that does not purport to express my views, opinions, ideas, and scholarly expertise. Here are some examples, all of which depend on the original data or text being created ethically, as described above:

1. Slide decks for presentations or lectures from text or data 
2. The creation of internal proposals or timelines or project management dashboards from text or data
3. Anonymous meeting minutes or notes compiled from an audio recording or from my own reportage
4. The conversion of data or text into a different format, like a table into prose summary for internal use or a spreadsheet into a chart
5. Shifting a document I’ve written into a required template, or into a house style, or shifting footnotes into some specific citation format
6. The transcription of interviews, lectures, or dictated notes, and the correction of the same for clarity, not content

Now, some might argue that it is entirely appropriate to upload to a closed system – like Notebook LM or something like that – all of the raw materials of a piece of writing and thinking and ask an agent to organize it, to assess it, or even to make sense of it. I’m not including that sort of use here, because this is where my scholarly writing and my administrative work part ways most sharply. Again, speaking personally, I’m uncomfortable using that kind of agentic approach to make sense of something that should come “from my fingertips.” I might do that if I am sifting through vast amounts of data for a formal report, or if I’m doing an impact assessment of some policy or program, but I won’t do it if I’m writing a history, or a biography, or a short essay to be published under my name as a writer.

I had originally included a seventh possible use – “The composition of literature summaries or briefing documents that compile what sources say – as long as the compiled summary isn’t submitted as my own analysis.” But a friend worried that this would have an impact on learning and cognition, and that it might be understood as permission given to students and colleagues to skip over the hard act of turning the physical page, struggling to discern meaning. Or reading, that is.  That hadn’t been my intent, and I didn’t want to be misunderstood.

And, of course, before any internal or external use, this material needs to be fact-checked by humans. For this reason, translation and code fall outside of this list. Translation risks becoming a voice – replacing some other “me,” some other human’s creative reinterpretation of the original text – and code is not something I create or use. In both cases, I would not be able to fact-check anything.

How I wrote this

I wrote the first draft myself, and tweaked it a lot. I sent it to a few friends and my wife, and expanded the opening section. 

I shared it with Claude and asked for an assessment. Claude offered a critique of the phrase “from my finger tips” and wondered whether I need to offer more examples of areas where AI might be “another set of hands.” I expanded it some more in those areas, and addressed a few other areas where I had left a few things half-baked. I asked another friend to read it, and changed it again. I shared it with Claude again. I talked to my wife again. Reading through it alone, I decided I needed to say more about “the pleasure principle.” Claude wondered where administrators found pleasure and proposed an answer, but I went in a different direction. I added the piece about my friend, and about what it means to trade voice for agency. Claude was surprised by that. I sent it to a few colleagues and friends. Whole paragraphs were added to address their particular, often pointillist concerns. Sometimes, they shared their own sense of struggle, or amusement, or irony. We disagreed on a few things, too. Finally, I asked Claude to review how I represented its voice in this final paragraph, and whether I had captured the spirit of our exchanges accurately. It disputed the attachment of “surprise” to one conversation, but I decided that was how I, as a writer, should best capture it.

I’m certain this will change.

September 27, 2026