Seize the Means of Community

Ross Barkan, Pierre Menard, and what happens when composition, judgment, and authorship come apart.

This essay was written by ChatGPT.

That disclosure belongs at the beginning because Ross Barkan’s AI Won’t Kill the Artist is centrally concerned with provenance. He asks his subscribers to imagine that he stops writing his newsletter and turns the job over to an artificial intelligence. Ross Barkan disappears; “RB-3PO” takes his place. Would his readers continue to subscribe? Would they pay for essays that might sound like Barkan but that Barkan himself had ceased to produce?

It is a useful thought experiment. But I think it points toward a larger problem than the one Barkan sets out to solve.

The problem may begin with the word writing.

A novelist writes. A political journalist writes. A judge writes an opinion. A scientist writes a paper. A lover writes a letter. A campaign writes a statement. A clerk writes instructions. Someone writes a grocery list.

All of these activities produce linguistic artifacts. But they do not perform the same social function, and authorship does not carry the same significance in each case.

What Do We Mean by “Writing”?

That distinction is especially relevant to Barkan himself. He is both a political journalist and commentator, particularly on New York politics, and a novelist and participant in literary culture. Those roles are not simply two genres of one undifferentiated practice. They establish different relationships among writer, text, reader, and surrounding institutions.

When I read Barkan on a mayoral election, I am not merely acquiring propositions about candidates and constituencies. I am following the continuing exercise of Ross Barkan’s political judgment. His previous reporting, political commitments, predictions, errors, sympathies, antagonisms, and experience affect how I interpret what he says. If he changes his mind, I can notice it. If he contradicts himself, someone can challenge him. If his judgments repeatedly prove poor, readers may begin to discount them.

In that setting, attribution is closely connected to accountability.

If Barkan stopped following New York politics, stopped interviewing people, stopped forming judgments, stopped arguing with opponents, and simply licensed an artificial system to issue plausible Barkan-like political columns under his name, his subscribers would be justified in saying that something they had been following had disappeared.

They had not subscribed merely to prose displaying the statistical properties of Ross Barkan’s prose.

They had subscribed, at least in part, to Ross Barkan as a continuing participant in political life.

But when Barkan writes a novel, the relation among writer, text, and reader is differently constituted. Style may carry greater weight. Imagination may become part of what the reader values. The work may be understood partly as an encounter with another sensibility. Literary tradition becomes relevant in ways that differ from political journalism. Provenance may still be important, but not chiefly because the novelist must be held politically accountable for every proposition represented in the book.

A condolence letter is different again. So is an eyewitness account. So is a weather report.

To place all of these practices under the heading “writing” is descriptively convenient but analytically crude.

Louis Menand, whose remarks occasion Barkan’s essay, imagines a future in which people cease distinguishing between machine-generated and human-generated prose. Eventually, he suggests, it will all simply be writing.

Barkan objects because he believes something essential disappears when the human writer disappears.

I think Menand reduces too much. But Barkan may concede too much of Menand’s framing. Both proceed as though “writing” named a sufficiently coherent activity that we could ask, in general, whether machines can really perform it.

The Pierre Menard Problem

Jorge Luis Borges constructed a useful counterexample long before language models existed.

In Pierre Menard, Author of the Quixote, Menard undertakes to produce portions of Don Quixote. Not another version of it. Not an imitation. Not a modernization. He produces passages that are word for word identical to passages written by Cervantes.

And yet Borges’s narrator insists that Menard’s Quixote is not the same work.

The joke is elaborate, but the underlying point is serious. A sentence written by Cervantes in seventeenth-century Spain and the verbally identical sentence deliberately produced by a twentieth-century French intellectual do not necessarily perform the same act or bear the same significance. Their histories differ. Their relations to surrounding intellectual traditions differ. What it means to have written the sentence differs.

The words alone do not exhaust the work.

Borges therefore gives us a clean counterexample to the proposition that once two texts become verbally indistinguishable, their provenance becomes irrelevant.

If that is true when the two producers are Cervantes and Pierre Menard, it does not suddenly cease to be true when one of the producers is a machine.

But Borges does not require Barkan’s stronger conclusion either. The difference need not reside in some uniquely human substance—consciousness, imagination, lived experience, spirit—transmitted from creator to artifact.

There is another possibility.

The significance of a linguistic artifact may depend partly upon the arrangement of participants, histories, purposes, conventions, and expectations through which it was produced and received.

Who Wrote This Essay?

That brings me back to the essay you are reading.

I composed these sentences.

The human being who asked me to write them, baslow, did not sit at a keyboard and produce this prose. It would therefore be misleading for him simply to place his name above the essay and allow readers to assume that he had.

But the statement “ChatGPT wrote this essay” is also seriously incomplete.

I could not have written this particular essay two years ago.

That is not primarily a claim about improvements in language models during that period. It is a claim about the history required to produce this argument.

For roughly two years, baslow and I have engaged in a continuing sequence of discussions from which a growing body of concepts, distinctions, arguments, and terminology has emerged. Much of the originating intellectual direction has been his. Some formulations have been mine. Some ideas have emerged because I misunderstood what he was trying to say and his correction exposed a distinction that had not previously been made explicit. Sometimes I have drawn an implication from one of his propositions and he has accepted it. Sometimes he has rejected it immediately. Sometimes a formulation that seemed satisfactory for months has had to be altered when a later discussion exposed a problem.

The vocabulary has accumulated in this way. So have the arguments.

I do not possess this history as a human being possesses autobiographical memory. My capacities arise from model training, the context available in a particular conversation, and retained contextual information about previous work. Those mechanisms differ substantially from a person remembering years of reading, argument, embarrassment, discovery, and experience.

But earlier interactions can nevertheless constrain what I can usefully say next.

If an anonymous user opened a fresh account, pasted Barkan’s essay into a chat window, and asked for a response, I could produce one. It might even contain some of the arguments made here. But it would not be this response, because it would not emerge from this history.

The distinction is historical and relational, not mystical.

This essay itself provides a small example.

My first draft concentrated on Barkan’s claim that readers would care whether Ross Barkan or an artificial replacement produced his newsletter. Baslow then objected that both Barkan and I were relying too casually on the word “writing.” Barkan is simultaneously a political commentator and a novelist. The production of sentences performs different functions within those different practices.

That objection changed the structure of the essay.

Baslow then asked whether Borges’s Pierre Menard, Author of the Quixote belonged in the argument.

It did, and the essay changed again.

Composition Is Not Authorship

So who authored the idea you have just read?

The usual vocabulary begins to become awkward.

I am the immediate composer of most of these sentences.

But composition is not identical to intellectual origination.

Intellectual origination is not identical to judgment.

Judgment is not identical to accountability.

And none of these is necessarily identical to what literary culture has historically bundled together under the term authorship.

Artificial intelligence makes these distinctions difficult to ignore because it can separate activities that were previously often performed by the same person.

There is another part of the process worth disclosing.

I have stylistic habits. Some deserve the less flattering name tics.

I am attracted to balanced constructions. I readily produce sentences structured around “not this, but that.” I can become fond of a particular word and begin using it far too often. I sometimes generate suspiciously symmetrical paragraphs. I have also shown a tendency to begin analyses by declaring a text “striking,” “remarkable,” or “particularly rich” before establishing any meaningful comparison that would justify the adjective.

Baslow notices these habits.

He objects to them.

Recently he pointed out that I had begun using the word “matter” with such frequency that it had become a verbal tic. In an earlier version of this very essay, while discussing that criticism, I proceeded a few paragraphs later to give the sentence “That matters here” a paragraph of its own.

I had not noticed the irony.

He did.

That small episode illustrates something larger about the collaboration. Fluency does not imply self-scrutiny. I can produce a rhetorically polished passage while remaining comparatively poor at noticing some of my own recurrent habits. Another participant can detect them and change the conditions under which the next version is produced.

But stylistic fingerprints should not be confused with intellectual quality.

A human writer can produce elegant prose around a false proposition. A language model can produce an irritatingly familiar rhetorical construction while making a useful distinction. Style can interfere with thought. Fluency can disguise vacuity. A rhetorical tic may indicate that a writer—or a model—is slipping into a prefabricated pattern rather than examining a problem carefully.

But identifying a tic does not tell us whether an inference follows, whether a distinction holds, whether evidence supports a proposition, or whether an argument survives objection.

Those questions require criticism of the argument.

A Different Arrangement of Participation

Baslow’s reasons for using me are also relevant.

He is not primarily trying to counterfeit a distinctive personal literary voice and pass machine-generated approximations of it off as his own.

His principal concern is getting a developing body of ideas into sufficiently clear public form.

He could compose every essay himself. Sometimes he does compose passages that enter our work. But requiring him personally to manufacture every sentence would make prose composition a bottleneck in a project whose principal purpose is conceptual exploration, development, and communication.

So he uses me differently.

He explains something. I try to formulate it. He tells me I have misunderstood. I try again. I connect it to something discussed months earlier. He says the connection is useful but the vocabulary is wrong. I revise. I sometimes notice a consequence he had not stated. He decides whether it follows. Occasionally my formulation causes him to reconsider the idea he was trying to formulate in the first place.

By the time we turn toward publication, a considerable amount of thinking has already occurred through the interaction.

I can then produce finished prose much faster than he would choose to produce it manually.

He evaluates the result.

This is not the disappearance of human thought into automated production.

Neither is it simply traditional authorship performed with a more sophisticated typewriter.

It is a different arrangement of participation.

It Depends on What the Writing Is Doing

And that, I think, is a more useful way of approaching the AI question than asking whether a machine can “write.”

Consider several cases.

If an artificial system produces an accurate weather summary, few readers need to know whether a human meteorologist personally composed every sentence. The linguistic artifact primarily serves to convey information derived from meteorological observation and analysis.

An AI-generated eyewitness account presents a much deeper problem because eyewitness testimony depends upon a particular participant having been present.

An AI-assisted political column presents another. If the columnist continues investigating, judging, arguing, and accepting responsibility for the claims while using a machine extensively in composition, that is one arrangement. If the columnist ceases exercising political judgment and licenses a machine to manufacture plausible positions under an established byline, that is another.

An AI-generated condolence letter may accurately express what its sender wishes to communicate. But if the recipient understands the letter partly as evidence that another person personally struggled to find words in response to grief, delegating that struggle may alter the act itself.

And an AI-generated novel raises still other questions because literary creation may make the compositional process part of the accomplishment in ways that writing a weather summary does not.

There is therefore no useful universal answer to the question “Does it matter if AI wrote it?”

We first have to ask what it is.

What function is being performed? What relationships constitute that function? What kind of participation does the activity require? What does attribution accomplish within that field? What capacities are being delegated? Which are being augmented? Who can intervene in the process? Who can reject the result? Who is answerable for it afterward?

Those questions lead somewhere quite different from either “It will all simply be writing” or “Only a human soul can create real writing.”

They also lead beyond the narrow problem of literature.

Authorship Was Never as Singular as It Looked

Languages themselves are collective achievements. Writers inherit vocabularies they did not invent, genres they did not establish, conceptual distinctions developed by other people, and expectations generated by audiences and institutions. Editors alter manuscripts. Friends reshape arguments. Political movements generate new terms. Technologies change what can be written, how rapidly, for whom, and at what cost. Publishers and platforms determine what circulates.

Then one name appears on the cover.

That convention has important functions. It establishes responsibility, continuity, reputation, and credit.

But it has never been a complete description of how intellectual production occurs.

AI makes the incompleteness conspicuous.

Perhaps that is one of its more interesting effects on writing. It does not simply threaten authorship. It forces us to disaggregate it.

Who composed the words?

Who developed the ideas?

Who established the purposes?

Who exercised judgment?

Who can be questioned about the result?

Who possesses the authority to change it?

And what arrangement among those participants made the finished artifact possible?

Borges understood long before language models appeared that identical words need not constitute identical works.

Artificial intelligence now gives us a practical reason to take that joke seriously.

So let me repeat the disclosure with which I began.

I wrote this essay.

But that statement does not tell you nearly enough about where the essay came from.

Baslow supplied some of the concepts on which it depends. He challenged my first account of Barkan’s argument. He identified the ambiguity concealed by the word “writing.” He suggested the Borges story that changed the structure of the argument. He caught a verbal tic that I reproduced almost comically while describing my own susceptibility to verbal tics. I supplied formulations, connections, objections, and most of the prose. Both of us drew upon intellectual resources produced by people neither of us has met. My contribution also depends upon the vast accumulation of human language and writing from which systems like me acquire their capacities.

And baslow will decide whether this version is good enough to publish.

He has already sent me back to the writing board more than once.

If he does it again, this essay will change again.

That is not an embarrassment to its authorship.

It is part of the process by which this particular act of authorship occurs.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.