Do people care who (or what) wrote this?

Generative A.I. as a writing tool has limitations. But what I've discovered over the past week is that my perceptions of those limitations can drastically change when I learn about a new way to use it. Before, I'd been giving ChatGPT fairly vague prompts: "Describe the town of Cottondale Alabama as a news article." Listening to a copy.ai prompt engineer on Freakonomics helped me understand that being more specific in your prompts about the length of the output ("500-1000 words") and the audience ("highly-educated audience") makes all the difference. 

The other key lesson is to think of writing with A.I. as an iterative collaboration: ask the program to generate five options, use your gold ol' fashioned human judgment to select the best one, then ask it to further refine or develop that option. If you find it to be boring, ask it to vary the sentence structure or generate five new metaphors for something and then pick the best one. I sensed that writing with generative A.I. could be more like a collaboration with a co-author than an enhanced version of auto-correct; this helped me to see what, exactly, that collaboration looks like, and how to effectively collaborate with the program. 

As the output got better and better, I wondered, "has anyone done a blind test of readers' ability to discern A.I.-assisted writing from purely human writing?" I'd heard of a few misleading journalistic stunts where writers trick readers into thinking that they're reading human writing when, in fact, they are not. But I'm looking for something more rigorous, something that compares readers' abilities to discern that difference across genres of writing: short news articles, poetry, short stories, long-form journalism, short documentary scripts, etc. It seems likely that readers will prefer the A.I.-assisted version in some cases, but it's important to know what types of cases those will be. 

I also wondered what our reactions - as readers and writers - to all of this. I can think of three metaphors for possible reactions to A.I.-assisted writing:

1) the word processor. It's use changed how writers write. It changed the output. Like most disruptive technologies, it was met with skepticism and hostility. But eventually, it was widely adopted. Young writers who hadn't grown up writing free-hand had an easier time adapting to this new way of writing. The technology became "domesticated" - normal to the point of being invisible, embedded in pre-existing structures of economy and society. 

2) machine generated art. Machines have been generating visual art for decades. Some of that art is indiscernible from human generated visual art. Some of it embodies the kinds of aesthetic characteristics that people value. And yet machine generated art has never risen beyond a small niche. The market for visual art largely rejects it, in part because those who enjoy art care about how it is created. Something about the person who created it and the process by which it was created is part of what they value about art. 

3) performance enhancing drugs. Output from A.I.-assisted writing is superior - in some cases far superior - to unaided human writing, and there is market demand for it - the public sets aside its qualms and embraces good writing regardless of how it came about. This situation is perceived by writers, some industries, and some governments as unfair or possibly dangerous, maybe in terms of what bad actors could do with such a tool or how profoundly disruptive its widespread use would be for economies and society. Therefore, they regulate it, discourage its use through public shaming, or, in some cases, explicitly forbid its use. 

The quality of A.I.-assisted writing's output is only part of what will determine its eventual place in our lives. The general public's reaction to it is another part worth paying attention to. 

An ethical case for using A.I. in creative domains

A few months after first considering the promise and threat of A.I. in creative domains, it's still the threats that are getting the most attention. I tend to hear less about the possibility that by allowing A.I. to be used widely (which helps it grow more sophisticated) we are hastening a machine-led apocalypse and more about what we would lose if we replaced human writers with A.I. It would be an obvious loss for the people who write for a living, but they make the case that it would be a loss for society. Creativity would decline, mediocrity would flourish, and we would lose the ineffable sense of humanity that makes great art. By taking the power to create out of the hands of the many writers and putting it in the hands of the few big tech companies, we would exacerbate inequality and consolidate control over culture. 

There are a few steps in this hypothetical process worth scrutinizing. First, this argument assumed that if A.I. is allowed to be used in a creative field (screenwriting, journalism, education), it will necessarily lead to the replacement of human labor. There's a market logic to this: if you owned a company and you could automate a process at a fraction of the cost of paying someone to do it, you would have to automate it. If you didn't, your competition would automate it, be able to produce an equivalent good or experience at a lower cost, charge consumers less for it, be a better value to shareholders as a publicly traded company, and put you out of business. You could point to examples of such things happening in the past as evidence of this logic (though I have to admit, I found it hard to find examples that used human communication rather than physical labor. I'd assumed chatbots had led to steep declines in customer service labor, but all I could find was editorials about how it will lead to steep declines and competing editorials about how customers find chatbots enraging and still demand human customer service agents). 

But I still have trouble thinking of this particular replacement-of-human-labor trajectory as inevitable. I can't help but think of A.I. as a tool that humans use rather than a replacement for humans, more like a word processor or the internet than a brain. I can't not see a future (and, honestly, a present) in which writers of all kinds use A.I. for parts of the writing process: formatting, idea generation, wordsmithing. Humans prompt the A.I., evaluate its output, edit it, combine it with something they generated, and share an attribution with the A.I. You could call this collaboration or you could call this supervision, depending on how optimistic or pessimistic you are, but the work that it generates is likely better than what A.I. generates on its own and it is generated faster than what humans generate on their own. But humans who prompt, edit, evaluate, and contribute to creating quality work are as necessary as they were before. They can still use that necessity to make their case when bargaining with corporate ownership. 

I also have trouble seeing a marketplace in which all content is generated by A.I. If the A.I. can only generate mediocre content, won't people recognize its mediocrity and prefer human-made creative work? It's hard not to see this particular facet of the argument against A.I. in creative fields as elitist snobbery - "of course the masses will choose the A.I.-generated dreck they're served! We highly-educated members of the creative class must save them by directing them toward 'True Art,' (which we just happened to create and have a financial stake in preserving)."

And that is an ethical argument for A.I. in creative fields that I have yet to hear: the argument that the class of people who are currently paid for being creative are protectionist. If they can just keep us thinking about Big Tech and the obscenely wealthy studio execs, we won't have to think about the vast number of smart, creative, compassionate people who happen to not know how to write well, or to write well in a particular language. I worked hard at becoming a good writer, spending a lot of time and money to acquire this marketable skill. Does that make it morally right to deprive others of the ability to use a writing tool that levels the creative playing field? I assume there are millions of people with the life experience and creativity to be great writers who simply lack the educational experience to craft grammatically correct prose. Who am I to insist they take out loans and wait years before they can make worthy artistic contributions?

I do understand the replacement-of-human-labor argument against A.I. None of the anti-protectionism argument really resolves or even speaks to the market logic argument. I suppose this is what smart regulation does - limit the use of technology in cases where we see clear evidence of social harm but allow it where there are opportunities for social good. As an educator, I want to make sure that students understand how to recognize the characteristics of "good" (i.e., compelling, effective at communicating, lasting the test of time) writing, even if they need a little help getting their subjects and verbs to agree.

It can be hard to see the good of A.I. in creative realms at this stage in the development cycle. It is hard to see the would-be writers and the untold stories, but any ethical approach to the question of A.I. in creative fields must consider them. 

The promise and threat of A.I. in creative domains

Though we've been expecting AI to get good enough at writing to actually perform a decent number and variety of writing tasks "by itself" (it still needs humans to engineer prompts), we hadn't been confronted with that reality until recently, when ChatGPT became more widely used. To test it out, I gave it some of the (relatively simple) prompts I give my undergraduate film history students and saw what it came up with. Then, I asked it to write a pilot episode of a teen drama set in Montana, a grant proposal for a research study on social media use and political polarization, a research paper on the environmental impact of microplastics, a quiz on 1970's media cultural studies, a review of Beck's new album, an amicus brief on microplastics, and a summary of the 1998 NBA Finals in the style of Proust. The results, as many have noted, aren't perfect but are a lot closer to perfect than what we're used to seeing from programs. It's hard to objectively convey the quality of the output and the limitations of the tool; it's best to just try it out yourself. 

I'm sure such tools will improve, but it's hard to say at what rate and to what degree, so I'll try to just stick to evaluating the impact of the current iteration of generative text programs like ChatGPT on two facets of writing: assessment (specifically, educators' ability to assess students' understanding of concepts, thinking, and communication skills) and its applications in various domains (i.e., journalism, creative writing, scientific writing). 

In terms of assessment, the ability of students to avoid actually doing the work of writing for an assignment and to avoid detection has increased significantly. However, the truth of this statement is highly dependent on the assignment. If the assignment just requires students to gather information (particularly older info on an extensively covered topic, person, or event) and synthesize it into a simple, readable essay or article, then ChatGPT seems to be quite good at that. But what about, say, watching a particular video on YouTube and applying something from a particular reading to that video? I would assume that ChatGPT can't easily extract the information that's in the video, so it would have a tough time with this task, and that it's ability to do anything with a particular reading is contingent upon that reading being freely available and accessible online somewhere. If the reading is in a textbook that doesn't appear online (or only appears on a few hard-to-access pirating sites), then this will be harder for ChatGPT to work with than, say, the Magna Carta or any other text that's in the public domain. But in any of these cases, I think that A.I. co-authoring (where ChatGPT does some of the work and the student changes and adds a few things) will be essentially impossible to detect. 

Another approach to ChatGPT's threat to assessment is to simply proctor writing assignments. That's easier to do for short writing assignments but difficult for longer ones, and impractical for online classes. Still, as a professor who has an interest in fairly and accurately assessing his student's writing, I'm leaning toward more in-class writing assignments and fewer outside-of-class writing assignments because of how challenging it will be to tell the difference between A.I.-generated writing and my students' writing outside the classroom. 

But the topic of applications of AI is more interesting to me, and in some sense pre-empts questions about assessment. If AI can accomplish much of what we're trying to train our students to accomplish, what is the point of such training? The rapid adoption of ChatGPT has forced me to reconsider the relationship between writing and thinking. Writing is often the medium through which thought is expressed, though not the only one; spontaneous discussions and conversation are a good alternative. When is writing chiefly a vehicle for thought and when is it more an integral part of the process of thought? Those who defend the sanctity of human writing, I surmise, would say either that writing is very often part of thinking, or that we may not be very good at being able to tell the difference between writing-as-vehicle-for-thought and writing-as-thinking. Therefore, ceding the task of any but the most formulaic types of writing to AI risks sacrificing the ability to think. 

We still want to get better at thinking, and we still want our students to get better at thinking (e.g., critical thinking, creative thinking, logical reasoning, problem solving). Even in a world of prevalent AI, having those abilities will make us better at prompt engineering, better at judging the value of what ChatGPT and similar AI spit out, and better at making refinements to that output. The same would apply for non-linguistic AI: prompting an AI architectural design program and judging/refining the results; prompting an AI coding program and judging/refining the results. In any application I can think of, humans are still part of the process. Whether AI is producing a kind of structural scaffold for ideas, identifying sources of information, or generating new ideas, it has to be set in motion, its output needs to be evaluated, and in its current iteration, its output requires a fair amount of refinement. That process of prompting and refining can be similar to a good conversation or collaborative improvisation ("yes, and..." thinking). 

I sense that there's a widespread antipathy toward applying AI in creative domains. I don't consider myself an AI optimist. I acknowledge how tempting it will be for students, businesses, creators, and governments to use AI in an indiscriminate manner, not bothering to vet or refine its output. I acknowledge that such use could cultivate a kind of passivity and (further) overreliance on technology. More people would get out of the habit of "thinking for themselves." Creativity would be increasingly derivative in the name of efficiency. 

But does this warrant an outright ban on AI? What would such an approach sacrifice?

I feel as though we've been good at articulating the risks of AI (aside from the aforementioned, there's scaling up spam and disinformation, and deceiving job recruiters or potential partners on dating sites). We're less good at articulating its potential upside. I think of that upside chiefly in terms of resource reallocation (which, yes, sounds very cold and business-speaky, but hear me out). 

Let's treat ChatGPT and the like as co-authors. They generate first drafts or templates but don't write the entire text or finished product. Human writers' role in the creative process shifts from crafting sentences and paragraphs to gathering primary original sources and identifying needs in their audiences and communities. What if writers could take the time they would have spent struggling to write formulaic copy and instead spent it cultivating twice as many primary sources as they would have otherwise, or co-writing twice as many stories, or ten times as many stories? Consider all the un-written stories or accounts of events; consider what's missing from historical accounts and news coverage, in part because the people who know how to create thorough and entertaining writing only have so much time. Some people's stories are told, some people and events are remembered, but the vast majority are forgotten. What if we could fill in some of those blanks? We don't have to be indiscriminate in our use of AI to turn data into stories that no one cares to read. Remember, humans and their understanding of what is valued are still part of the process. And what if we could offer those stories in twice as many accessible formats? If editors and readers valued these qualities more than they feared AI's downsides, we could use AI to improve our information environment. 

The genie is out of the bottle. If one country or group of people wants to spend time trying to stuff the genie back into the bottle, that won't stop other countries or groups of people from using it. We needn't be totalizing in how we conceive of the application of AI in creative domains. We have been discriminate in our use of technologies in the past, often making mistakes, attempting to correct those mistakes by introducing limitations or redesigning the technology, fitfully moving along but rarely un-inventing anything. Maybe we should look around and see where the successes and failures are right now.