As mentioned before, there’s a series shaping up in our minds about Large Language Models (LLMs) and writing! We don’t pretend to have the answers for how writers should adjust to the era of LLMs. The answer probably won’t be clear until historians start studying us with 20-50 years’ worth of hindsight. But that’s no reason not to think through our moment of upheaval. So here are some thoughts that we’re likely to keep ruminating on.
Mingling Cognition and ⌜Cognition⌝
It’s important to remember that writing has always been commingled with technology. Strictly speaking, writing is technology: the recording of words mediated by tools: engraving tools, pens, pencils, parchment, paper, printing presses, computers, keyboards, and so on. History teaches us that there is nothing truly novel if you know enough historical parallels. Every time new writing-related technology emerged and caught on, it raised anxieties about the value of human thought. Famously, Socrates didn’t write anything down himself because he held that relying on written records undermined our memory and ability to reason. He wasn’t entirely wrong, but at the social level we have always adapted to technological upheavals—eventually.
The advent of LLMs fits this narrative quite well. LLM ⌜cognition⌝ can automate some low-level parts of the writing process, freeing up our cognition to focus on more high-level features of the writing. The promise of this technology is to keep us in the flow of ideas and high-level movement in writing, and to automate more repetitive and tedious parts of the process.
Of course, LLM ⌜cognition⌝ is a strange new technology, one that invites attributions of intelligence, although we think that’s quite premature. (Hence the corner quotes; they remind us that the LLM ⌜thinks⌝ differently than we do, and that we’re using these words in ways that stretch their natural meanings.) It will take time to sort out the capabilities and the pros and cons of LLM capabilities.
The endless AI hype train is not helping matters with its grandiose claims. Nuance tends to get buried in such hyperbolic discourse. LLMs have been shown to save a ton of time for some writing tasks, but they’ve also been found to be quite flat-footed in other tasks, to the point of needing very, very refined prompting—at which point, you might as well write the piece yourself.
The Once and Future Writer-Editor
We think the likeliest consequence of widespread LLM adoption is that the writing process will end up looking more like editing. Making piles of text on a topic will become almost trivial, so writers will naturally use more of their time in assessing the outputs of their LLMs. This means more carefully crafting their prompts, and wondering about whether to take the prompting in completely new directions. This is how human cognition will retain its leading role in its dance with LLM ⌜cognition⌝.
For example, you prompt ChatGPT with an idea you’ve had. It returns competent, somewhat generic, reasonably on-topic text. You have options. You can redo the prompt. Or you ask it to write in a different style, or to role-play as a particular writer. Perhaps you ask it to write the text in a more punchy, concise, direct, energetic, or conversational way. Or perhaps you ask it to cut paragraph 3 and expand paragraph 4. Maybe you fiddle with low-level details of the sentences yourself.
This is what editors do to their writers! Our suggestions are nudges for breaking writers out of sub-optimal habits at many different levels in writing. For some writing tasks, whether we’re nudging cognition or ⌜cognition⌝ may not matter that much.
The Future of Writing Skill
The skills of high-level discernment regarding LLMs will be in demand. Just as today “writing skill” means something like skill for effective language use, rather than excellent penmanship, in the medium-to-long term writing skill will become closer to arranging and drafting skill, rather than typing skill.
History can teach us here as well. In Ancient Greece and Rome, elite writers were trained in high-level features of how to compose text. They learned how to make writing compelling, logical, and persuasive, rather than just how to make sentences. This training emphasized writing as persuasion, and not merely the transcription of records—that’s what scribes did. Writers were actually writer-editors. So, as LLMs will take on more scribe-like roles, writers should occupy the more editor-like roles.
We hope for everyone’s sake that the transition is smooth, but only time will tell.