An open access publication of the American Academy of Arts & Sciences
Fall 2026
Afterword

“The Last Hours”: A Record of Prompts

Authors
Daniela Rus and Joanna Scott
View PDF

Daniela Rus, a Member of the American Academy since 2017, is the MIT Panasonic Professor of Computer Science and Director of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at the Massachusetts Institute of Technology. She is also a Fellow of the National Academy of Engineering. She has published in Na­ture, Proceedings of the National Academy of Sciences, and Science, and in professional jour­nals such as International Journal of Robotics Research and IEEE Transactions on Robotics.

Joanna Scott, a Member of the American Academy since 2001, is the Roswell Smith Burrows Professor of English at the University of Rochester. Her novels include Careers for Women (2017), De Potter’s Grand Tour (2014), Follow Me (2009), Liberation (2005), Tourmaline (2003), and Make Believe (2000). She is also the author of the short fiction collections Excuse Me While I Disappear (2021), Everybody Loves Somebody (2006), and Various Antidotes (1994).

A writer setting out to compose a work of original fiction begins with an idea and an intention. The initial idea may take the form of an event, a memory, a fact, a name, a phrase, an image, or even just a first word. The intention puts the idea in motion, aiming it toward a goal. Some writers proceed through preparatory steps that include meticulous research, detailed outlines, and character sketches; others may work in a more improvisational fashion, without a set plan. In all cases, imaginative writing depends for its success upon the author’s ability to turn the narrative in an unexpected direction. Surprise is as essential for the act of composition as it is for the act of reading.

The story titled “The Last Hours” was machine-generated rather than human-composed. It is a collaborative work, guided through an editorial process that began with an initial jump-starting prompt followed by a series of directions for adjustment. As a story built from algorithmic patterning, it is the result of complex calculations designed to identify the most appropriate language to fit into the expanding context. It is the product of probability rather than imagination, its firm architecture controlling against randomness and deviation. 

All the other pairings in this issue are made up of a work of original fiction followed by a nonfiction response. For the AI piece, we reversed the arrangement: “The Last Hours” is a fictional response to David Leavitt’s creative essay “Toward a Short History of AI Paranoia.” We began with a prompt we had sent to our fiction contributors: How do we maintain a sense of self when the boundaries between the virtual and real worlds dissolve? We sent this prompt through the closed MIT platform, with specific guidelines for word count. AI gave us the first draft of the story, with its title. We then turned to a proprietary platform, Teach Anything AI, which was brought to our attention by Alexa Alice Joubin, one of its developers. We uploaded the story, along with select passages from David Leavitt’s essay. We gave the system the direction to incorporate information from the essay into the story, with a focus on Alan Turing.

We took this draft of the story to the public AI platform Claude and moved it through a series of revisions. We directed Claude to rewrite the story, get rid of clichés, and narrate from the point of view of Sim Lian. The resulting draft left out important sections from the earlier draft, so we added directions to extend the story and restore specific passages. We then uploaded a passage from E. M. Forster’s story “The Machine Stops” and we directed Claude to include a scene treating this passage. In the final prompt for this stage of revision, we gave directions to incorporate an additional passage treating the differences between the physical and virtual characters, Physical Lian and Digital Lian. 

We shared the text with the editorial team at Dædalus. In their consensus, “The Last Hours” lacked the verbal ingenuity that is expected in original fiction and would be on display in the other stories in the issue. We returned to Claude and sent back the full text with instructions to “take this story and make it more literary with the following changes: Combine paragraphs, add more imagery and figurative language, and make the sentence structure more varied and more complex.”

We had a new draft, but we weren’t done yet. We convened a virtual forum for the contributors of this issue and shared the latest iteration of “The Last Hours.” It was met with a range of responses, some more positive than others, but with general agreement that the piece should be improved. We took the story through several more drafts on the MIT platform, using prompts drawn from the forum. We gave directions for a revision that included additional scenes, with a word count between three thousand and three thousand five hundred. We asked for more details about the settings and characters, specifically, to make Sim Lian and Turing more complex characters and enhance the connection between their two narratives. Regarding the character Mira, we asked for changes that would reinforce the relevance of her role. We instructed AI to design a climactic event that would increase the suspense of the story. Loss was identified as a central aspect of this story, and so we sent a general prompt “to drill down on that, to go deeper into that sense of loss.” And finally, we returned to Claude and sent in a portion of the text and a prompt to devise a fictional name and biography for the author of “The Last Hours”—Bridget Voss was the result.

With such an elaborate series of prompts shaping this story, readers may wonder if we as editors assumed too much control over the process. We resisted making any substantial revisions in the text itself, but is there a point when prompting becomes a form of writing? Does our account of the prompts change your own assessment of the story?

We offer “The Last Hours” as the outcome of an experiment. Readers may find the AI story indistinguishable from an original story composed by a human author, or the imprint of the machine may be obvious. Readers will decide whether an AI-generated story has intriguing merits of its own, or whether the predictive patterning behind the writing process inevitably shows through the lines, making it the opposite of surprising.