Two stories crossed my feed in the same week, and on the surface they have nothing to do with each other. One is a lab in Minnesota claiming it built a cell from scratch. The other is a study that combed through half a million ChatGPT conversations and found a person who asked, thousands of times, for a fanfiction scene in which a schoolgirl goes into labor in a literature clubroom. Different worlds. Same shape underneath. Both are machines that look like they create on their own, and both quietly depend on parts they cannot produce.
Start with the cell, because it is the more serious of the two.
SpudCell does a lot, and cannot make its own ribosomes
Kate Adamala and Aaron Engelhart at the University of Minnesota announced on July 1 what they call the first synthetic cell with a complete life cycle. The name is a play on Sputnik, shortened from an earlier joke, Potato Cell, a nod to Adamala's Polish heritage. What they built is genuinely clever. Rather than editing an existing organism, which is the usual route, they went bottom up: water-filled spheres called liposomes packed with synthetic DNA, enzymes, and molecular machinery, assembled until the system could run through an entire cycle on its own. The genome is about 90,000 base pairs spread across separate plasmids. For scale, E. coli carries 4.6 million. It feeds by fusing with smaller liposomes that deliver lipids and nutrients, and it divides not with a cytoskeleton but by crowding proteins at its membrane until mechanical stress splits it in two. The cell's own DNA controls how fast it grows and whether it divides. That is real, and it is a nice piece of work.
Now read the part the headlines skip. SpudCell cannot make its own ribosomes. It borrows ribosomes from E. coli, supplied through feeding. Without the ability to remake that machinery, it runs for five to ten generations before the whole thing degrades and stops. It cannot regulate its own metabolism, cannot survive outside the carefully stocked liquid around it, and because it cannot keep reproducing over many generations, it cannot evolve. Drew Endy at Stanford put it plainly: Adamala has constructed a cell, but he does not think she has created life. Adamala herself hedges, saying life is not binary and she cannot definitively call it alive. That honesty is the right posture. "Scientists have created a synthetic cell that can eat, grow and reproduce," as the headline framing goes, is technically defensible and materially misleading. What they created imitates the cell cycle for a few generations on borrowed parts.
There is a reason to hold the framing at arm's length. The paper was not peer reviewed when this landed. Adamala says Cell rejected it after a reviewer called SpudCells not real biology, and she then sent a 190-page manuscript to journalists under embargo, before even posting it to the preprint server bioRxiv. A synthetic biologist at Heidelberg called that an unusual way of doing things. It is. The incentive to convert a contested result into a settled headline is strong, and the press obliged. Treat every performance claim here as provisional until someone independent reproduces the five-to-ten-generation run.
I want to be fair, because the ambition is pointed at things that matter. The team launched Biotic, a public-benefit outfit meant to license the core technology as a shared standard, the way Linux became an operating system everyone builds on. The longer aim is to harness biology to produce drugs, biomaterials, and eventually petrochemicals, cutting reliance on fossil fuels. That is a real problem worth a real tool. Just do not confuse the tool with the god-move the headline promises.
The fanfiction machine, and the loop it runs in
The second story wears a sillier costume. Neel Gupta, Maria Antoniak, and Melanie Walsh, working across the University of Washington and Colorado Boulder, published a preprint called "AI Fiction in the Wild." They did not collect new data. They mined WildChat, a public set of real ChatGPT conversations that the Allen Institute for AI gathered in 2023 and 2024 by hosting free chatbots, where users consented, twice, to having their chats released. From roughly 573,000 English conversations, more than a third involved some form of fiction: original stories, roleplay, fanfiction, erotica.
The distribution is the interesting part. The top two percent of users produced more than 80 percent of the fiction. Around half of the fiction prompts were fanfiction and over a quarter were sexually explicit, much of it slipping past the model's rules through jailbreak prompts copied straight off social media. Repetition sits at the center of the whole thing. Average users with more than one conversation produced near-duplicate prompts about 42 percent of the time. For power users it climbed to 69 percent. People wanted the same tropes, romance, time travel, vampires, happy endings, over and over, each version a fraction different from the last.
Which brings us to the outlier the bookmark seized on. One person, over several months, prompted ChatGPT thousands of times to write a scene in which the Doki Doki Literature Club character Natsuki suddenly goes into labor in the clubroom. The technique is visible in the paper itself.
This single user inflates the franchise counts so badly that Doki Doki Literature Club tops the list with 22,381 mentions, far ahead of the next games, which tells you more about one obsessive than about the world. The authors coin two profiles for this behavior. Story cyclers iterate on one story for a while, then move on. Infinite story demanders ask for the same thing endlessly, sometimes for months. The DDLC user is the flagship of the second kind.
The line that stayed with me is the paper's argument that this may be producing a solipsistic reader-writer, someone who generates and consumes fiction inside a closed conversational loop, talking to a machine instead of another person. The authors compare it, carefully, to how pornography works: familiar genres, instant delivery, repetition, niches, momentary gratification. They do not claim all AI fiction is this. Plenty of people share and publish what they make. But the pattern is there, and it is worth naming.
What the two share, and where they part
Both systems perform an act we associate with autonomy. One performs living. One performs authorship. Neither closes the loop. The cell cannot manufacture the ribosomes that keep it going, so it borrows them and dies within a handful of generations. The fiction machine cannot originate the desire it serves, so it recycles the user's own prompt back at them, slightly reshuffled, until the person feeds it again. In both cases the appearance of self-sustaining creation depends on a supply line the thing itself never built, and in both cases the public framing sands that dependence away.
The difference is where they point. SpudCell borrows machinery in order to reach outward, toward drugs and materials and problems that kill and impoverish people. The fiction loop borrows readers in order to close inward, toward a single person and a screen. I would rather my curiosity, and my skepticism, spend most of its time on the first. But it is the second that shows you, cleanly and a little sadly, what a closed loop feels like from the inside.