Gemini's rewrites turn a Chandler passage into horror after 31 iterations

The author says the blog rarely covers generative AI, since little written about it says anything of consequence, but explains, from a security background, a specific interest in how AI models fail. Recursive processes, in which an AI system feeds on its own output instead of fresh input, get special attention: the author calls this the core mechanic behind AI agent loops, and says it also sits at the center of two opposing predictions about where AI is headed, the model collapse that AI skeptics warn of and the rapid self-improvement that AI optimists expect.

As a first illustration, the author made a short video by asking the image and video model Nano Banana 2 to generate a fake movie poster, then repeatedly asking it to make a localized edit to just the title. The author's point: this kind of recursive AI image and video editing tolerates only a limited number of tries before the accumulated errors break down into what the author calls pure chaos.

The author then asks whether the same collapse happens with text or code, and answers yes and no. Text input and output encoding is lossless and injects little per-token randomness, so a model asked to make a single, well-specified change, such as swapping one word in a paragraph, can almost always do it cleanly, with no damage to the rest of the passage. Open-ended rewrites behave differently: most do not diverge forever, the author says; instead the output thrashes for a while and then settles into a stable result the author calls "peak LLMese," though that settled result can still end up far from where it started.

To show this, the author took a passage from Raymond Chandler describing an oversized, mirror-and-crystal-filled room and the narrator's appraisal of a woman named Mrs. Regan, then asked Google's Gemini to "boldly rewrite" it to "improve tone, clarity, and flow." Each subsequent round fed the model only its own immediately preceding output together with the same instruction, never the original passage and never the full history of earlier rewrites.

On the very first pass, Gemini reverted to what the author calls its worst habits: metaphors that sound sophisticated but are just obtuse. The room became a "vast, towering tomb of bleached opulence," the drapes "pooled greedily on the carpet," and the air "sat thick with the threat of rain." By iteration 8 the language had turned, in the author's words, "truly weird": mirrors "weaponized the brutal glare," Mrs. Regan's legs were "sculpted for ruin," and her mouth carried "an insatiable hunger for destruction." By iteration 12, which the author calls "uncharted territory," the carpets were "suffocating the room in sterile, unyielding wealth," the storm was "waiting to shatter the floor-to-ceiling glass," and Mrs. Regan had become "pure, unapologetic malice ... effortlessly cruel." By iteration 16, opulence "didn't merely occupy the room; it annihilated it," the carpet "devoured the space," Mrs. Regan's legs "cut through the air like stiletto blades," and her eyes were "burning with the singular urge to destroy."

Around iteration 20 the escalation stopped accelerating and instead began cycling through a small, repeating set of near-synonymous phrases: the decor variously "stripping," "scrubbing," or "eradicating" traces of humanity, and Mrs. Regan as "pure venom" or "concentrated venom" who intends to "dismantle," "tear apart," "ruin," or "flay" the narrator. By iteration 31 the text locked onto one fixed version, which the author quotes in full: the room now "turned its opulence into a weapon," the light "scorched pristine silk, razor-honed chrome, and spotless carpet," and Mrs. Regan, described as "venom wrapped in high fashion," delivers "a single guarantee: she was about to ruin me."

The author closes by joking, "who knew that Raymond Chandler had such a knack for horror," and adds, for transparency, that the post will fail Pangram's AI-writing checks, since it necessarily quotes text actually written by AI: everything in the post except the marked Gemini excerpts is human writing. The author signs off by noting a rare departure from a usual policy of not writing about large language models, and points to an unrelated, recent woodworking project as a way of making up for that lapse.

Key facts

The author repeatedly fed Google's Gemini its own previous output plus the same "boldly rewrite ... improve tone, clarity, and flow" prompt, applied over and over to a single passage from Raymond Chandler.
By iteration 8 the rewrite had turned "truly weird" (mirrors that "weaponized the brutal glare," legs "sculpted for ruin"); by iteration 16 it had escalated further, with opulence that "annihilated" the room and legs that "cut through the air like stiletto blades."
Around iteration 20 the drift stopped compounding and began cycling through a small set of near-synonymous violent phrases; by iteration 31 the rewrite locked onto one fixed, horror-toned version, quoted in full in the post.
A separate test with the image and video model Nano Banana 2, repeatedly asked to edit the title on a faux movie poster, shows the same recursive-feedback breakdown hits images even faster, within a handful of edits.
The author argues text degrades far more slowly than images because lossless encoding and low per-token randomness make small, well-specified edits nearly deterministic, and frames the exercise as a hands-on look at the recursive self-feeding that also drives AI agent loops and fuels both model-collapse warnings and rapid-self-improvement hopes.

Why it matters

Recursive self-feeding, where an AI system is handed only its own last output and asked to act on it again, is the literal mechanism inside AI agent loops, and it sits at the center of two opposing stories about where AI is headed: the model collapse that skeptics warn of, and the rapid self-improvement that optimists expect. This post does not settle that debate, but it gives a small, reproducible, and oddly funny answer for one narrow case: repeated open-ended rewriting does drift, sometimes into outright violent language, yet it does not diverge forever. It thrashes, then locks onto a stable pattern that can still be far from the starting point. Anyone building a pipeline where a model's output becomes the next step's input is looking at the same dynamic, just with prompts and outputs less colorful than a Raymond Chandler passage turned horror story.

Who it affects

Builders and users of multi-step AI agent pipelines, where one model call's output feeds the next; anyone running iterative AI rewriting, tone-polishing, or image and video editing tools that loop a model's own output back into itself across several passes, intentionally or not; and readers following the model collapse and rapid-self-improvement debates who want a concrete example rather than another abstract argument. The stakes are illustrative rather than operational: this is one blog author's own passage and prompt, not a benchmark result anyone is shipping against.

How to use it

There is no product or setting to configure here, only a pattern worth borrowing. The post suggests keeping unsupervised, open-ended self-referential rewriting passes few: real drift is visible within single digits of iterations, and the image and video case broke down within just a handful of edits. Narrow, well-specified edits, such as changing one word in a paragraph, appear safe to repeat many times, since the author found those close to deterministic. And when a pipeline must run many recursive passes regardless, it is worth checking whether the output has settled into a stable pattern rather than assuming it will keep drifting without limit, since that is what happened in both the text and, eventually, the image experiment.

How solid is it

This is one person's own demonstration, posted to a personal blog and submitted to Hacker News, where it drew 53 points and 14 comments, not a peer-reviewed study or a benchmark with repeated trials. It runs exactly one text example (one Chandler passage, rewritten by an unspecified version of Gemini, through 31 logged iterations) and one image and video example (Nano Banana 2 editing a poster title, with no iteration count given), each a single run rather than a series averaged across prompts or models. The author is explicit that model collapse and rapid self-improvement are offered as context for why the topic is interesting, not as something this particular experiment claims to measure or prove.

Risks and caveats

The result is anecdotal: a single passage, a single unversioned Gemini setup, and no comparison against other models or prompt phrasings, so the exact iteration counts (8, 12, 16, about 20, 31) describe this one run rather than a general rate. The post does not name the novel the Chandler passage comes from, does not date when the experiment ran, and gives no iteration count for the separate Nano Banana 2 video, so that example cannot be compared quantitatively to the text one. The escalation into violent, threatening language is also a caveat in its own right for anyone chaining generative rewrites unattended: nothing in the prompt asked for violence, yet repeated self-feeding produced it anyway before settling down.

“The room turned its opulence into a weapon, stripping away every trace of human warmth. Blinding, clinical light scorched pristine silk, razor-honed chrome, and spotless carpet. Beyond the glass, a bruised storm raged; within, leaded crystal and mirrors shattered the glare into jagged shards.”

— Gemini, rewriting a Raymond Chandler passage after 31 recursive iterations