Paras Chopra, an independent researcher who publishes the newsletter Inverted Passion, ran a small experiment testing whether a modern language model can detect, from a few sentences alone, that another AI system wrote them.

He fed a dozen September 2026 news headlines to GPT2-medium, a small text-prediction model OpenAI released in 2019, and let it write a continuation. Then he handed that same continuation, with GPT2’s identity hidden, to Qwen3’s 4 billion parameter base model and asked it to keep writing. Separately, he asked Qwen3 to write its own continuation of the same headline from scratch, with no GPT2 text involved.

The comparison is the whole experiment. When Qwen3 picked up where GPT2 left off, did its writing resemble what GPT2 itself would have written next, or did it resemble Qwen3’s own natural style? Chopra reports that across several ways of measuring text overlap, Qwen3’s completions of GPT2’s prefixes lined up more closely with GPT2’s own hidden continuations than with Qwen3’s completions of its own prefixes.

He ran a second check to rule out coincidence. He appended the phrase “Year: 20” to full GPT2-authored articles and to full Qwen3-authored articles on the same headlines, then asked Qwen3 to fill in the rest of the year. GPT2 launched in 2019, so its training data predates that year. Qwen3 consistently guessed an earlier year for the GPT2 text than for its own, even after Chopra removed passages that already contained a year and could have simply repeated it.

In one example built around a headline about a town testing a rotating bus schedule, GPT2’s hidden continuation described a poll displayed on “electronic signs at stops.” Given that same GPT2 prefix, Qwen3 continued with a similar civic-poll mechanism. Given only the bare headline with no GPT2 text at all, Qwen3 invented a different mechanism, a text-message vote. The pattern repeated in a second example about air quality and child development: Qwen3’s completion of GPT2’s prefix tracked GPT2’s own hidden ending more closely than it tracked Qwen3’s unprompted version of the same story.

Chopra frames this as tentative support for a hypothesis first raised in posts on the forum LessWrong: that because so much of the modern internet is now AI-generated text, current models may have absorbed rough internal models of other AI systems, the way they model human writers, in order to predict what comes next more accurately. He is careful to call his own work “a quick-and-dirty exploratory study,” not proof, and notes he tested only a 4 billion parameter base model plus a repeat run on a 14 billion parameter Qwen base model, not an instruction-tuned assistant, since those models do not perform plain text completion well.

The finding has not been replicated by an academic lab or reproduced at scale, and Chopra’s own methodology, hand-picked headlines and manual overlap comparisons, leaves plenty of room for confounds. But the underlying question matters beyond one blog post: if large models build internal representations of other models’ outputs, that same mechanism could eventually support tools for flagging AI-written text, or work against them if models learn to imitate a target system’s fingerprint on demand. Anyone building AI-detection products should treat this as a reason to test whether their classifiers hold up against a model that has learned to write like its rival, not just against generic machine text.

Reported by Paras Chopra on Inverted Passion (Substack) on 22 September 2026.