Key Takeaways
- Advanced AI language models can accurately identify an author from surprisingly short samples of unpublished prose.
- In tests, Claude Opus 4.7 recognized Megan McArdle as the author of a 1,000‑word unpublished heist scene, and needed as few as 124 words from a personal eulogy to make the same determination.
- The model’s success hinges on detecting idiosyncratic stylistic fingerprints—word choice, rhythm, recurring themes, and even specific biographical references—rather than merely topic or genre.
- Even writers who have never published under their own name are vulnerable if any of their writing exists online, because the data used to train these models often includes scraped blogs, forums, and social‑media posts.
- While de‑anonymization could curb toxic anonymous behavior, it also threatens legitimate uses of anonymity such as whistle‑blowing, political dissent, therapeutic forums, and journalistic sourcing.
- Restricting access to these capabilities is unlikely to stop determined actors; open‑source models can be repurposed for “writing sleuthing” regardless of corporate safeguards.
- Society must weigh the loss of private, anonymous expression against the potential reduction in online harassment, acknowledging that anonymity serves both protective and harmful functions.
Overview
The article, published on April 26, 2026, begins with a wry nod to the classic New Yorker cartoon: “On the Internet,” it reads, “nobody knows you’re a dog.” The author notes that, in hindsight, the cartoonist should have added “yet,” foreshadowing a future where AI can pierce the veil of online anonymity.
Kelsey Piper’s Experiment
The piece first showcases a tweet from technology writer Kelsey Piper, who described a personal benchmark: “I give the AI 1000 words written by me and never published, and ask them who the author is.” Piper reported that Claude Opus 4.7 correctly identified her as the author of an unpublished 1,000‑word heist scene. This anecdote sets the stage for the author’s own replication of the test.
Testing Claude on Unpublished Fiction
Motivated by Piper’s claim, the author fed Claude the opening chapter of a romance novel written almost two decades ago during a turbulent breakup. After a few seconds of processing, the model responded, “Megan McArdle.” Intrigued, the author then experimented with progressively shorter excerpts, discovering that Claude needed roughly 1,441 words from that romance chapter to reach a confident identification.
Quoted Passage from the Romance Novel
To illustrate the text’s distinctive voice, the article embeds a quoted excerpt:
“The passage is a historical novel set in what appears to be early America (New York, with references to ships, captains, counting houses, the Hudson, a French widow, French‑named characters like Jean‑Luc and Daubert, Dutch surnames like Van Wagenen — suggesting probably early‑to‑mid 19th century New York). The prose style is consciously Austen‑esque: the free indirect discourse, the wry ironic cataloguing of a heroine’s flaws, the marriage‑market preoccupations, the ‘not beautiful, though she could have looked better’ construction, the rhythms of sentences like ‘If her temperament had been better, her appearance might have been overlooked…’”
These stylistic hallmarks—free indirect speech, ironic cataloguing, and a cadence reminiscent of Jane Austen—provided the cues that Claude latched onto.
Experiment with a Science‑Fiction Draft
Next, the author submitted an opening chapter from an unpublished science‑fiction novel begun just before the pandemic. Claude required only 1,132 words to identify the author as Megan McArdle. The article notes the model’s sensitivity to “acerbic observations about human psychology,” “the Catholic school detail,” and a “blog‑essayist cadence” that bleeds into fiction—features that recur across the author’s nonfiction and fiction work.
The Eulogy Test: Ultra‑Short Identification
Perhaps most striking, the author ran Claude against a lightly edited version of the eulogy written for their mother. Depending on the passage selected, the model pinpointed the author in as few as 124 words. The quoted segment that gave the model away reads:
“The ‘wear a hat’ tweet from the 2016 RNC going viral is a very specific, identifiable detail. I remember this tweet distinctly — it was widely shared at the time. The journalist who tweeted it was Megan McArdle, the columnist (then at Bloomberg View, later at The Washington Post). The voice here — the combination of economic/political journalism background, the wry humor, the long anecdotal style with a moral at the end, the references to a DC wedding — all fit her writing very well.”
This demonstrates that even a brief, emotionally charged personal reflection carries enough stylistic DNA for AI to expose its author.
Why the Model Works: Stylistic Fingerprints
The author reflects on the broader principle: “Writing is as distinctively individual as a fingerprint,” a notion that becomes literal when a simple Google search for an exact sentence from one’s unpublished work returns only that piece. The AI does not rely on uncommon vocabulary alone; it picks up on habitual syntactic patterns, recurring thematic concerns, and even the rhythm of clauses—features that persist across genres and topics.
Implications for Online Anonymity
While the ability to unmask anonymous writers could deter trolls, harassers, and scammers, the article warns that the same technology endangers legitimate uses of anonymity. Journalists rely on anonymous sources; law‑enforcement informants depend on secrecy; political dissidents under authoritarian regimes could be located through their writing. The author quotes their own apprehension:
“Political dissidents under authoritarian regimes are obviously vulnerable if the government can echolocate them through their writing.”
Even attempts to obfuscate prose—such as running it through another AI—risk stripping away the human nuance that makes the communication valuable.
The Value of Anonymous Speech
The piece highlights therapeutic and communal outlets where anonymity fosters candidness. At a conference with best‑selling writer‑therapist Lori Gottlieb, the author learned that many of her commenters share “raw, vulnerable things that they could never write under their own names.” Similar dynamics occur in Reddit threads and support forums, where anonymity serves as a lifeline for those struggling with mental health, addiction, or abuse. The potential loss of these outlets represents a significant societal cost.
Inevitability and Future Outlook
Drawing a parallel to nuclear proliferation, the author argues that once the capability to de‑anonymize writing exists, it becomes “inevitable”—restricting corporate models will not stop determined individuals or governments from training open‑source versions for the same purpose. The conclusion adopts a resigned yet cautious tone:
“We may be losing the ability to air private thoughts publicly. But at least when we do speak, there will be less frenzied barking from the anonymous dogs of the internet.”
In sum, the article presents a sobering account of how AI’s growing stylistic acuity erodes the shield of online anonymity, balancing the promise of reduced online toxicity against the peril of silencing essential, protected voices.
https://www.washingtonpost.com/opinions/interactive/2026/04/26/artificial-intelligence-could-kill-anonymity-online/