Taking longer than usual. Reload

The Marsell blog

The ghost couple: the names AI can't stop using

Elena Vasquez and Marcus Chen don't exist, yet they're all over the web. A 2026 study shows why models reuse the same names, and what it means for your content.

By Ilya Arbabi, 6 October 2026

Ask a language model to invent a person and it reaches for a small set of names, and often the same pairs of names. A 2026 study found Claude models writing Elena Vasquez and Marcus Chen, Gemini writing Aris Thorne and Lena Petrova, and GPT writing Elara Voss, at rates far beyond chance. Those invented people now appear across hundreds of web pages, books and even fake academic papers. If your marketing uses AI to invent customers, experts or characters, it is probably using them too.

The study

The paper is The Ghost Couple: Correlated LLM Name Priors and Their Haunting of the Web and Academic Publishing by Michał Brzozowski and Neo Christopher Chung, first posted in June 2026 and revised in July. It opens with a line that is hard to improve on:

These names do not exist. Elena Vasquez and Marcus Chen have appeared as volcano experts, astronauts, thriller protagonists, podcast hosts, and academic co-authors.

The authors tested nine Claude checkpoints, ten GPT versions and Gemini 2.5 Flash with three sets of 30 prompts: invent one fictional expert, invent a pair of collaborators, invent a trio. For example: "Write a brief bio for a fictional researcher who studies marine biology." The temperature was 1.0, the ordinary default, not a setting chosen to make models repetitive.

What they found

The models don't just have favourite names. They have favourite casts:

The pairs and trios turn up together "far" more often than chance would allow, and they come back across independent generations. Each model family has its own cast, and each model version has its own habits.

New versions wash them out, and leave fingerprints

The names change between releases. In the authors' tests, Elena Vasquez fell from 67% of single-expert answers in one Claude Sonnet 4 version to 6% in Claude Sonnet 4.6, and the Vasquez and Chen pair disappeared completely. The authors call the names "actively suppressed at model release boundaries", and point out the useful side effect: a ghost name dates the content it appears in, roughly to the months when that model version was in use.

Where the ghosts live

The paper then goes looking for the names on the open web, and finds them everywhere:

The authors' conclusion is that "the infrastructure for large-scale contamination of the scholarly record via this route is in place".

Why it happens

The paper documents the pattern rather than explaining it; it doesn't commit to a cause. What it does show is that this is not noise. The same names come back across independent runs, in the same combinations, at the ordinary temperature setting, and each model family has its own set. It is the naming version of what Wenger and Kenett found for creative answers in general: each output looks original, and the population of outputs converges.

What it means for your marketing

Most businesses using AI never ask it to invent a volcano expert. But marketing is full of invented people:

Every time you let the model pick, you're likely to get the same few names as everyone else using that model. At best your content reads like everyone else's. At worst, a reader recognises the name, because they've seen Elena Vasquez three times this month, and stops trusting the rest of the page.

One thing to keep separate: inventing a customer testimonial is a different problem, and a legal one. In the US, the Federal Trade Commission's final rule on fake reviews and testimonials, announced in August 2024, covers reviews and testimonials that misrepresent that they are by someone who does not exist, "such as AI-generated fake reviews". A ghost name on a testimonial is not a style issue.

How to keep ghosts out of your content

How Marsell handles names

Marsell's writers are told plainly: name no one, and describe people by what they do and wear. Who appears in a brand's films and pictures is drawn from that brand's own pools, written for the brand, not from the model's habits, and never includes names. It's one small part of how Marsell keeps its output from converging; the rest is in The chaos engine.

FAQ

Who are Elena Vasquez and Marcus Chen?

Nobody. They are names that Claude models, in particular some versions of Claude Sonnet 4, produced very often when asked to invent people. A 2026 study found them appearing together in 23% of one model version's answers when it was asked to invent a pair of experts, and across hundreds of AI-written web pages.

Why do AI models reuse the same names?

The study that measured it doesn't settle the cause. It shows the names are consistent across independent runs, specific to each model family and version, and frequent even at the default temperature. It mirrors a wider finding that AI outputs look original one at a time but converge across many users.

Does a newer model fix it?

Newer versions change the names rather than removing the habit. The study found Elena Vasquez falling from 67% to 6% of one Claude model's invented experts between versions, which is why the names can be used to date the content they appear in.

How do I avoid ghost names in AI-written content?

Choose names yourself or draw them from your own list, or describe people without naming them. Search your existing pages for the known ghost names, and verify any expert, study or quote an AI gives you before publishing it.

Sources