On June 21, a debate erupted over a phenomenon dubbed AI 'ghost names' — fictional personas like 'Elena Vasquez' that generative AI systems repeatedly produce. The discussion, which played out across tech forums and social media, centers on why these systems lean on such fabricated identities and what that means for model reliability. Prediction market Polymarket currently places an 84% probability on Anthropic being the company most associated with the issue, though the debate itself extends beyond any single firm.
Why AI keeps using fake names
Generative AI systems, including large language models, generate text by predicting the most probable next word based on training data. Sometimes that probability calculus lands on a name that appears nowhere in the real world — a pure statistical artifact. 'Elena Vasquez' is one such example, a character that has shown up in outputs from multiple models. Researchers note that this isn't a glitch in the traditional sense; it's a byproduct of how these systems learn patterns from vast datasets.
The problem intensifies when these ghost names get reused across interactions. A feedback loop takes hold: the AI sees its own made-up name in prior outputs, treats it as a valid signal, and becomes more likely to generate it again. Over time, the fictional identity can become self-reinforcing, nudging the model away from factual accuracy and toward a kind of digital hallucination.
Polymarket's bet on Anthropic
Polymarket, the decentralized prediction platform, shows an 84% probability that Anthropic — the AI safety company behind the Claude model — is the key player in the ghost names debate. The market, which opened shortly after the June 21 discussion, reflects bettors' views on which organization's systems are most entangled with the issue. No official statement from Anthropic has been released, and the company hasn't confirmed any internal findings related to ghost names.
The high probability doesn't necessarily mean Anthropic is the worst offender; it could signal that the company's research transparency makes it a focal point. Critics argue that ghost names aren't unique to any one model — they appear across OpenAI's ChatGPT, Google's Gemini, and others — but the Polymarket figures suggest traders see Anthropic as the company most likely to address or be affected by the phenomenon.
The feedback loop problem
At the heart of the ghost names debate is a broader concern about AI behavior loops. When a model generates a plausible but fictitious name, and that name later appears in its training data or in user prompts, the system can strengthen its association with that fake identity. This isn't limited to names; it can apply to any invented fact. The difference here is that ghost names are particularly sticky because they're human-like and easy to remember.
The feedback loop can degrade output quality over time, especially in applications that require consistent factual grounding. For customer service bots or news aggregators, a model that insists on referencing 'Elena Vasquez' might frustrate users or spread misinformation. Developers are still figuring out how to break the cycle — one approach involves filtering out low-probability entities from training data, but that risks removing rare but real names.
The debate on June 21 didn't produce a resolution. Instead, it set the stage for further investigation. Researchers are calling for more systematic tracking of phantom names across models and for clearer documentation of how these artifacts emerge. The Polymarket odds may shift as new evidence comes to light. For now, the question remains open: can AI systems be trained to stop inventing people, or will 'Elena Vasquez' keep showing up in conversations?
