OpenAI's GPT-5.2 Allegedly Cites Grokipedia, Revealing Tangled AI Training Data

OpenAI's latest model appears to be citing its competitor's AI-generated content as a source. According to a report from Engadget, GPT-5.2 has been observed referencing Grokipedia—the AI-generated encyclopedia created by Elon Musk's xAI—in its outputs. If verified, this represents a significant data point in understanding how today's frontier models are trained and how they're increasingly feeding off each other.

GPT-5.2 Surfaces Amid Training Data Questions

The report's existence confirms what many had suspected: OpenAI has quietly rolled out or is testing GPT-5.2, a model that hadn't been formally announced. The company's last major release was GPT-5, which launched in late 2025 to significant fanfare. A 5.2 iteration suggests incremental improvements—possibly to reasoning, knowledge cutoff, or response quality.

But the real story here isn't the version number. It's what GPT-5.2 apparently ingested during training.

Grokipedia launched as xAI's answer to Wikipedia—an AI-generated knowledge base powered by Grok, designed to provide real-time information synthesis. Unlike Wikipedia's human-edited model, Grokipedia content is machine-generated, which makes its appearance in a competitor's training data particularly noteworthy.

The Ouroboros Problem Gets Real

AI researchers have warned for years about "model collapse"—the degradation that occurs when AI systems train on AI-generated content. The concern isn't theoretical anymore. If OpenAI's models are ingesting Grokipedia content, and xAI's models are presumably trained on web data that includes ChatGPT outputs, we've entered a feedback loop where AI systems are increasingly learning from each other rather than from human-generated sources.

This creates several problems:

• Provenance becomes impossible. When a model cites Grokipedia, it's citing AI-generated synthesis of sources—not the sources themselves.
• Errors propagate. AI hallucinations in one system become training data for another, potentially cementing false information across models.
• Quality standards blur. Human-curated knowledge bases like Wikipedia have editorial processes. AI-generated ones don't—or at least, not the same kind.

What This Reveals About OpenAI's Data Practices

OpenAI has been notoriously opaque about its training data sources. The company faced multiple lawsuits in 2024 and 2025 over alleged copyright infringement, with publishers and authors claiming their work was used without permission. OpenAI has consistently declined to detail what's in its training corpus.

The Grokipedia citation suggests OpenAI's data pipeline includes recent web scrapes that capture AI-generated content—either intentionally or as an unavoidable consequence of crawling the modern web. This raises a question the company will eventually have to answer: Does OpenAI filter AI-generated content from its training data, and if so, how effectively?

The Competitive Angle

There's also an irony worth noting. Elon Musk has been one of OpenAI's most vocal critics, filing a lawsuit against the company and repeatedly attacking CEO Sam Altman on X. That his company's AI-generated content is now apparently showing up in OpenAI's outputs adds a layer of absurdity to an already contentious rivalry.

For users, the practical implication is straightforward: when you ask an AI a question, you may now be getting an answer that's been filtered through multiple layers of machine synthesis, with no clear path back to original human sources.

The AI industry has spent years debating whether models can be trusted. This revelation suggests a different question might be more pressing: Can they even cite their sources accurately when those sources are themselves AI-generated?