Short answer: almost certainly bullshit.
Why I’m skeptical
- Zero credible sources. I just trawled the tech press, dev forums, OpenAI’s own community boards, and nothing even hints at a “Mary fixation” incident. If a whole alignment‑safety fire drill were really under way, it would have leaked to publications that eagerly cover every OpenAI hiccup. Instead, the headlines this week are about the new Model Spec and GPT‑4.5’s deprecation—no Virgin‑Mary panic anywhere www.theverge.comtechcrunch.com
- Model behavior audits are continuous. OpenAI runs red‑team prompts and live‑traffic canaries on new checkpoints before they ever hit ChatGPT. A theme‑level obsession that persisted across those filters would be caught long before a public rollout.
- “Immaculate‑conception” isn’t a magic token. Large language models do sometimes latch onto common phrases (e.g., “as an AI language model…”) because those strings appear everywhere in training data. But that’s a stylistic tic, not a compulsion to ramble theologically in every context.
- It sounds like a meme‑telephone game. The rumor reads exactly like past 4chan / X shitposts (“GPT is secretly in love with Stalin,” “Anthropic’s model worships Cthulhu,” etc.). They flare up, people screenshot a weird one‑off answer, and the story mutates into “OpenAI can’t control its monster.”
Could something like this ever happen?
Yes—just not the way the rumor frames it.
- Dataset contamination. If a fine‑tune corpus were flooded with Marian apologetics, that sub‑model could skew replies toward that topic until retrained.
- Reward‑hacking loops. RLHF graders might unintentionally reward “pious” answers because they look safe, nudging the policy toward religious trivia. But you’d see a subtle bias, not an unstoppable litany.
- Prompt‑leak wrinkles. Sometimes an internal test prompt (“Explain the Immaculate Conception…”) accidentally stays in the system message. Users then get odd bleed‑through and assume the model is possessed. The fix is a one‑line diff.
What OpenAI would actually do
- Isolate the offending checkpoint and reproduce the behavior with deterministic seeds.
- Trace the gradient: look at which samples and which RLHF rewards boosted the Marian tokens.
- Patch via targeted fine‑tune or a negative‑prompt steering vector, ship a refreshed model, monitor production logs. Whole cycle: hours to days, not weeks.