💻 computer science

The Oracle's Fingerprint: Correlated AI Forecasting Errors and the Limits of Bias Transmission

This paper reveals that while three independently developed large language models exhibit highly correlated forecasting errors creating an "epistemic monoculture," this shared bias has not yet been transmitted to human forecasters, who instead continue to rely on their own pre-existing biases and rational updates toward ground truth.

Original authors: Theodor Spiro

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Theodor Spiro

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Three Wise Men" Who Are Actually One

Imagine you are trying to predict the weather. To get the best answer, you ask three different experts: a meteorologist from New York, one from London, and one from Tokyo. You expect them to give you three slightly different opinions because they have different training and data. If they do, you can average their answers to get a very accurate forecast. This is the "Wisdom of Crowds."

This paper asks a scary question: What if those three experts are actually just one person wearing three different masks?

The researchers tested this by asking three of the world's most famous AI models (GPT-4o, Claude, and Gemini) to predict the outcomes of 568 real-world events (like "Will a specific political event happen by a certain date?").

Study 1: The "Echo Chamber" Effect

The Finding: The three AIs were not independent. They made the exact same mistakes, in the exact same direction, at the exact same time.

The Analogy: Imagine you ask three different fortune tellers to predict a coin flip. If they are truly independent, sometimes one says "Heads" and another says "Tails." But in this study, the three AIs were like three parrots that had all learned the same song. When the coin was actually going to be "Tails," all three parrots confidently screamed "Heads!"

The Stat: Their errors were correlated at 0.77. In the world of statistics, this is incredibly high. It means that if you consult all three models, you aren't getting three different opinions; you are getting one opinion with a little bit of static noise.
The Conclusion: We have built an "epistemic monoculture." Just as planting only one type of crop makes a farm vulnerable to a single disease, relying on these three AIs makes our collective judgment vulnerable to a single type of error.

Study 2: Did Humans Start Copying the AIs?

The Question: Now that these AIs are available to everyone, are human forecasters starting to copy their mistakes? Did the "parrots" start teaching the humans to sing the wrong song?

The Experiment: The researchers looked at a group of elite human forecasters on a website called Metaculus. They compared how these humans predicted things before ChatGPT was popular (November 2022) and after.

The Finding: Surprisingly, no. The humans did not start shifting their predictions to match the AIs.

The Analogy: Imagine the AIs are a loud, confident tour guide leading a group of hikers in the wrong direction. The researchers checked if the hikers started following the guide. They didn't. The hikers kept walking their own path, which happened to be the correct one.
Why? The humans on this specific website are experts. They are like seasoned detectives who know how to spot a lie. They didn't blindly trust the AI.
The Caveat: The study wasn't perfect at spotting small changes (it was a bit like trying to hear a whisper in a noisy room), but it found no evidence that the experts were being "brainwashed" by the AI yet.

Study 3: The "Fingerprint" Reveal

The Question: If the humans aren't copying the AI, where did the AI get its weird biases in the first place? Did the AI invent new, alien ways of thinking?

The Finding: The AI didn't invent anything new. It simply mirrored the biases humans already had.

The Analogy: Think of the AI as a mirror. Before you even put on a funny hat, the mirror showed you a face that looked like you. The researchers found that the pattern of mistakes the AI made before it was even widely used was almost identical to the pattern of mistakes humans were already making.

The Twist: After ChatGPT launched, the humans actually started to diverge from the AI. The humans seemed to realize, "Hey, this AI is overconfident about technology topics," and they corrected themselves, while the AI kept making the same old mistakes.
The Conclusion: The AI is not a new source of error; it is a bias amplifier. It took the biases humans already held (like thinking technology will solve everything or underestimating geopolitical risks) and made them louder and more confident.

The Final Takeaway

The paper concludes with a warning: The "Monoculture" is built, but not yet activated.

The Infrastructure is there: The three major AIs are essentially the same brain. If you use them all, you aren't diversifying your thinking; you are just reinforcing the same blind spots.
The Danger isn't "New" Errors: The AI isn't going to teach us crazy new ways to be wrong. It's going to confirm the wrong ways we are already thinking, making us feel more confident about them.
The Current State: On expert platforms, humans are still doing better than the AI and aren't blindly following it. But as AI becomes a standard tool for everyone (not just experts), the risk is that we will stop checking our own work and start trusting the "Oracle" that is actually just a mirror of our own collective biases.

In short: The AI is a very confident mirror. It reflects our own mistakes back at us with such authority that we might stop checking if the reflection is true.

Technical Summary: The Oracle's Fingerprint: Correlated AI Forecasting Errors and the Limits of Bias Transmission

Problem Statement

The "wisdom of crowds" relies on the statistical independence of individual errors; when forecasters err in idiosyncratic directions, their errors cancel out in aggregation. However, the widespread adoption of Large Language Models (LLMs) for decision support introduces a risk of epistemic monoculture. If human forecasters consult a small set of AI models that share systematic, correlated failure modes, the diversity of the "crowd" collapses. The paper investigates three necessary conditions for this monoculture to cause harm:

LLMs must exhibit systematic forecasting biases.
These biases must be highly correlated across different models (even those from different organizations).
Human decision-makers must absorb and propagate these biases into their own forecasts.

Methodology

The study utilizes data from Metaculus, an online forecasting platform with resolved binary questions and verifiable ground truth. The research is divided into three studies:

Study 1 (LLM Bias Map): The authors queried three major LLMs (OpenAI's GPT-4o, Anthropic's Claude Sonnet, and Google's Gemini 1.5 Pro) on 568 resolved binary questions. They calculated signed errors ( $e = p - o$ ) and measured pairwise error correlations between the models. To address potential training data leakage (where models might know the answers from their training data), they applied a leak detection heuristic and analyzed a "clean" subset of 414 questions.
Study 2 (Temporal Shift): Using a within-question design, the authors tracked community prediction shifts across the ChatGPT launch boundary (November 2022). They analyzed 179 questions open before and after this date. They regressed the community prediction shift ( $\Delta$ ) against the "resolution pull" (movement toward ground truth) and the "LLM pull" (the direction the LLM would pull the prediction). This controlled for rational updating to isolate specific LLM influence.
Study 3 (Bias Fingerprint): The authors categorized questions into five topical groups (Technology, Geopolitics, Economics, Health, Science/Environment). They compared the category-level bias patterns of LLMs against human community biases in pre-ChatGPT and post-ChatGPT periods. They calculated the correlation between the LLM "fingerprint" (vector of biases across categories) and the human bias vector for each period to determine if human biases were converging toward the AI pattern.

Key Results

1. High Inter-Model Error Correlation (Condition 1 & 2)

The primary finding is that distinct LLMs function as a single oracle with noise rather than independent sources.

Correlation: The mean pairwise error correlation between GPT-4o, Claude, and Gemini was $r = 0.77$ ( $p < 0.001$ ) on the full dataset and $r = 0.78$ on the clean subset.
Implication: This correlation is 3–5 times higher than typical correlations among human forecasters ( $r \approx 0.1–0.3$ ). The authors calculate that an ensemble of three such models is effectively equivalent to only 1.3 independent forecasters, offering negligible diversity gains over consulting a single model.
Bias Direction: All three models exhibited a positive directional bias (overestimating "Yes" outcomes) and overconfidence in the mid-range probabilities.

2. Lack of Detectable Transmission to Elite Forecasters (Condition 3)

Study 2 found no statistically significant evidence that LLM biases have propagated into Metaculus community forecasts beyond rational updating.

Regression Analysis: While the raw correlation between community shifts and LLM pull direction was positive ( $r = 0.20, p = 0.007$ ), this was fully explained by the community moving toward the ground truth (Resolution Pull, $\beta_1 = 0.061, p < 0.001$ ).
LLM Influence: The coefficient for LLM-specific influence ( $\beta_2$ ) was not significant ( $\beta = 0.023, p = 0.36$ ).
Caveat: The authors note low statistical power (23%) for detecting small effects and acknowledge that Metaculus forecasters are an elite, highly calibrated population, representing a "best-case scenario" for independence.

3. Inversion of the Fingerprint Hypothesis (Condition 3)

Contrary to the hypothesis that AI would imprint new biases onto humans, Study 3 found that human biases already matched LLM biases before ChatGPT was widely available.

Pre-ChatGPT: The correlation between the LLM bias fingerprint and human bias was strong ( $r = 0.87$ ).
Post-ChatGPT: The correlation weakened to $r = -0.28$ .
Interpretation: LLMs did not introduce novel biases; they inherited existing human cognitive biases from their training data. The weakening of the correlation post-ChatGPT may indicate an "awareness effect," where expert forecasters (particularly in Technology) recognized and corrected against AI overconfidence.

Significance and Claims

The paper concludes that an epistemic monoculture is "built but not yet activated."

The Infrastructure of Failure: The most robust finding is that the AI ecosystem lacks diversity at the model level. Consulting multiple major LLMs provides almost no additional independence compared to consulting one, as they share the same failure modes.
Bias Mirrors, Not Sources: The authors argue that LLMs act as bias mirrors. They reflect and rigidify the biases already present in the human information ecosystem (media, public discourse) rather than injecting foreign errors.
The Risk of Amplification: The danger lies not in AI introducing new errors, but in providing authoritative confirmation of existing human biases. When a human forecaster consults an LLM that shares their bias, the "corrective mechanism" of encountering a disagreeing perspective is lost. This creates a "bias lock" structurally analogous to, but more potent than, social media echo chambers due to the perceived objectivity of AI.
Current State: On elite platforms like Metaculus, human forecasters remain more accurate and independent than the AI models. However, as AI tools become integrated into workflows for less expert populations, the risk of large-scale correlated failure increases.

The paper emphasizes that while the transmission of bias to elite forecasters is not yet detectable, the preconditions for systemic risk are in place: the models share the same biases, and they are increasingly treated as authoritative sources.