[Research Request] “Prestige Tax”: Quantifying Epistemic Silence in AI Safety Teams

Summary

Empirical research in organizational psychology suggests that hierarchical rank is the strongest predictor of whether employees speak up or self-censor. In the high-stakes, high-prestige context of AI safety, this "silence" likely functions as a significant information bottleneck. I am conducting independent research to quantify this effect within the safety community.

If you are a junior researcher, intern, or have <5 years of experience at a safety org, please consider taking this 4 minute survey.

The Problem: Epistemic Bottlenecks

The expected value of an AI safety organization is a function of its internal information flow. However, truth-seeking is often at odds with social signaling. In labs with steep prestige hierarchies—where a junior researcher may be sitting across from someone whose seminal papers they studied in University—the social cost of "looking naive" or "being wrong" can lead to significant signal loss.

The literature suggests this isn't just a minor friction; it’s a systemic failure mode:

• The 85% Gap: A foundational study (Milliken, Morrison & Hewlin, 2003) found that 85% of employees withheld important information from supervisors due to fear of social consequences.
• The Decay of Candor: Longitudinal data (Edmondson et al., 2024) indicates that new hires enter with high willingness to speak up but "learn" silence within their first 12 months as they calibrate to the organizational hierarchy.
• Non-Routine Work: The costs of this silence are highest in non-routine research environments—exactly the conditions of alignment and evaluation work.

Why this matters for AI Safety

In this field, the counterfactual impact of a missed clarifying question, technical criticism or a "crazy" but correct alignment theory is disproportionately high. If junior staff feel they must "filter" their thoughts to match the Overton window of their leads, the organization loses the "long tail" of creative insights.

Managers often talk about "speak-up culture," but the empirical evidence suggests that culture-messaging (e.g., "our doors are always open") is largely ineffective at overriding the structural reality of prestige. 

The Research Goal

I am currently investigating whether these findings from broader organizational psychology hold true within the AI safety community. Specifically, I want to identify:

1. What is being withheld? (e.g., technical disagreements, "basic" questions, or organizational critiques).
2. Why is it being withheld? (e.g., reputational risk, fear of wasting senior time, or lack of confidence).
3. The Magnitude: Is this a marginal issue or a primary bottleneck for lab productivity?

Call to Action: 4-Minute Survey

To ground this research in actual data rather than theory, I need input from the front lines.

If you currently work at an AI safety/eval organization (e.g., METR, ARC, Apollo, Redwood, CAIS, etc.) in a junior or mid-level capacity, please fill out this survey:

[4 Minute Survey]

• Privacy: Responses are completely anonymous. No individual data will be shared with your organization.
• Output: I intend to share the aggregated, anonymized findings back with the Forum to help labs optimize their internal communication protocols.

The Institute for Classical Dialogue (ICD) is an independent organization exploring protocol-based interventions to improve epistemic quality in high-stakes research teams. Feel free to reach out in the comments or via PM.