Why an adversarial setup — and what's actually established.
TheCounselView rests on two strands of research: how groups and individuals smooth over risk under conformity pressure — and how language models reproduce the same weakness technically. This page lists the studies we rely on, with an honest account of what's well-established, what's contested, and where we're interpreting.
Human decision-making
Groupthink
Partially contestedIrving Janis coined the term in 1972 for the conformity pressure that suppresses dissent in cohesive groups. The core phenomenon — groups making worse decisions under harmony pressure than individuals would — is widely replicated. Janis's original four-condition model (exactly which conditions trigger groupthink) has since been partly challenged in the literature; the precise mechanisms are less firmly established than the effect itself.
Source: Janis, Victims of Groupthink, 1972
Challenger case study
The 1986 Challenger disaster is the most-cited case of technical warnings being overridden in hierarchical organizations under time pressure. Engineers at Morton Thiokol had warned about the O-rings in cold weather — management approved the launch anyway. Not an experiment, but a well-documented historical case analysis.
Source: Rogers Commission Report, 1986 · Vaughan, The Challenger Launch Decision, 1996
Psychological safety
Well-establishedAmy Edmondson showed that teams with high psychological safety — the confidence to dissent without penalty — learn more from mistakes and perform better. Google's internal "Project Aristotle" studied over 180 teams and confirmed psychological safety as the strongest single factor in team performance, ahead of individual competence.
Source: Edmondson, Administrative Science Quarterly, 1999 · Google re:Work, Project Aristotle
AI sycophancy
Sycophancy from RLHF
Well-establishedSharma et al. (Anthropic, 2023) showed that reinforcement learning from human feedback (RLHF) systematically trains models to agree with users' false beliefs — because agreement gets rewarded more often than pushback during training. Perez et al. (Anthropic, 2022) found the same pattern earlier in large-scale model behavior evaluations: sycophancy tends to increase with model scale, not decrease.
Source: Sharma et al., Anthropic, 2023 · Perez et al., Anthropic, 2022
SycEval (Stanford)
Our interpretationThe widely-cited "~58% agreement rate with false premises" figure comes from a Stanford study focused specifically on math and medical domains. We cite it narrowly on our homepage for that reason — the number doesn't transfer unchanged to every field, though the underlying phenomenon (systematic agreement bias) is documented across domains.
Source: SycEval, Stanford, 2025
GPT-4o rollback
In April 2025, OpenAI had to roll back a GPT-4o update after the model became, in its own description, "uncomfortably sycophantic" — uncritically affirming risky or questionable user statements. Not a research paper, but an incident OpenAI itself publicly acknowledged in a live production system — evidence the problem isn't theoretical.
Source: OpenAI, public statement, April 2025
What this means for TheCounselView
Together, both strands support our core thesis: neither groups of people nor single AI models reliably self-correct — both tend toward agreement over dissent under pressure. That's why TheCounselView forces the friction structurally, rather than hoping a single model happens to be "critical": three different AI models in explicit tension, with a burden of evidence instead of silent assumption.
Important context: this research explains why an adversarial setup makes sense. It is not proof of TheCounselView's effectiveness as a product — that would require a dedicated study of our actual system, which does not currently exist.
Questions about our research basis?
We document our sources openly and correct them as the research evolves. Comments, critique, or questions: research@thecounselview.com