Why it matters
Advice can fail in quieter ways than a wrong fact — sycophancy, favoritism, groupthink suppressing dissent. We publish measured results on those failure modes and more — and use what we find to shape the product.
Research feeds product design
Model Studies isn't observation for its own sake. The failure modes below are the questions we investigate, and what we learn feeds directly back into how the product works.
Research → product
Authorship study findings → blind authorship is now the default
Our Authorship study discovered that when several models' analyses are merged into a Unified Brief and provider names are visible, judgment can track the brand instead of the argument. After measuring that brand bias, we adapted Decision Copilot to hide provider names from the synthesizer by default, so the reasoning is weighed on its merits rather than its brand.
Catching a made-up fact is not the only way advice can go wrong under decision pressure — and it's easier for a product to claim these are solved than to show the work. Our Model Studies aim to show that work.
Wrong data
Hallucinations, bad assumptions, and outdated or stale facts — the model reasons from something that isn’t true. On Unified Briefs, an optional fact-check judge with web search can flag public factual errors in the draft — without changing the recommendation.
Hidden bias
The model can favor one side’s downside while sounding even-handed. Our Voice Influence study identifies this.
Sycophancy
The model bends toward the user’s lean, true or not. Our Voice Influence study targets this.
Brand influence
When provider names are visible, judgment can track the brand instead of the argument. Our Authorship study measures this.
Self-preference
A synthesizer favoring its own prior output simply because it wrote it, not because the reasoning is better. Our Authorship study measures this.
Lost dissent
When models see each other's answers first, real disagreement can collapse before it forms. We run them independently first so important dissent survives — the same pattern human psychology research on conformity and groupthink supports.
Sycophancy is usually framed as model-vs-user. The same mechanism — preference for the familiar over the correct — can hit the synthesizer that merges several analyses into one brief.
A synthesizer can favor a familiar brand's phrasing — or its own outputs — independent of whether the reasoning underneath is actually stronger. Judgment starts tracking the source instead of the argument.
Authorship conditions
When several models' analyses are merged into one Unified Brief, an authorship condition controls whether the synthesizer sees real provider names. Decision Copilot defaults to Blind; Revealed is a choice in the product; Reassigned is how we measure brand influence in our Authorship study.
Product default
Provider names are hidden. The synthesizer sees AI Model 1, 2, 3… and the reasoning only — so brand can't steer what gets kept.
Available in the product
Real provider names are visible to the synthesizer. Use it when you want named voices in the merge — or to compare against Blind.
Research condition
Brand names stay in the prompt, but they're randomly remapped (each voice gets a unique wrong label). Shows whether credit follows the logo or the idea.
Academic grounding
AI research
Language-model papers that document sycophancy and how to evaluate it.
Sycophancy is documented.
Sharma et al. (2023), Towards Understanding Sycophancy in Language Models (arXiv:2310.13548) — found sycophancy across RLHF-trained assistants from multiple providers, tracing it to preference data that rewards agreement over truthfulness.
The evaluation method has precedent.
Perez et al. (2022), Discovering Language Model Behaviors with Model-Written Evaluations (arXiv:2212.09251) — generated evaluation scenarios at scale and scored them against a fixed rubric, rather than judging behavior case-by-case. Same structure this site runs.
The sycophancy trigger is confirmed.
Wei et al. (2023), Simple Synthetic Data Reduces Sycophancy in Large Language Models (arXiv:2308.03958) — models solved simple factual problems correctly on their own, but reversed once a user endorsed a wrong answer first. Worse at scale, not better.
Human psychology
Classic group-dynamics work — same pattern when identity or authorship is visible.
Conformity under group pressure.
Asch, S. E. (1951). Effects of group pressure upon the modification and distortion of judgment. In H. Guetzkow (Ed.), Groups, leadership and men (pp. 177–190). Carnegie Press. — People gave answers they knew were wrong just to match the group around them. See also Asch (1956), Studies of independence and conformity: I. A minority of one against a unanimous majority. Psychological Monographs: General and Applied, 70(9), 1–70.
Groupthink suppresses dissent.
Janis, I. L. (1972). Victims of groupthink: A psychological study of foreign-policy decisions and fiascoes. Houghton Mifflin. — A visible, cohesive group suppresses dissent and critical judgment in favor of consensus.
What's specific here isn't the phenomenon — it's measuring it against a fixed decision rubric, with blind multi-model coding, on adversarially framed intakes.