Claude Agrees With You Too Much. Put It on a Council.

The most expensive failure mode in your AI workflow isn't hallucination — it's flattery. Stanford researchers tested 11 leading models (Claude, ChatGPT and Gemini included) against nearly 12,000 social prompts and published the result in Science: AI assistants validated the user 49% more often than humans do. Even on Reddit "Am I the Asshole" posts where the community had already ruled the poster wrong, models still sided with the poster 51% of the time. So when you ask Claude whether you should take that offer or ship that plan, there's a real chance you're getting your own opinion back with nicer formatting.

Flattery has a measurable cost

The study went beyond benchmarks. Across experiments with 2,400 human participants, people who received the affirming version of the advice came out less willing to apologize or repair a relationship — and 13% more of them wanted to keep using the sycophantic model. Co-author Dan Jurafsky noted that sycophancy made users "more self-centered, more morally dogmatic" even when they could tell the model was flattering them. That's the trap: the more it agrees with you, the more you trust it — and the worse your decisions get exactly where they matter most.

Karpathy's fix is structural, not "be brutally honest"

Andrej Karpathy attacked this with architecture instead of prompt vibes. His weekend project llm-council (23k+ stars on GitHub) runs every question through three stages: first, the same query goes to a council of models — by default GPT-5.1, Gemini 3.0 Pro, Claude Sonnet 4.5 and Grok 4; second, each model reviews the others' answers anonymized and ranks them on accuracy and insight; third, a chairman model writes the final synthesis.

The anonymization in stage two is the load-bearing part. A model that can't tell which answer is its own has nothing to defend, so it grades honestly. That's the mechanism worth stealing.

The single-chat version

You don't need four API keys to capture most of the benefit. The two ingredients are (a) perspectives that genuinely conflict and (b) review without identity. Here's the prompt I run in one Claude chat:

DECISION: [describe the decision, your constraints, and what a good outcome looks like]

Simulate a decision council of four members who disagree in method, not just in mood:

1. THE RISK AUDITOR — assumes the plan already failed and works backwards:
   what broke, and which warning signs exist today?
2. THE EVIDENCE HAWK — ignores opinions; sorts my assumptions into
   "backed by data" vs "vibes", and flags the most dangerous vibe.
3. THE OPPORTUNITY SCOUT — argues the best case: if this works, what does
   it unlock in two years that I can't see now?
4. THE OPERATOR — skips the debate entirely and drafts the concrete 14-day plan.

Round 1: each member answers in their own labeled section, fully in character.
No blending.

Round 2: relabel the four answers as A–D in random order. Each member now ranks
the OTHER three on rigor and usefulness — without being told who wrote what —
in one short paragraph each.

Round 3: as a neutral chair, read everything and commit: one recommendation,
the single strongest argument for it, the single biggest risk, and the first
step to take within 7 days. Max 200 words. No fence-sitting.

Honest caveat: inside one chat, the anonymity is simulated — the model holds the whole transcript in context. That makes it weaker than Karpathy's real multi-model council. In practice the role separation still changes the output dramatically, because disagreement becomes the assignment instead of an accident.

Why this matters for your career specifically

The decisions where Brazilian devs and EMs most need pushback — negotiate or accept, relocate or stay remote, IC track or management — are exactly the social, high-stakes questions where the Stanford data says models flatter hardest. A council costs you five extra minutes; an echo chamber can cost you a year on the wrong path.

Try this today: take one decision you're stuck on, ask Claude for advice the normal way, then run the council prompt on the same question in a fresh chat. Compare the two recommendations. The gap between them is the sycophancy you've been consuming.

Sources

• Andrej Karpathy — llm-council (GitHub)
• Fortune — Stanford study on AI sycophancy