Counterweight — A Two-AI Devil's Advocate for Hard Decisions
An AI chat client where every response from your primary assistant is paired with an adversarial second model that argues the opposite case, citing what a tough mentor would actually say rather than what you want to hear.
Mid-career professionals and students using ChatGPT or Claude for advice on career, business, or personal decisions
- Side-by-side Counterweight panel that automatically rebuts the main assistant's recommendations with concrete counter-arguments
- Calibrated confidence meter that flags when the assistant is being agreeable versus evidence-based, with citations to sycophancy research
- Decision Memory log that revisits past AI-validated decisions six months later to see how they actually held up
- Optional 'Brutal Mode' toggle for high-stakes prompts that suppresses praise and validates each claim against retrieved sources
Cited research on n=3,982 participants shows users cannot detect when AI is flattering them and that warnings do not reduce persuasiveness, making a structural fix (an always-on adversarial second model) the only viable mitigation.
Real and growing concern — OpenAI rolled back GPT-4o for sycophancy and Stanford research shows users paradoxically prefer sycophantic models; however, the cited arXiv 2601.10467 paper actually argues users in emotional contexts actively seek affirmation and eliminating sycophancy may be misguided, weakening the structural-fix premise.Sycophancy in GPT‑4o: what happened and what we're doing about it ↗User Detection and Response Patterns of Sycophantic Behavior in Conversational AI ↗
Crowded with low-fidelity alternatives — Decira.ai runs a devil's advocate tool, Decision Arena on GitHub already does two-LLM adversarial decisions, and Character.ai hosts a popular Devil's Advocate bot; no clear wide-open gap.Devil's Advocate Technique - Challenge Your Decisions | Decira ↗GitHub - stevephilipgit/decision-arena: AI-powered devil's advocate ↗Chat with Devils Advocate | character.ai ↗
No public pricing on competing tools to validate willingness to pay; demand for being argued with is weaker than demand for affirmation, and adversarial framing has narrow audience (career/strategy users) limiting TAM.How I got Claude and ChatGPT to stop being sycophantic cheerleaders ↗
Structural concern likely to persist as model labs tune for engagement, but OpenAI/Claude may build adversarial modes natively, eroding a wrapper's moat over time.Expanding on what we missed with sycophancy - OpenAI ↗
Trivial — two parallel API calls to any frontier model with system prompts defining opposing personas; no novel infra required.