AI debate for investment decisions: a devil's advocate for your thesis
The dangerous moment in any investment decision is when everyone in the room already agrees. A single AI assistant makes that worse, not better — ask it whether your thesis holds and it will build the bull case fluently, quote plausible numbers and rarely volunteer the reason you are wrong.
An investment view is only as good as the strongest argument against it. LLM Debate Council forces that argument into the open: several models from rival vendors take opposite sides, a verifier checks the claims against live sources, and you get a bull case, a bear case and a reasoned verdict you can actually defend to a committee.
Force a real bull-and-bear split
State the thesis — 'this company's margins are durable because of switching costs' — and run Debate mode with the Opponent role on. One agent builds the strongest bull case; another is instructed to argue the bear case in good faith, hunting for the non-obvious risk rather than nodding along.
Because the models must respond to each other's strongest points, the weak arguments get pressure-tested in round two instead of surviving into your memo. Agreement is tracked numerically per round, so you can see whether the models genuinely converged or one simply wrote with more conviction.
- Thesis stress-test → Debate mode with the Opponent (devil's advocate) on
- Risk discovery → Red-team pass attacks the emerging conclusion from several angles
- Independent reads → Expert panel: each model takes a position, then a merged verdict
- Comparing options → Brainstorm the angles, then Synthesis mode drafts the memo
Fact-check the numbers, not just the story
Investment write-ups live and die on specific figures — growth rates, margins, market size, comparable multiples. After the debate, a verifier extracts each factual claim and checks it independently on the web, returning a claim-by-claim table marked VERIFIED, DISPUTED, FALSE or UNVERIFIABLE with the sources beside each verdict.
For anything time-sensitive, turn on web access so the models ground their claims in current filings and reporting rather than stale training data. A number that shows up DISPUTED is exactly the one to check before it reaches your allocation.
Different vendors as a bias control
Using GPT, Claude, Gemini and Grok together is a practical hedge against any one model's blind spots. Their training data, alignment choices and risk appetites differ, so a conclusion that survives all of them at once is harder to write off as one model's optimism. Council mode (Pro) adds anonymous peer ranking, where models grade each other's arguments without knowing whose they are.
A transcript your committee can audit
Every debate produces a full record: who argued the bull and bear case, which claims were verified, which were disputed, and how the view converged. Export it to PDF or DOCX and attach it to the investment memo — the decision stops being 'we had a feeling' and becomes a documented, checkable process a committee can review.
FAQ
Is this investment advice?
No. LLM Debate Council is a reasoning and research tool — it structures a rigorous bull-and-bear debate and fact-checks claims, but the decision and any regulated advice remain yours. Treat verdicts as inputs to audit, not recommendations to follow.
Can the models see current market data?
With web access on, agents search current sources and cite them; for paywalled data (broker research, filings behind a login), paste the key excerpts into your question and the models will debate the material you provide.
How do I stop the models from just agreeing with my thesis?
Turn on the Opponent role and, on Pro, the red-team pass — these explicitly assign a good-faith bear case and attack the emerging conclusion, which counteracts the sycophancy a single model shows.
Free tier · bring your own API keys · 15 languages