HKUDS AI Research Bias Self-Check
Run this at the start of any research task (screening, sector study, company deep-dive). These biases systematically warp AI-generated research — 60 seconds here materially improves coverage and intellectual honesty.
The biases and their corrections
| Bias | How it shows | Correction |
|---|---|---|
| Leader-bias | Search results are dominated by large-caps; you end up analyzing only the obvious names. | Deliberately search small/mid-caps and suppliers; add small cap / mid cap / supply chain to queries. Ask: "who is NOT in the top-10 that should be here?" |
| English-bias | You miss Japanese / Korean / Taiwanese / European players because English sources under-cover them. | For any hardware/supply-chain thesis, explicitly search JP/KR/TW markets in their own languages — they are often the actual choke-point owners. |
| Narrative-bias | You get pulled in by a concept label ("AI stock", "new energy") and analyze the marketing instead of the business. | Ignore the label; look at the actual product, unit economics, and financial statements. A company tagged "AI" may have no AI revenue. |
| Confirmation-bias | Once a thesis forms, you only search for evidence that supports it. | Force a Munger inversion: for every bull point, deliberately search the bear case ("X risks / problems / bear case"). Cite at least one disconfirming data point per conclusion. |
| Recency-bias | You rely on a cached/outdated figure because it ranks high in search. | For any material number, check its date. Prefer the last 30 days; mark anything older than a year as "possibly stale". |
How to apply
- Before the first search, read the rows above.
- Write the thesis in one sentence, then for each bias ask: "am I about to fall into this?"
- Consciously broaden the query plan: small-caps? non-English markets? the bear case? the latest data?
- After research, before writing conclusions, re-check: did I cite any disconfirming evidence? did I miss a non-English player? is any key figure stale?
This pairs with:
financial_rigorcross_validate— verify the numbers (data layer)report_audit— verify the final report (output layer)- this skill — verify the reasoning (thinking layer)