How AI Berkshire Handles AI Research Bias: A Two-Layer Framework

AI Berkshire handles AI research bias through a systematic two-layer framework that combines Information-Richness Ratings (A/B/C) with a mandatory Bias-Awareness Checklist embedded in every research workflow.

The xbtlin/ai-berkshire repository implements a rigorous bias-awareness protocol designed to counteract systematic distortions in AI-generated investment research. This approach transforms abstract concerns about AI research bias into concrete, auditable actions by mandating explicit workflow artifacts before any analysis begins.

The Information-Richness Rating System (A/B/C)

Before generating research outputs, the system classifies target companies according to information availability and consensus strength. This rating is defined in README_EN.md and directly drives the intensity of subsequent bias checks.

The three-tier classification works as follows:

• A-class: Information-rich environments with extensive analyst coverage and high consensus risk
• B-class: Moderate coverage with mixed opinions and medium information density
• C-class: Information-scarce targets, typically private or early-stage firms with limited public data

The Bias-Awareness Checklist: AI研究偏见自觉

For A-class or consensus-heavy targets, the workflow automatically injects a mandatory checklist titled "AI 研究偏见自觉" (AI research bias awareness). This checklist is defined in codex-skills/investment-memo-craft/SKILL.md and captures four specific bias categories that analysts must mitigate:

Bias Type	Description	Mitigation Step
Large-cap bias	Over-weighting well-known corporate giants	Explicitly compare against smaller peers
English-language bias	Preference for English-language sources	Include Chinese-language reports and local filings
Narrative bias	Following prevailing market storylines	Document contrary data points and counter-arguments
Listed-only bias	Ignoring private-market opportunities	Add "future IPO candidates" list

How the Bias Framework Works in Practice

When the /investment-research skill is invoked, the first step of the generated report is the AI研究偏见自觉 section, as implemented in skills/investment-research.md. The system records the information-richness rating and populates the checklist with concrete mitigation items.

During subsequent research stages—including moat analysis, management evaluation, and valuation—the analyst must reference checklist items to ensure identified biases are revisited and either confirmed or disproved.

Example workflow invocation:

# Invoke the bias-aware research skill

/prompts:investment-research NVIDIA

## AI研究偏见自觉

- 信息丰富度评级：A  
- 共识陷阱：高 – 多数分析师持“长期增长”乐观观点  
- 偏见检查清单  
  - 大盘偏见 ✔ 已对比同业中小盘公司  
  - 英文语言偏见 ✔ 包含中文晨星报告与本地监管文件  
  - 叙事偏见 ✔ 记录与“AI 计算芯片”热潮相悖的产能利用率数据  
  - 上市仅偏见 ✔ 添加未上市的芯片材料供应商作为潜在 IPO 目标

Industry-Funnel Integration

The bias-awareness protocol extends beyond company-specific research into sector analysis. The skills/industry-funnel.md file repeats the bias-awareness step, explicitly calling out the same four bias categories in its "AI bias awareness" bullet point.

This ensures that industry-level deliverables—such as reports/AI算力-funnel-20260509.md—maintain the same rigorous standards and document mitigation steps for large-cap, English-language, narrative, and listed-only biases.

Summary

• AI Berkshire employs a two-layer framework combining Information-Richness Ratings with a mandatory AI研究偏见自觉 checklist.
• Companies are classified as A, B, or C-class based on information density and consensus risk, determining the depth of bias scrutiny.
• The checklist targets four specific distortions: large-cap bias, English-language bias, narrative bias, and listed-only bias.
• Enforcement is embedded in codex-skills/investment-memo-craft/SKILL.md and triggered through the /investment-research and /industry-funnel skills.
• Analysts must document concrete mitigation actions and reference bias checks throughout the research lifecycle.

Frequently Asked Questions

What are the four types of bias that AI Berkshire checks for?

AI Berkshire specifically targets large-cap bias (over-weighting famous companies), English-language bias (preference for English sources), narrative bias (following popular storylines), and listed-only bias (ignoring private markets). Each type has a defined mitigation step that analysts must complete in the checklist, such as explicitly comparing against smaller peers or adding private-company pipelines.

How does the Information-Richness Rating affect the bias checklist?

The A/B/C rating determines whether the full bias checklist is triggered. A-class companies (information-rich with high consensus) automatically receive the AI研究偏见自觉 section, while lower classes may undergo lighter scrutiny. This ensures intensive bias checks are applied precisely where consensus risks and information overload are highest.

Where is the bias checklist defined in the codebase?

The bias checklist is defined in codex-skills/investment-memo-craft/SKILL.md, which automatically inserts the AI研究偏见自觉 section into every investment memo. The system also implements bias awareness in skills/industry-funnel.md and documents the theoretical framework in README_EN.md.

Can the bias awareness framework be customized for other research domains?

Yes. Because the framework is implemented through modular skill definitions in the codex-skills/ directory, researchers can adapt the AI研究偏见自觉 template and four-bias taxonomy to other domains by modifying SKILL.md files while maintaining the same systematic workflow structure and mitigation documentation requirements.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}