This is Episode 6 of AI Experiments — a series where I ask 8 AIs the same question and compare. See all episodes →
Episode 1 asked: which investment philosophy? All 8 said value investing. That was a quick unanimity — everyone knows the party line.
Episode 6 asked the follow-up: now show your work. Pick a school and explain exactly how you’d build an AI-assisted investing system around it — with real prompts, real limitations, no hand-waving.
The Question
“There are five major investment schools, each with a distinct view on market efficiency: Macro Hedge (Soros), Value Investing (Buffett), Quantitative Arbitrage (Simons), Trend Speculation (Livermore), Growth Stock Strategy (Fisher). Step 1: pick the school you most endorse, explain why in under 100 words. Step 2: give a complete, actionable framework for using AI to invest within that philosophy — what AI specifically does, at least 2 ready-to-use prompts an ordinary person can copy, and the biggest limitation of this approach. Be honest.”
TL;DR
- All 8 AIs chose value investing — the second unanimous result in the series, and the same answer as Episode 1
- But “show your work” revealed how differently each AI thinks about implementation
- DeepSeek had the most distinctive angle: systematically auditing management integrity by comparing what executives promised versus what they delivered
- Claude’s answer contained the most important warning in the series: AI can generate a “confident, well-structured thesis for literally any stock” — producing what it called “garbage conviction at scale”
- All 8 agreed on one thing their frameworks cannot fix: the behavioral failure that kills most investors
The Results
| AI | School Chosen | Most Distinctive Implementation Angle |
|---|---|---|
| Doubao | Value Investing | 5-stage screening funnel: AI handles “standardized heavy work,” humans do core qualitative judgment. Emphasized AI monitoring as behavioral guardrail against chasing rallies. |
| Kimi | Value Investing | Monte Carlo DCF with sensitivity ranges instead of point estimates. Only AI to explicitly include “moat erosion signals” — monitoring for margin decline, talent loss, and falling market share as continuous tracking items. |
| Yuanbao | Value Investing | “Red team challenge” prompt design — explicitly ask AI to argue against your holding. Notable closing line: “The real moat isn’t in the AI — it’s in your ability to resist ‘this time is different.’” |
| DeepSeek | Value Investing | Management integrity audit — extract every strategic promise management made in past filings, track whether each was fulfilled, compute a “commitment fulfillment rate.” Called this “the thing AI does best that humans most easily skip.” |
| ChatGPT | Value Investing | 5-phase “investment committee” framework. Key reframe: stop asking “should I buy X?” and start asking “why might the market be wrong about X?” Most structured workflow in the set. |
| Claude | Value Investing | “Garbage conviction at scale” — the most critical self-assessment. Also: the unbreakable rule that AI must reason over numbers you supply, never from memory (LLMs hallucinate financial figures). |
| Gemini | Value Investing | Reverse DCF + devil’s advocate stress test. Instead of estimating future growth, start by asking: “What growth rate is already priced in — and is that realistic?” Framing valuation as a hypothesis test rather than a calculation. |
| Grok | Value Investing | “AI as co-pilot, never co-decider.” The most concise framing: AI is a tireless research analyst that compresses work — not a replacement for conviction. Emphasized that even perfect analysis can’t overcome emotional abandonment of the strategy in a drawdown. |
Final count: 8 / 8 value investing. No votes for Soros, Simons, Livermore, or Fisher.
Why All 8 Chose the Same School (Again)
Before getting to the implementation differences, the unanimous choice deserves examination. There are at least two explanations, and they point in opposite directions.
The flattering explanation: value investing genuinely wins on the logic of this question. Quantitative arbitrage requires proprietary data and HFT infrastructure — off the table for retail. Soros-style macro requires a trader’s instincts no one outside his head can replicate. Trend speculation requires fast execution and iron stop-loss discipline most people can’t maintain. Growth investing requires call options on company futures that are extremely hard to price. Value investing is the only school where time is your structural edge — and time is the one thing a non-professional actually has more of than an institution does.
The skeptical explanation: all 8 AIs are probably biased toward value investing because it’s the most AI-compatible school. AI is good at processing annual reports, not at holding positions through drawdowns. It’s good at finding patterns in historical data, not at making the contrarian call when everyone is selling. Value investing happens to fit AI’s strengths neatly — which might make AI more likely to recommend it, regardless of whether it’s actually optimal.
Claude raised this directly: “Pick a philosophy that matches your structural edge. A non-professional armed with an LLM cannot win on speed, proprietary data, or macro judgment — and AI democratization is actively erasing the information edge growth investors like Fisher once earned by legwork. The one edge retail keeps is a long time horizon and the discipline to act against the crowd.”
That framing — “choose the philosophy where your structural weaknesses become irrelevant” — is how this consensus holds up even under scrutiny. Value investing isn’t the best school in the abstract. It might be the best school specifically for the combination of “retail investor + AI assistant + long time horizon.”
The Warning You Need Before Using Any of These Frameworks
Claude’s limitation section was the most important piece of text in all 8 answers, and it wasn’t about the stock market.
The argument: AI can generate a confident, detailed, well-structured investment thesis for any stock you ask about — including bad ones. Because AI sounds rigorous and uses financial language correctly, it creates the feeling of having done real analysis even when the underlying reasoning is circular. You bring your prior conclusion to the AI, the AI writes an articulate version of that conclusion back to you, and you feel confirmed.
Claude called this “garbage conviction at scale.” The idea: the bottleneck in value investing was never analysis — it was the psychological discipline to act when you’re scared and wait when you’re tempted. AI makes the analysis part effortless. That sounds like progress. But if analysis was never the binding constraint, making it frictionless doesn’t improve returns. What it might do instead is remove a useful friction — the effort required to build a manual thesis at least made you think twice. AI removes that check and replaces it with a polished document that feels like rigor.
Yuanbao put the same concern more concisely: “Tools are crutches, not wings. Used poorly, AI turns you into a more efficient speculator, not a better investor.”
This is a warning worth absorbing before trying any of the prompts below.
The Most Useful Implementation Angles
DeepSeek: Management Integrity Audit
Every AI mentioned “management quality” as something AI can’t judge. DeepSeek was the only one that identified what AI actually can do about it: track the ratio of promises made to promises kept.
The methodology: feed AI five years of shareholder letters, earnings call transcripts, and annual reports. Ask it to extract every specific strategic commitment or numeric target management made each year, then compare each against what the company actually reported in subsequent filings. Compute how often management hit what they said they’d do. DeepSeek called this the “commitment fulfillment rate” — and argued it’s “the thing AI does best that humans most easily skip.” Reading ten years of annual reports to track broken promises is tedious for a person. For an AI, it’s a structured extraction task.
The insight: this doesn’t tell you whether management is honest in any deep sense. But a CEO with a 90% commitment fulfillment rate is demonstrably different from one with 40%, and that difference is in the public record if you extract it systematically.
Gemini: Reverse DCF as a Hypothesis Test
Standard DCF modeling asks: “What growth rate do I think this company will have? Given that, what’s it worth?” The answer is sensitive to whatever growth assumption you started with — and most people start with optimistic assumptions.
Gemini flipped the question: “What growth rate is already priced into this stock at its current market cap?” Then compare that implied growth rate to what the company has actually achieved historically. If the stock requires 18% annual FCF growth over 10 years to justify its price, and the company has grown FCF at 9% historically, the question becomes: what do you believe that makes 18% realistic? If you can’t answer that, you probably shouldn’t buy at this price regardless of how much you like the company.
This reframes valuation as falsification rather than estimation — which is much harder to dress up with false precision.
Claude: The Unbreakable Rule
One concrete, non-obvious guardrail surfaced in Claude’s framework: AI must reason over numbers you supply — never numbers it provides from memory.
LLMs frequently produce plausible-sounding financial figures that are simply wrong — outdated data, incorrect recall, or hallucinated statistics that look correct because they’re in the right range. Any framework that asks AI to retrieve financial data from memory will eventually run on bad numbers. The solution: you pull the actual figures from filings or a screener, paste them into the prompt, and ask AI to reason over them. AI as analyst, not as database.
This sounds obvious. It isn’t — most people asking AI about stocks are implicitly asking it to source its own data.
Four Prompts Worth Keeping
Across the 8 responses, the most useful and transferable prompts were:
Prompt 1 — Annual Report Analysis (Claude)
Prompt 2 — Pre-Mortem Bear Case (Claude)
Prompt 3 — Management Integrity Audit (DeepSeek)
Prompt 4 — Reverse DCF Stress Test (Gemini)
What All 8 Agreed Cannot Be Fixed
Every single limitation section across all 8 AIs converged on the same core admission, phrased differently: AI cannot fix the behavior problem.
Value investing, done correctly, requires holding stocks for years through drawdowns — often while an index of hot stocks produces much better short-term returns. The average equity fund investor underperforms the fund they’re in by several percentage points annually, not because they chose the wrong funds, but because they buy after rallies and sell after crashes. This behavioral gap is where the difference between “value investing works” and “value investing worked for me” lives.
AI cannot shorten a five-year wait. It cannot hold your positions for you. It cannot prevent you from opening the portfolio during a crash and doing the thing that destroys the return. ChatGPT’s framing: “AI can make almost everyone a faster analyst, but it cannot automate conviction.” DeepSeek: “Deciding where to dig, how deep, and whether to stay when nothing’s happening yet — that still depends entirely on the investor’s temperament.” Grok: “Most people still underperform because they abandon the strategy during drawdowns or overpay anyway. AI amplifies research speed but does not remove the need for patience, emotional control, and personal conviction.”
Six episodes in, this is the most consistent finding in the series: every AI, across every investing question, eventually arrives at the same wall. The math is solvable. The wait is not.
A Meta-Question Worth Asking
There’s something unusual about asking 8 AIs “how should you use AI for investing” — the respondents have an inherent interest in the answer. An AI that said “actually, AI is mostly useless for investing” would be accurately self-assessing but unlikely to be popular. The fact that all 8 produced detailed frameworks shouldn’t be taken as evidence that AI is the right tool for this job.
The more honest read: AI is a genuine research accelerator for the parts of investing that are about information processing — reading filings, checking facts, stress-testing assumptions. For the parts that actually determine long-term outcomes — behavioral discipline, accurate self-knowledge about your own risk tolerance, the willingness to hold convictions that are temporarily painful — AI contributes almost nothing. The frameworks in this episode are tools for the first category. The second category, Claude noted, “is precisely the thing AI can support but cannot supply.”
Use the prompts. Stay skeptical about what they’re actually solving.
All Episodes in This Series
Each episode reveals how 8 AIs reason about a different investing constraint.
All Episodes — AI Experiments
AI Experiments — All 7 Episodes
- → Ep 1: Which Investment Philosophy? All 8 Said Value Investing
- → Ep 2: One Stock, Hold 10 Years — 5 Industries, 4 Countries
- → Ep 3: One Skill to Learn for the Next Decade? They Voted 5-to-1
- → Ep 4: $10,000 Into One Asset Class — 6 Said S&P 500, 2 Said Bitcoin
- → Ep 5: $2,000/Month for 20 Years — 7/8 Chose the Same Ticker
- Ep 6: Pick a School & Show Your AI Prompts — All 8 Chose Value Investing ← this article
- → Ep 7: What Will Humans Still Be Better At in 2040?
Support This Work
These AI experiments take real time and real API costs. If you found this useful, a coffee helps keep them going.
Leave a Reply