If you've ever asked ChatGPT or Claude "is Apple a good Buffett pick" or "what would Warren Buffett think of Tesla," you've probably gotten a confident-sounding paragraph with phrases like Buffett would say... or Munger has repeatedly emphasised... — and you have no way to verify any of it. The quote sounds right. The framing sounds right. The actual citation? There isn't one. The LLM made it up.
This post explains the technical difference between generic AI stock-analysis (a chatbot asked to roleplay a famous investor) and RAG-grounded AI stock-analysis (a system that can only quote from an indexed corpus of real source material). The difference is materially important for investing — and it's the core engineering decision behind invest-like's Ask Buffett, Buffett Brain, and Boardroom features.
The hallucination problem in plain English
Modern LLMs are trained to produce fluent, plausible-sounding text. When you ask "what did Buffett say about Apple," the model produces text that sounds like a Buffett quote. It doesn't know whether the quote is real or fabricated; it doesn't have an internal flag for "I made this up." It just generates the most-likely-next-word, conditioned on the patterns in its training data.
For most tasks, this is fine. For investing, it's a serious problem. Two specific failure modes:
Failure 1: Fabricated quotes attributed to real investors
Ask ChatGPT: "What did Warren Buffett say about NVIDIA in his 2024 letter?"
You will get a confident-sounding answer with a "quote" that may or may not appear in the actual letter. The model doesn't know. It's pattern-matching on what Buffett-style language sounds like, combined with what an NVIDIA bull or bear case sounds like. The result is an articulate sentence that no real person wrote.
For an investor, this is catastrophic. You're being given fake evidence to support a real decision.
Failure 2: Confident reasoning about ratios it doesn't have
Ask ChatGPT: "What's Apple's current ROIC?"
The model gives you a number. It might be the right number, the wrong number from last quarter, or made-up. ChatGPT has no realtime data access in its base mode; it can only generate text based on what was in its training corpus. The ROIC you get may be from 2023 — and you have no way to know.
What RAG actually does, technically
RAG = Retrieval-Augmented Generation. The architecture has three stages:
-
Index: We pre-process the source material (Berkshire shareholder letters from 1977 to 2025, plus Munger's published talks, plus our own scoring methodology) into chunks. Each chunk gets an embedding vector. The vectors live in a database.
-
Retrieve: When a user asks "what did Buffett say about Apple in the 1996 letter," we embed the question, find the chunks in the corpus with the most similar embedding, and pull those chunks.
-
Generate: We feed the retrieved chunks to the LLM as context, with strict instructions to only quote from the provided context and cite the source. The LLM still generates fluent text, but it's now constrained.
The user-facing result: every Ask Buffett answer cites a specific letter year and section. If the user clicks through, the source passage is highlighted. If a question can't be answered from the corpus, the system says so instead of inventing.
Same architecture for the Boardroom: each investor persona (Buffett, Graham, Lynch, Greenblatt) can only retrieve from its own corpus. Buffett can quote Buffett's letters; Graham can quote The Intelligent Investor; Lynch can quote One Up on Wall Street. Cross-attribution is blocked by design.
Side-by-side example: "What did Buffett say about premium prices for quality?"
Generic ChatGPT (with no tools / RAG)
"Warren Buffett has frequently emphasised that paying a fair price for a wonderful business is far better than paying a wonderful price for a fair business. He has noted that quality compounds and that the cost of patience is often outweighed by the cost of compromise on the underlying business. In his 1985 letter, he reportedly stated..."
(Made up, sounds plausible, you have no way to verify which sentences are real and which are fabricated.)
Ask Buffett on invest-like
Based on the 1989 Berkshire Hathaway Shareholder Letter:
"It's far better to buy a wonderful company at a fair price than a fair company at a wonderful price."
Source: 1989 letter, section on Salomon Inc.
Additionally, the 1996 letter expands on this: [retrieves specific chunk from 1996 letter with timestamp metadata].
[The user can click any citation to see the source passage.]
The Ask Buffett version is grounded, falsifiable, and verifiable. If the LLM gets the citation wrong, the source PDF is one click away and the user catches the error. If the LLM gets the quote wrong, the source passage is highlighted and the discrepancy is visible.
This is the architectural difference: not "smarter AI", but "AI that can only cite the real source."
The Buffett Brain layer on top of RAG
Ask Buffett is one layer. Buffett Brain is a second layer that combines the RAG-grounded letter quotes with our deterministic 7-framework scoring engine.
When you open a Buffett Brain verdict on a stock, the system:
- Computes the seven framework scores deterministically from FMP fundamentals data (this part is pure math, no AI)
- Pulls the most-relevant Buffett letter quotes for the specific pillar findings (RAG)
- Asks the LLM to synthesise the deterministic scores + the cited quotes into a single A-to-F grade with reasoning
The synthesis is the only AI-generated step. The underlying numbers are deterministic. The citations are real. The user can audit every link in the chain.
The Boardroom: multi-investor RAG with skeptic
The Boardroom feature extends the same architecture to four investor personas at once. Each persona has its own corpus, retrieves from its own corpus only, and cites accordingly. A separate "skeptic" persona has access to a different corpus (focused on counter-arguments, risk research, and known framework limitations) and challenges every bullish claim.
The technical insight: the skeptic isn't just "negative ChatGPT". It's a separately-trained constraint that retrieves from a deliberately-different corpus to surface counter-arguments. This is the part that no generic LLM can do — generic LLMs are trained to be helpful, not to argue against themselves.
Why this matters for investing
Most decisions in life don't require citation. If ChatGPT helps you write a birthday card, the worst-case is awkward phrasing.
Investing decisions are different. You're allocating real money based on the analysis. If the AI you're consulting is hallucinating citations, you're making decisions on evidence that doesn't exist. Over time, this compounds badly.
That's why invest-like was built differently from day one. Every claim attributable to a famous investor cites a specific source. Every fundamental number comes from a licensed data feed (Financial Modeling Prep) and is timestamped. Every Buffett Brain verdict can be traced back to deterministic framework rules.
This is the engineering difference between "AI-flavoured fluff" and "AI-grounded analysis." For investing, it matters.
How to verify this yourself
Open Ask Buffett on any stock. Ask a question. Click the citation. The source passage is highlighted in the letter. If the quote doesn't match the source, the system has a bug — email the founder and it will be fixed.
Compare the same question to ChatGPT's answer. Look at whether ChatGPT cites a source. Look at whether the citation is verifiable. Decide for yourself.
The honest read: ChatGPT's value-investing content is fine for casual reading. For decisions involving real money, you want grounding.
Disclosure
Educational tool. invest-like uses Anthropic's Claude and OpenAI's GPT-4 family models for generation, with RAG architecture grounding all citations against actual indexed source corpora (Berkshire shareholder letters 1977-2025, plus the cited books of Graham, Fisher, Lynch, Greenblatt, Munger, and Smith). The licensing of these source materials is documented at our methodology page and reviewed against German § 51 UrhG Zitatrecht standards (we are based in Kiel, Germany).
The author is Zaid Ghazal, indie founder. Quotations from Warren Buffett are © Berkshire Hathaway Inc. and used for educational commentary under fair-use / Zitatrecht doctrine.