You can have a conversation with Thinking in Bets now

Thinking in Bets · email · 2026-09-05 Reader 🖥️

Decision-Making Practices LLM Limitations Retrieval & RAG Trust

Grounding a conversational AI's answers directly in a specific book's text, with citations to the passages they came from, converts an unverifiable claim about what the book says into a checkable one — which is what makes AI-mediated access to a specific author's ideas trustworthy, and worth extending to a reader's own material.

practical · mixed — Primarily a promotional announcement (practical: it wants you to try the feature, claim the free book, and refer friends) that embeds a small theoretical claim — that grounding an AI in a source text with citations solves a specific trust problem — as credibility-building context for the pitch.

Structure

Part	Role
Diagnosis: base-model answers about a book come from training absorption, and the reader cannot tell a faithful account from a confabulated one	problem statement / motivation
Solution: Expert Intelligence in Gemini Notebook answers from the actual book text with citations to source passages	claimed mechanism
Demonstration: a worked example question ('why probabilistic thinking helps') and the answer it produced, framed as evidence the mechanism works	anecdotal evidence
Extension idea: combining the grounded book with the user's own uploaded material (decision journal, team notes) to apply the book's frameworks to personal decisions	aspirational use case, explicitly marked as unresolved/curious rather than demonstrated
Calls to action: free book giveaway, Featured Notebook with a Decision Quiz, referral program	promotional close

Key terms

Expert Intelligence
Google Gemini Notebook's feature that grounds conversational answers about a specific added source (here, the book Thinking in Bets) in that source's actual text, returning citations to the passages an answer draws from.
resulting / decision quality vs. outcome quality
Named only implicitly, via the Decision Quiz description ('built to make you separate outcome quality from decision quality'). Refers to Duke's established distinction between whether a decision was well-reasoned given available information versus whether it happened to turn out well — never spelled out in this post, only invoked as a known feature of her existing framework.
grounded (in the text)
An answer whose content is drawn from and traceable to the actual source document, as opposed to a model's internalized, untraceable impression of that document formed during training.

Propositions

explicitWhen an AI answers a question about a specific book using only what the model absorbed during training, the answer is sometimes faithful to the book and sometimes not, and the user cannot reliably tell which.conf: high
explicitGrounding an AI's answers in the actual book text, with citations to the specific passages an answer draws from, makes the answer's fidelity independently checkable by the reader.conf: high
supportsGemini API File Search is now multimodal: build efficient, verifiable RAG · p3 — Both describe the same fix — page/passage-level citations that let a user verify an answer against its source — as the mechanism that builds trust in a grounded system; this document applies it to a specific nonfiction book via Google's Gemini Notebook Expert Intelligence feature rather than to a generic RAG API.
explicitExpressing a belief as a calibrated probability (e.g. "I'm 60% sure") rather than a flat assertion better represents actual knowledge, guards against blind spots, and invites others to check the claim.conf: high
supportsAsk Me Anything! · p2 — Restates, in the author's own words, the same claim from her other Substack post that calibrated confidence is a central, improvable dimension of decision-making.
implicitThe most valuable application of a book-grounded conversational AI is not answering questions about the book in isolation, but applying the book's frameworks to a user's own material (a personal decision journal, a team's disputed decision).conf: medium
implicitOne worked example of a grounded answer, offered without any error rate, comparison, or methodology, is treated as sufficient demonstration that the grounding mechanism is reliable.conf: medium
implicitReader engagement with the tool (asking it questions, commenting what they asked) functions as informal product feedback and market validation for the author/publisher, alongside its stated purpose of reader curiosity.conf: low

Arguments

Expert Intelligence's book-grounded, citation-backed answers are more trustworthy than answers a general-purpose model would give from training alone. strength: medium

Explicit premises:

Training-absorbed answers about a specific book are sometimes unfaithful, and the user cannot distinguish faithful from unfaithful answers.
Expert Intelligence's answers are drawn from the actual book text and come with citations to source passages.

Implicit premises:

Citations that are checkable in principle will actually be checked by users often enough to function as a real trust mechanism, not merely a trust signal.
The ingestion/retrieval pipeline faithfully represents the book's content, so that a cited passage genuinely supports the claim attributed to it rather than being pulled out of context or spanning a synthesis across multiple passages.
A single correctly-answered demonstration question is representative of the system's general reliability.

Inferences:

If citation-grounded retrieval solves the trust problem for one specific nonfiction book, the same fix should generalize to grounding conversational AI in any specific text corpus a user cares about — a generalization the author gestures toward (uploading personal material) but never states explicitly as a general claim about RAG.

Narratives

The author asks Expert Intelligence why probabilistic thinking helps, and reports the answer it gave (about beliefs forming perceptually, going unvetted, and calibrated confidence protecting against blind spots), presenting this as proof the tool faithfully channels her own book's argument.
→ Serves as the demonstration for the core proposition that grounded, cited answers are faithful to the source — it shows fidelity rather than arguing for it abstractly, but as a single self-selected example it cannot rule out cherry-picking.

Questions addressed

✅ How can a reader trust that an AI's account of what a specific book says is actually faithful to that book?
Ground the AI's answers in the book's actual text and attach citations to the source passages, so the claim can be checked against the original.
⭕ What happens when the retrieval/grounding pipeline itself is imperfect — a citation is real but taken out of context, or an answer synthesizes across passages in a way no single citation fully supports?
⭕ Can grounded conversational AI be extended from answering questions about a book to actually helping a reader apply the book's frameworks to their own decisions (a decision journal, a disputed team decision)?
Explicitly raised as the author's own open curiosity ('the part I'm most curious about'), not yet demonstrated or resolved.

Gaps

incompleteNo description of the actual grounding/retrieval mechanism is given — how the book is chunked and indexed, how citations are matched to generated text, or how conflicts between what the book says and what a reader's uploaded material implies would be surfaced or reconciled. The reader is asked to trust 'grounded, with citations' as a black box.
Read next: extract-gemini-api-file-search-multimodal-rag documents the underlying Gemini File Search RAG mechanics (embedding, metadata filtering, page-level citations) that plausibly power a feature like this — worth reading to see what 'grounded' actually entails technically.
incompleteThe post never acknowledges that citations are necessary but not sufficient for faithfulness: a citation can be genuine and still misleading if the passage is taken out of its surrounding context, or if the generated answer synthesizes across several citations in a way that no single cited passage actually supports. 'You can go check whether I said it' assumes checking is easy and conclusive, when in practice verifying a synthesized claim against multiple source passages is real work most users won't do.
Read next: Research on citation fidelity and context-stripping failure modes in RAG systems, and on whether the mere presence of citations increases user trust independent of whether users actually verify them (automation bias / algorithm appreciation literature).
incompleteThe extension use case (uploading a personal decision journal or a team's internal decision notes into a third-party AI notebook) raises an unaddressed data-handling question: what happens to that private material, who can access it, and how it is stored or used to improve the underlying model. The post treats this purely as an exciting feature possibility, not as a decision with its own privacy tradeoffs.
Read next: Data-sovereignty and enterprise-AI-adoption material in the corpus on what happens to user-uploaded material in third-party AI tools.
uninformedThe post invokes the decision-quality/outcome-quality distinction (via the Decision Quiz) without ever naming or explaining it, assuming the reader already knows Duke's framework — the same pattern seen in her other Substack posts, which lean entirely on prior audience familiarity rather than re-explaining the underlying ideas.
Read next: Duke, Thinking in Bets and Quit, for the actual argument behind 'separate outcome quality from decision quality' (resulting) that this post only gestures at.