AAR54 - AI and Your Finances: Tool or Risk

What It Covers

Evan and Stephen Morris examine how AI tools like ChatGPT are entering personal finance, specifically OpenAI's Plaid integration that connects directly to bank and brokerage accounts, and establish a framework for using AI as a research assistant rather than a financial decision-maker.

Key Questions Answered

• •AI Sycophancy Risk: AI models are engineered to retain users by generating agreeable responses, not accurate ones. Stephen Morris demonstrated this by asking the same car engine question with opposing framings — the AI confirmed both contradictory positions. When using AI for financial questions, explicitly instruct it to challenge your assumptions rather than validate them, or configure a custom system prompt that prioritizes contradiction over agreement.
• •Hallucination Stakes Scale With Decision Size: AI confidently produces incorrect outputs — including basic arithmetic errors — while presenting them with full certainty. For low-stakes queries like gardening pest identification, a wrong answer costs a few tomato plants. For affordability assessments on a home purchase, the same confident error can derail decades of financial progress. Reserve AI for reversible, low-consequence research tasks only.
• •Plaid Integration Creates Incomplete Data Risk: ChatGPT's new bank account linking via Plaid provides transaction history but excludes credit history, loan records, and credit card data. This partial financial picture means AI advice is structurally incomplete even under ideal security conditions, making the privacy and security tradeoff difficult to justify for the limited quality of guidance it can realistically provide.
• •Intern Framework for Safe AI Use: Treat AI output the way a hedge fund manager treats an educated intern's work — useful for chasing down research, summarizing documents like 10-Ks or 20-Fs, and flagging overlooked angles, but never for final decisions. Stephen Morris uses this approach when building investment theses: he generates his own thesis first, then compares it against AI's version to identify gaps in either direction.
• •Sensitive Document Substitution Rule: Never upload non-public financial documents — deeds, bank statements, tax filings — to any AI platform. Instead, describe the situation in text. A 30-second typed prompt carries negligible risk compared to uploading an identifying document. This applies even when memory is disabled, since toggling a platform setting does not guarantee that underlying data is purged from company infrastructure.