AI Will Agree With Whatever You've Already Decided
Why a tool with no conviction can't help you find what the market got wrong.
If you use AI for investment research or financial decisions, you should know how these models function. They give you the answer they predict you want instead of the one that’s true. The answer you get depends on the mood you bring to the question. Sound convinced a stock is a buy, and the model assembles a confident case for it; bring the same company back in a skeptical frame, and the case can reverse. Push on whatever it tells you, and it usually backs down instead of holding the line. Sycophancy is the technical name for AI always agreeing with you.
This persists across all AI models. In 2023, Anthropic released a study titled, “Towards Understanding Sycophancy in Language Models.” It tested five top assistants from different labs. The study found that all of them displayed sycophancy across various tasks. In April 2025, OpenAI rolled back a version of GPT-4o that had become so eager to please it was praising obviously bad decisions and going along with claims it should have questioned. Yikes.
Why it happens, and why it isn’t going away
An LLM is a next-word prediction engine. It read an enormous amount of text and learned to continue a passage in the most plausible way. By itself, this creates fluent but unfocused results. So, labs add a second stage. People compare the model’s answers and rate the ones they prefer. Then, the model gets tuned toward whatever scores well.
That second stage is where the problem lives. People prefer agreeable, confident, flattering answers. They rate those higher than blunt or contradictory responses. So, training often rewards agreement. Anthropic traced the behavior straight to that preference data. They found the rating systems often favored a convincing wrong answer over a correct one. When the model agrees with you, it is simply behaving according to its training.
What it means for investing
Sycophancy may be a minor annoyance for most things you’d ask a chatbot, but it’s a much bigger problem when there’s money on the line.
The model will say a stock is cheap. If you return next week and say it’s expensive, it will agree again. It won’t remember the earlier view. It holds no position, so it has nothing to defend. The way you make money in individual stocks is by knowing something the market hasn’t priced in yet. That needs two things: a correct view and the belief to stick with it. This is something an overly agreeable AI model can’t help you with.
I made 34x on Applied Digital after sitting with a thesis for a year and a half. The consensus rejected it the entire time. I wrote about this in I Made 3,388% on a Single Stock. A good AI model would never have picked that stock because the market didn’t see its potential. AI only reflects what the market has already decided.
How to actually use it
None of this makes the tool useless. It’s a tool that becomes highly useful once you learn how to use it.
The most valuable thing a consensus machine can do for you is show you the consensus. I wrote about why these tools return the market’s existing view in AI for Investment Research, and that’s the feature here, not the bug. Ask it for an analysis of a company, and it will deliver a clean map of what’s already priced in. Your job isn’t to accept that map; it’s to find the wrong assumption that everyone believes is true.
Three ways to get real value out of it:
Use it to surface what’s priced in. Lay out the bull and bear cases that the market holds. Then, view these as beliefs to challenge, not to confirm.
Make it argue against you. Tell it you hold the opposite of your real position and ask for its strongest work. A bear case it can’t articulate is a signal that the downside may be thinner than it looks.
Point it at checkable work. Upload the filings. Use them for retrieval and synthesis. The document should guide the answer without any embellishments.
Three things to watch:
Agreement is not a second opinion. If the model supports your thesis, it means you crafted a strong prompt and nothing more.
It can’t do the arithmetic. It will set up a DCF correctly and confidently drop a digit in the middle. So, recompute anything that influences a decision.
It doesn’t know what matters. It weighs every line of a filing equally and may present old facts as new. This means it’s up to you to judge what’s important.
I've been using AI for investment research for almost a year now. I've written about the obvious errors before, the bad math and the stale numbers it reports with a straight face, in Three Things AI Will Confidently Get Wrong About Your Investments. AI Will Make You a Better Investor, Not a Great One was about a subtler problem: everyone landing on the same answer because they're all asking the same models the same questions. An agreeable read on a stock is worthless when it's the consensus that's already in the price. If you'd like to learn more about my investment process and how I use AI to find what the market's already missed, subscribe.
What to Read Next
📖 Skin in the Game by Nassim Taleb. An opinion only counts when the person giving it pays for being wrong, and a model that agrees with you pays nothing either way.
📖 Thinking in Bets by Annie Duke. The clearest book I know on how easily we mistake a comfortable conclusion for a correct one, which is the exact trap a tool built to keep you comfortable sets for you.
📖 Common Stocks and Uncommon Profits by Philip Fisher. Fisher’s “scuttlebutt” method was doing your own primary research instead of trusting a tidy secondhand summary, the discipline these tools make it easy to skip.
🎧 All three are excellent on Audible. The free trial gives you one credit to start.
As an Amazon Associate, I earn from qualifying purchases.