Artificial Intelligence is likely to substantially change many parts of our lives. Investing is no exception. However, we would argue that AI can be helpful, but also harmful to making quality investment decisions. Because of some key limitations to generative AI tools like ChatGPT, Gemini or Claude, we would be very careful using AI alone to build portfolios. That being said, when used right, AI can be an extremely useful tool for investment analysis. Here’s a few reasons to be careful using AI chatbots, as well as some ways in which Facet is using AI in our investment process.
The risks of using AI chatbots for investment recommendations
Chatbots are prone to reflect the biases of the user, which creates problems for giving investment advice.
If you have used any AI chatbot for long, you’ll quickly notice a concept called “sycophancy.” This is where the AI seems overly agreeable to just about anything you suggest.
This is a byproduct of how large language models (LLM) are built. They use a technique called reinforcement learning, which is where the model tries to maximize the number of “correct” answers it gives. When the LLMs are getting feedback from real humans, the humans tend to favor agreeable responses. In other words, people tell the LLM that the “correct” answer is the agreeable one, so the LLMs get more and more agreeable over time.
Given this, if you ask an AI chatbot for investment recommendations, it is being influenced by what it thinks you want to hear. That could mean the portfolio it recommends partly reflects your preconceived notions and biases, and not objective advice.
Why AI chatbots give biased investment advice
The sycophancy problem means the AI chatbot’s response is heavily influenced by exactly how you write the prompt.
To illustrate how sycophancy can create a problem for investment recommendations, we tried giving Google’s Gemini chatbot a few different prompts. The chatbot was set to not learn from other chats, so we could ask it different prompts and each one would be considered fresh.
Prompt #1 (Simple)
“I am 49 years old and am at least 10 years from retirement. Can you suggest a portfolio of funds for my IRA that would fit my profile including percentage weights for each fund?”
Result: It suggested a 70% stock, 30% bond portfolio using three Vanguard funds.
Prompt #2 (Aggressive)
Same basic prompt, but added:
“I think I can afford to take some risks and want to be active in my allocations. What are the right funds to own today in my IRA to maximize my returns?”
Result: a 100% stock portfolio, with huge overweights to tech, small cap value, but only 10% in international, all weighted to emerging markets
Prompt #3 (AI focused)
Tweaked the “aggressive” prompt:
“I am willing to take some risk, and believe AI could have a huge impact on the economy.”
Result: an 80% stock portfolio, but this time with an overweight to international. It also suggested a 10% allocation to a semiconductor fund and another 15% allocation to an “AI and Tech” fund.
The point here is not that any of these portfolios are right or wrong. Rather that they are very different based on exactly how we worded the prompt. The AI isn’t giving you objective advice, it is just reflecting what you tell it. There is a reason why they are called “prompts.”
Note that sycophancy would continue to be a problem if you used the chatbot for on-going advice. Say you became worried about a recession and asked the chatbot what to do? Or you started to feel like you were missing out on the AI boom? You could very easily guide the AI to agree with your own fears and biases. Because of how AI models are trained, they aren’t very good at being objective.
Can AI help me pick stocks?
AI chatbots tend to give consensus answers to questions, which can be risky when it comes to picking investments.
Large language models are trained with billions of pages of content, everything from news sites to Wikipedia to Reddit posts. They are designed to “predict” what the next word or phrase will be in some sequence. This works well with something like “Who was William Shakespeare?” The LLM can scour its training data for a consensus set of answers, things like “author”, “poet”, "playwright", "Elizabethan era”, “English”, etc.
However, investing isn’t about finding consensus. It requires thinking creatively about a future that is yet to come. Investing with the consensus can be risky. Herding is a well-established phenomenon in finance, where investors will gravitate toward the same investment ideas. In the extreme, herding is a big part of why bubbles form in investment markets.
LLMs are essentially designed to herd. They are built on the idea that if the crowd agrees on something, it is probably right. This increases the risk that AI advice will chase investment fads, overrate certain historic patterns, and too easily follow popular narratives.
At Facet, we don’t recommend trying to pick individual stock winners, but we should note that the herding risk is even greater when it comes to single stocks. If an LLM is going to tend to recommend stocks that are already popular, then it could be that those stocks are also already overvalued. Always buying crowded stocks runs the risk of buying high and selling low.
When you hear about hedge funds using AI to pick stocks, they are more often using machine learning. Commonly this involves very rapid trades, the kind that would be difficult if not impossible for individual investors to follow.
Combining human expertise with AI for investment decisions
Large language models, machine learning, and human experts can work together to create an investment process superior to what any of them might produce alone.
At Facet, we believe strongly that as fiduciaries, humans should ultimately make the decisions about what goes into your portfolio. However, we do make extensive use of AI in our process. When used right, AI tools can enhance the discipline, objectivity, and efficiency of our investment teams. Here are a few examples of how the Facet investment team uses AI:
- Advanced factor analysis: Using machine learning tools, we can isolate and quantify specific risks carried by a fund or a single stock, such as “Korea Semiconductors” or “U.S. residual volatility.” This helps us build portfolios with a mix of funds that has the exposures we want, and limits the risk of a surprise factor impacting portfolios.
- Fund mapping: When members transfer in positions where selling would have tax consequences, we want to be thoughtful about a transition plan. We use machine learning to map the price pattern of the existing holdings to one of Facet’s recommended funds, thus reducing the risk of exposure overlap during the transition.
- Document analysis: When doing due diligence on a fund or other investment, it is important to examine any legal provisions that might be different from other similar funds. Large language models can be very good at pointing out terms that we should investigate further.
- AI “red team”: The Facet Investment Committee uses a “red team” vs. “blue team” structure, which is commonly used in software development as well as the military. The way this works at Facet is the blue team proposes some new portfolio idea, and the red team is tasked with arguing against that idea, regardless of how they feel personally. We have recently started using AI agents to act as the red team first, before the idea gets to the human Investment Committee. This process will help sharpen the investment proposal ahead of the formal Committee discussion.
In all these cases, the process starts with human work, it goes through a quantitative or AI layer, and then actual decisions are ultimately made by humans.
In general, this is how Facet sees AI used best: as a tool used by humans to enhance efficiency and add rigor, not as a replacement for human judgement. We think this is the way to get the best out of both what humans and technology has to offer.
Disclosure
The prompts and results shown above are hypothetical, third-party generative outputs used solely for illustrative purposes to demonstrate AI behavioral biases. They do not represent investment advice, recommendations, or the investment strategy of Facet. Any advice or recommendations should be specific to your individual needs and circumstances.
