Thoughts on AI and financial advice (April 2023)

My views on AI, and specifically how it is likely to impact financial advice in the short- to medium-term, have evolved a lot in the last year or so.

Why? Because when the facts change, I try to change my mind. And over the past six- to twelve-months, the facts have changed when it comes to artificial intelligence (AI). This has changed my priors about whether, when, and to what extent, AI will impact our lives.

Later in this article I will talk about:

• What has changed my mind.

• Why we’re not quite at the point where AI will be able to provide high quality, comprehensive, holistic financial advice – but why I think we’re near.

• Some of the reasons why there might be a delay with AI financial advice becoming widespread, both technological and non-technological.

But first, I want to make some predictions about AI and financial advice, and share some of my previous thoughts about AI and financial advice.

And before I say that, I want to caveat everything by saying that this article is more of a contemporaneous record of my thoughts right now, and a gambit to talk about AI with thoughtful people. I’m sure a lot of my views will change radically, and perhaps sooner than later. I’m mindful that I might be buying too much into the hype of AI; I might also have a status quo bias. It’s hard to know!

What lays ahead

Here’s my best guess of what I THINK will happen:

• Financial advice will continue to be provided to clients in much the same way, at least for the short-term (say, five years). There various reasons for this. For one, the challenge of hallucination (aka AI bulls#!tting) is significant. Regulation (and fear relating to bad advice at scale) is also a significant factor.

• Over time, there will be services available that will be as good, or better, than almost any financial planner. My GUESS is that this will happen in five to ten years, but I can imagine this happening faster or slower.

• Even when we get AI of this nature, there won’t be one single model for providing advice. There will be models that have different biases and focuses. One challenge is that it will become difficult for consumers to identify ahead of time what is likely to be most aligned with what they are after. Incentives of the service providers will matter.

• I can imagine a scenario where the biggest challenges for organisations and for regulators will be quality assurance relating to advice that is given. Organisations will need checkers, and the regulator(s) will need to be able to check the checkers. The regulatory focus will, of necessity, have to relate more to appropriateness and quality of advice, more so than advice process, since the AIs providing the advice will operate largely as black boxes.

• In the immediate- to short-term, some financial advice individuals and organisations will use AI to be more efficient and productive. For instance, I can imagine a time not too far in the future where I can get ChatGPT (or some variation of it) to help me write client reports significantly faster than before. (At the moment, the biggest challenge is working out how to prompt it, and get it to generate reports without compromising privacy.)

• This MAY result in costs relating to advice being less expensive and more accessible. However, it could just as well result in greater profitability for those who utilise these tools effectively. My guess is that it will be a balance, with benefits accruing slightly more to the service providers.

• AI will mean that the impact of certain people and organisations will be amplified, when they are able to scale their advice more efficiently – for good and bad. There will be cases where this won’t work well. Privacy is an issue, and there will be leaks. There will also be people and organisations who use these technologies with ulterior motives. Others might do it for good motives – but will cause more harm than good. (“The road to hell is paved with good intentions”.)

My previous thoughts

Below is a summary of thoughts I’ve shared on this blog, dating back to 2017 and 2019 about AI as it relates to financial advice. (At the time I called it “roboadvice”, which is a term I’m ambivalent about.) In short:

• In 2017, I thought that “Roboadvisers will not make financial advisers redundant”.

• In 2019, I said “I have no fear of roboadvice in my industry – at least in the short-term future”. I then said that .”I think there’s a long way to go before robo advice can replace the advice I, and many other financial advisers, provide”.

In early 2017 I wrote about roboadvice as I understood it at the time. I said:

” I think roboadvice is a great service, and potentially disruptive to advisers who think of themselves as investment managers or stock pickers. Roboadvice services are typically very cheap, relative to, say, someone who provides the same service at many times the cost.

Roboadvisers will not make financial advisers redundant, however. “

I concluded:

“[Roboadvice] will challenge the value proposition of advisers who think their focus should be on products, not clients. But for those advisers who focus on clients, roboadvice and similar service will showcase the real value that good financial advisers provide.”

In saying this, I listed various types of “real value” that good financial advisers provide:

• “getting to know a client, and building a strategy that is tailored to their unique circumstances, needs, and objectives, which change over time”

• “helping clients to get clarity about their goals and priorities. This can involve helping to get couples onto the same page – as much as possible. “

• “helping clients to identify the uncertainties in their lives and develop a strategy for managing risks and gaining exposure to potential upside.”

• providing motivation and accountability as clients commit to a process and achieve long-term outcomes, which in many cases is simple but not easy.”

In 2019 I wrote about why I’m not afraid of roboadvice:__

” I have no fear of roboadvice in my industry – at least in the short-term future. In fact, I think there _’s a huge amount of space for roboadvice in the financial advice industry, and that effective use of technology is essential in order to serve people who have traditionally been underserviced ….___

BUT! I think there ’s a long way to go before robo advice can replace the advice I, and many other financial advisers, provide.

(Famous last words?) “

In the article, I talked about the types of conversations I have with clients. I talked about all of the things we cover, and how it “feels like I’m mining people’s souls”.

I said:

” Maybe, one day, a roboadvice service will be able to discuss these sorts of topics in a holistic, comprehensive way.

At the moment, the roboadvice services I ’ve seen are extremely narrow and myopic. They’re also created by product issuers who have an obvious incentive to recommend their own products – making them something of an automated sales tool.

Computers ain ’t great conversationalists.

And computers with ulterior motives are even worse.

For the foreseeable future, I ’m pretty confident that there’s no substitute for the human touch when it comes to advice.

Which is great for me – because it means I can continue to have terrific conversations with terrific people! “

Things have changed

When I wrote about roboadvice, I conceptualised it in terms of people being directed through a very specific set of questions, which would result in specific recommendations. There might be some fuzziness around the edges, but it would be a bit like someone going through one of those “pick an adventure” books, or being on clearly defined rails that would lead to a specific strategy or plan of attack, depending on some of the answers they gave. Basically, advice would be based on branched options, based on the information they shared.

I thought of roboadvice as being fairly myopic. Excellent for very narrowly-scoped advice, but not so much for holistic advice that considered a person’s broader circumstances, needs, and values and priorities.

The way I framed it was that “Computers ain’t great conversationalists”. And so long as that was true, there would be a need for people who knew their stuff and were able to have good conversations with their clients.

I’ve changed my mind on that.

Large Language Models

The main reason for this is the advent of Large Language Models (LLMs).

My understanding is that LLMs started to become prominent in around 2018, especially with the use of transformer architecture. (Don’t test me on the details!)

I first heard of LLMs when GPT-2 (released in 2019) was around, and became more familiar with them, in an oblique “that’s interesting and something I’d like to know more about” way when GPT-3 came out.

As with many people, the first true “holy s#!t” moment came when I first tried DALL-E 2, and it was the release of ChatGPT that really sealed the deal for me.

(All of these are OpenAI products, although there are other organisations and products that have my attention. Notably, there is Midjourney in relation to image generation, and Claude by Anthropic. I am also aware of DeepMind (the subsidiary of Alphabet) and its applications of AI to various fields, such as AlphaZero (beating world champions in chess and go) and AlphaFold [protein folding]. My understanding is that Alphabet and Meta have advanced models and capabilities up their sleeves. Having said this, I’ll focus on OpenAI for specific examples here.)

We’re not quite there, but we’re not far away

First up: one of my core hypotheses is that effective roboadvice becoming available is, or will be, less about technological issues and more about incentives. More on that later.

It might have been the case that computers weren ’t great conversationalists. That’s not the case now, and pretty soon I think they’ll be better conversationalists than humans in many contexts.

I already use ChatGPT to “think out loud” about things, and even in its current iteration have more interesting and useful conversations with ChatGPT-4 than I do with some other people. It has helped me generate insights, and influenced important decisions.

These chat faculties will get much better. Consider:

• When chat isn’t just via text, but becomes properly conversational. Ie, you can speak with it, and it can speak back. (Conversational input via high quality audio-to-text, and output via text-to-audio, executed so it sounds like an actual human rather than the robotic voices of today).

• When video comes into play: so you’re integrating with an avatar with natural movement, so it feels like interacting with a real person (or multiple people) via Zoom.

• When it can incorporate more information about you than just what you put in a text box, but it can hear you and/or see you, and take various subtle cues to calibrate how it responds to you. It might have algorithms to identify subtle changes in your breathing rate, temperature, sweat, heart rate, and pupil dilation. Imagine a service that can also extrapolate who you are based on what you share and your unique speech and body language – and calibrate to this.

Below are some reasons why I think this is somewhat inevitable. In short, I think AIs will become more capable than human beings in many ways, especially when it comes to conversation.

They won’t have many of the limitations we do

Computers don’t become tired, bored, or distracted in the way that humans do.

Computer systems are constrained in lots of ways: unless expanded, they are limited by their ROM and RAM and the size of their hard drives. There are limits by their compute (ie, the power of their CPUs and GPUs.) Historically, they’ve been much more constrained than humans.

The key words here are “unless expanded”. Computer systems can be expanded, and the hardware at the disposal of many of these AI companies is enormous.

Whereas humans are biologically limited, there is theoretically no limit to how computer systems can be expanded.

An AI will also be able to monitor more discrete things at a time. I think about this regularly when it comes to self-driving cars. We have a limited number of sensory inputs: two eyes and two ears for monitoring visual and audio aspects of our environments. A self-driving car can have dozens or hundreds of visual and audio inputs, and these inputs can track things that we are not capable of tracking.

As I write this, ChatGPT-4 has a maximum token limit of 32,000 tokens. That translates to approximately 25,000 words. This is a huge leap from GPT-3.5’s limit of 4,000 tokens (~3,125 words).

For reference, the following novels are less than 50,000 words, or less than twice what ChatGPT-4 can hold within its token limit:

• The Great Gatsby
• Slaughterhouse-Five
• Fahrenheit 451
• The Hitchhiker ’s Guide to the Galaxy

Another piece of context is that the pace of conversation is usually 100-150 words per minute. That’s 167 to 250 minutes, or roughly 3 to 4 hours’ of conversation.

Two things:

• I imagine that GPT will blow past 32,000 tokens sooner than later.

• Even if it doesn’t, I’m sure there are ways to make the most of these 32,000 tokens. When I have a conversation with clients, I don’t need to note down everything word for word. Not only this, I can use shorthand to remember various things. Another way of thinking about it is that JPG files save a lot of visual information in a very efficient way. I’m sure there are ways to make the most of 32,000 tokens to save a lot more substance than 25,000 words indicates.

They will become more persuasive than human beings

Some people are more persuasive than others. Maybe some people are born with this set of skills. However, it can also be learned and cultivated. Even someone who is naturally persuasive will usually become more persuasive as they calibrate the way they communicate to the way people respond to them.

The amount of experience an AI will be able to accumulate will be enormous compared to the experience of any one individual. In particular, an AI system can learn from many interactions simultaneously.

This is another area where my expectations have changed. For as long as I can recall, I’ve thought artificial companionship would eventually become a big deal. (Think: sexbots but more wholesome.) Until recently, I thought the issue will be emotional rather than physical. Now, I think the physical side is going to be much harder.

Many people are more willing to share personal information with an AI than another human being

Most of us don’t like sharing our dirty laundry with other human beings. Even if this is in a context where privacy and confidentiality are guaranteed. Even when we implicitly trust the people we are dealing with.

There is a desire to put our best foot forward. There is also a desire not to take up too much time talking about things that you don’t think are relevant, but might be.

When it’s “just a computer”, people seem to be prepared to share more than with other people. Take what we share with Facebook or Google, for example.

Not so fast!!

I think there are many things that will delay the widespread availability of AI-generated financial advice.

I could be underestimating the importance of genuine human relationships

Guilty as charged.

The problem of bulls#!tting

As it currently stands, LLMs are often excellent bulls#!tters. It isn’t uncommon for you to ask it a question, and it will make something up. (The term of art is “hallucinate”.)

What is particularly scary is that it hallucinates in very confident way. Unless you know your stuff, it’s hard to identify what is true and what is not.

I think there are or will be ways to address some aspects of this. However, for the time being it’s a big issue. So you need to be alert when asking it. It might make someone with subject knowledge more efficient, but it can misdirect someone else very easily.

You really need to “trust but verify”.

Regulatory reluctance

There is currently a regulatory regime that relates to giving financial advice. This includes advice that is facilitated by systems as well as individuals. The current regime in New Zealand is new. It took a long time to be drafted, and there was a long transitional period as well. For reference, the first reading of the bill was in 2017 – and that was after a period of consultation. The bill received royal assent in 2019. It was only on 17 March 2023 that it fully came into force.

Regulation tends to move slowly. I also think there will be reluctance to change the rules in certain domains. When the domain relates to personal finance – or health or legal rights – then this will especially be the case. There’s a reason people often qualify their statements by saying they’re not providing medical, legal, or financial advice.

To begin with, a good proxy is to ask where are the domains where we as society can afford to be messy. In creative areas, we can. When it comes to money, I’m not sure.

(However, there is a balance here. It’s easy to focus on risks of action – ie, instances of AI-generated advice being harmful. But there are also risks of inaction – ie, lots of people not having advice when it was possible for them to receive it. Ultimately, we are looking at the balance of harms.)

Underlying models seem to be in flux

You can build a service, but the underlying model might change, which might change the outputs significantly. There will be a need for constant monitoring.

Physical limitations

This is less relevant to financial advice, but I feel compelled to note it. I think any domain that requires physical presence will take a long time to be replaced. If your role involves fine motor skills that need to be used in very nuanced, varied ways, on top of a foundation of deep, specialised knowledge, my guess is that you’re going to be safe. (As opposed to fairly rote processes, which were probably already at risk of being automated. Although AI will probably expand the range of tasks that can be automated.)

The problem of incentives and bias

I actually think the challenges relating to AI-driven financial advice aren’t technological in nature, so much as bias- and incentive-driven.

I don’t think it’s any coincidence that the organisations who have sunk the most resources into “roboadvice” to date are product distributors. Perhaps I’m being too cynical, but there’s a part of me that thinks a lot of financial product providers think of roboadvice as yet another tool for product distribution, rather than a tool for providing the best possible advice they can (which might not be in their best interests).

I can also imagine scenarios where AI becomes really effective for scammers, or people with ulterior motives. Or simply people with good intentions but bad ideas. In fact, there are plenty of financial advisers (including people who are fairly prominent), where the idea of them scaling their views to address an even bigger audience, really concerns me.

Or: consider that there might be different versions of AI-facilitated advice, provided by different organisations. Think of a few other people who write or talk about personal finance in NZ, and imagine a scenario where each of us were given the resources to generate an AI-based system based on our own central presuppositions, biases, and views of the world. They would all be different. Some of them would be VERY different – with different weightings given to debt and certain asset classes (property versus financial assets versus cultivating human capital versus investing in own ventures). Some bias might be built into each of these models, and some models might be better equipped to talk about certain strategies very narrowly and specifically, versus others that might be better equipped to talk across different types of strategies, in a more general way.

Much as there are different versions of “good” advice, different AI models could spit up very different suggestions that might be within the realm of “shades of right”.

AI is a tool, and it will allow some advisers and organisations to service more clients, more efficiently. The efforts of these advisers and organisations can be magnified. This creates opportunities. However, there is an enormous risk of terrible advice being provided at scale.

Strong opinions, weakly held

This leads me to two strong opinions (weakly held):

• The limitations that will restrict widespread AI systems providing financial advice will not be technological. There will be several factors, including regulatory issues justifiably related to the possibility of bad advice at scale. I also think there will a correctly held fear from many consumers that the advice they are receiving from AI systems might suffer from hallucinations (bulls#!t) and/or won’t necessarily be in their best interests.

• There won’t be a single AI that serves everybody. My intuition when first thinking about AI was that it would be a “winner-take-all” situation, where the provider that gets a head start or provides the “best” system would end up serving the majority of clients. However, I think there will be different models that provide different advice in different ways, and it will be a competitive market. There might not be a lot of “big” providers. But I don’t think it will end up as a monopoly. A thousand flowers will bloom, and I have no idea what the precise outcome will be – but in general terms, I think the financial advice landscape could be very different in 10 years’ time to what it is today, and AI will be a significant factor.

Whatever happens, we’re living in interesting times! And financial advice is not the only domain that will be affected.