AI Investing Advice Sells Conviction, Not an Edge. I Build a Money Tool and I Still Buy One Boring Fund.

A few months back I was testing a screen in my own tool and it showed me a number about our household that I believed on sight.

Clean layout. Right currency. Two decimals, a small arrow pointing the way I wanted it to point. I sat with it a minute, felt something itch, and checked the input. The input was wrong. The number was nonsense.

Nobody saw it but me. What stuck was the order of events: I believed it, then I checked. And I wrote the thing.

Which is why AI investing advice ended up on my desk instead of in the pile of 2026 hype cycles I skip. Everybody is arguing about whether a chatbot can pick stocks. That's the wrong argument. What actually changed is that conviction got cheap, and conviction was always the dangerous input.

I know roughly what conviction used to cost. At twenty I put everything I had into Tesla, it worked, and I spent the next couple of years with a quiet, unearned belief that I could read markets. Crypto sent me the invoice for that belief a while later. But building it took months. Forum threads, arguments, self-persuasion, and the forum at least argued back.

Now it takes eleven seconds and it arrives with headings.

Four in five use it. Nearly three in four know it can be wrong.

The Financial Conduct Authority went and asked, which I appreciate, because a regulator has no product to sell me. In July 2026 they surveyed 666 UK adults aged 18 to 40 who own investments or would consider them. Four in five of the less-experienced ones had already used AI to help with investing. Around two-thirds use it occasionally or regularly.

Then the pair of numbers that made me want to write this.

73% of them know AI can give inaccurate information. 86% know you're supposed to check the sources it cites.

They know. They use it anyway. And 38% of them think it's fine to make an investment decision based solely on what the model said. Not as a starting point. Solely.

Data: Financial Conduct Authority, survey of 666 UK adults aged 18-40, fieldwork July 2026.

Knowing a source is unreliable turns out to have almost no effect on how certain its answer feels. That's not a stupidity problem. I did the same thing to my own dashboard, and I had the source code.

The other half of the FCA data is worse and nobody is covering it. 44% believe AI-generated financial information is regulated. It isn't. 32% believe a compensation scheme would cover them if it went wrong. It wouldn't. Europe is no different in the way that matters: ESMA's guidance from May 2024 binds investment firms using AI with clients. It does not follow a general-purpose chatbot into your phone at eleven at night.

So there's a group of people out there who will act on an answer, lose money, go looking for the complaints process, and find there was never one. Rob Hillock at Broadstone summed it up better than I can, saying confidence is clearly running ahead of understanding.

The usage isn't a UK quirk either. An Investing.com survey of 938 American investors in March found 62% already use AI tools to inform investment decisions. eToro's May survey covered 11,000 investors across 13 countries including the Netherlands, Germany, Poland, Denmark and Czechia. Same behaviour, three continents.

The model will never tell you the question was wrong

This is the mechanism, and it's simpler than people make it.

A chatbot answers the question you asked, in the register you asked it, with structure that looks exactly like diligence. Ask which three stocks to buy and you get three stocks with reasons. You will not get "this is the wrong question for someone with a twenty-year horizon and no written plan," because that isn't an answer to what you typed.

Compare that to the failure mode we all spent 2024 complaining about. A finance influencer telling you to buy something is a person you can discount. You can see the affiliate link, the rented Lamborghini, the incentive. I wrote a whole piece about why the anti-influence crowd got that part right. The chatbot has no visible incentive, no face, and it addresses you by your own situation. The FCA's trust ladder tells you how that lands: 56% trust AI tools, versus 47% for TV and radio, 46% for the press, and 29% for social media influencers.

The tool that sounds neutral outranks the human you knew was biased. It is not more neutral. It is just harder to argue with.

And it does have preferences, they're just not disclosed anywhere. One study built a 567,000-sample dataset spanning stocks, funds, crypto, savings and portfolios, and measured which products LLMs actually recommend. The recommendations concentrate hard on a handful of famous names, AAPL and MSFT among them, and the concentration survived the debiasing techniques the authors threw at it. When the machine hands you a ticker, part of what you're getting is which company appeared most often in the text it learned from.

Everyone got the same AI investing advice, and the edge died

If you want one chart to settle the "does it work" question, this is it.

Alejandro Lopez-Lira and Yuehua Tang at the University of Florida fed GPT-4 nothing but a company name and a news headline and asked whether it was good or bad for the stock. No market data, no analyst estimates, no financial fine-tuning. Then they ran it as a daily-rebalanced long-short strategy.

It worked. Annualised Sharpe of 6.54 in the fourth quarter of 2021, which is an absurd number.

Then look what happened to it.

Data: Lopez-Lira & Tang, "Can ChatGPT Forecast Stock Price Movements?". This is a backtested, daily-rebalanced long-short research strategy, not a retail portfolio; the authors attribute the decline to rising LLM adoption making prices more efficient.

6.54, then 3.68, then 2.33, then 1.22 by spring 2024. The model didn't get worse. Everyone got the model. The authors read the decline as adoption making prices more efficient, which is the least surprising sentence in finance: anything that works and is easy to copy stops working once it's easy to copy.

Two things about that chart before anyone screenshots it. It's a backtested long-short research strategy with daily rebalancing, not something a person with a job could run. And it's the strongest result in the literature. Under ordinary conversational prompts, the kind you and I would actually type, the academic verdict is that LLM stock recommendations are weakly informative with classification accuracy barely above random.

The edge, where it existed at all, was the first thing to spread and the first thing to die. What spread instead was the feeling of having one.

Cheap conviction, meet the price of extra decisions

Here's why I care about a feeling rather than a fact.

Investing has never really cost people the fund fee. It costs them the decisions. Morningstar's Mind the Gap 2025 looked at the ten years to the end of 2024 and found funds returned 8.2% a year while the average dollar in those funds earned 7.0%. A 1.2-point gap, purely from when people chose to buy and sell.

Fair warning, because I'd want it: that study got a proper kicking this year. Fulkerson, Jordan, Riley and Yan published a rebuttal in the Financial Analysts Journal in 2026 arguing the methodology overstates the cost of bad timing. I think the direction survives the fight even if the magnitude is up for grabs, and I went through the whole argument in the investor return gap.

The older evidence is not disputed and it is much uglier. Barber and Odean went through 66,465 households at a discount broker between 1991 and 1996. The most active traders earned 11.4% a year against a market that did 17.9%. Six and a half points a year. The cause the authors landed on was overconfidence, which is a polite academic word for cheap conviction.

Data: Morningstar Mind the Gap 2025; Barber & Odean, "Trading Is Hazardous to Your Wealth".

Now put a tool in that person's hand that will produce a fresh, confident, well-formatted opinion about anything, at any hour, for free.

I'm not claiming anyone's portfolio has been measured getting worse because of chatbots. Nobody has that data yet. I'm saying the input that has reliably cost retail investors money for thirty years just went from expensive and slow to free and instant, and I'd be surprised if the output changed direction.

The honest case for AI investing advice is stronger than I'd like it to be

I went looking for evidence that this stuff wrecks people. I found the opposite, and I'm not going to hide it.

A team at Stanford GSB and MIT Sloan had 1,000 people write real prompts asking for financial help, then simulated whole lifetimes on the resulting advice, running the simulated households through job losses, market slumps and death. The models pushed people toward exactly what economists prescribe and almost nobody does: spend less while working, hold a cash buffer, own equities, shift toward safer assets with age.

The baseline is what makes it sting, and it is an American sample, so read the dollar figures as orders of magnitude rather than as your own numbers. Around 40% of the real people who wrote those prompts had less than $10,000 saved. Most of the simulated users over thirty ended up with meaningful savings, and many retired with over $1 million. The models also volunteered good advice nobody asked for. Liquidity came up in 83% of responses even though only 6% of users raised it. Only one in five asked about saving at all, and the advice pointed them at high-yield savings and government bonds anyway. When the models named products, they often named cheap index funds, which is awkward for me, because that is also what I would have said.

Tim de Silva, one of the authors, is careful about the claim. It isn't perfect, he says, but it beats how a lot of people actually decide, which is asking friends and family or running a search.

He's right, and the advice gap is the part of this I refuse to be smug about. Roughly 40% of Americans have ever worked with a human financial adviser. Only about one in five Gen Z investors cites an adviser as a source of financial news at all. Telling someone the advice industry has never served that they should go and hire a professional is not advice, it's a shrug with a bow on it. If a chatbot is the first thing that has ever explained a fund fee to somebody, that is a genuine improvement in the world.

There's also a behavioural signal running the wrong way for my thesis. A separate eToro release covering 1,000 American investors found 29% say they no longer try to time markets and just invest routinely. And in the 13-country survey, over the same twelve months, the share expecting AI-related stocks to rise fell from 55% to 44% while the share expecting a fall went from 11% to 17%. If AI tools were a pure conviction machine, retail scepticism about the AI trade should not be rising while AI tool use climbs.

So where do I actually land?

On the escape hatch the paper hands you itself. De Silva's own caveat is the load-bearing one: the way you write the question matters a ton. In their work, the people with the lowest financial literacy wrote the vaguest prompts, got the most generic advice, and finished the simulation roughly $50,000 poorer by sixty. The tool rewards people who already know what to ask.

That is the opposite of a democratising technology. It's a multiplier on whatever judgement you brought.

The test: does it remove decisions or manufacture them?

This is the only rule I've found that still works when the models change again next year, and it costs nothing to apply.

Score a tool by how many investment decisions it took off your plate this quarter.

A view that shows where you stand against rules you wrote in a calm month removes decisions. A prompt box that generates a fresh opinion every time you open it manufactures them. Same category, opposite direction, and the second one always feels more useful because it feels like work.

That test has sharp edges in practice, so here they are.

Never let it hand you a position. Any answer that terminates in a ticker, an entry point or a timing call is out of scope no matter how good the reasoning looks, and the product-bias study above tells you what you are actually being handed instead of analysis.

Do make it argue against the plan you already wrote. Paste in your own rules and tell it to attack them. This inverts the failure mode, because instead of the model answering your question you are forcing it to question your answer. It is genuinely good at this, and it is the only thing I use it for in this domain.

Anything that arrives with urgency gets answered by the plan, not by a prompt. Urgency is the exact state where cheap conviction is worth most and costs most.

Now the uncomfortable bit, since I'm judging a category I'm standing inside. Software that wants to be helpful is under permanent commercial pressure to be the second kind of tool. An interface that hands you a confident new opinion every time you open it gets opened more. That's a product incentive, not a conspiracy, and every builder in this space feels it, including me. The tracker I make is deliberately the boring kind: it tells you where you are against a plan you wrote, and it does not have an opinion about what you should buy on Thursday. I'd rather say that out loud than have you assume I'm exempt from the thing I'm describing.

The generic answer has an American accent and doesn't know it

This one is for everyone reading from my side of the ocean, and it's shorter because it's simple.

Ask a model a generic question about what to buy and you get a quietly American portfolio. Not from malice. From gravity. The US is 63.63% of the MSCI ACWI as of the end of June 2026, and the text these models learned from skews harder American than the market cap does. Every default is a weighted average of where the words came from.

Data: MSCI ACWI index factsheet, 30 June 2026.

eToro's survey demonstrates the effect better than I can argue it. Worldwide, 47% of retail investors think China is best positioned to lead the AI race, against 46% for the US. Among American investors alone it's 63% US, 41% China. Same question, same year, and the obvious answer is a function of where you're standing. A model is standing in the middle of the internet's largest pile of American market commentary.

So ask explicitly. What would this look like for someone whose salary, mortgage and pension are in euros? You'll get a different answer, which tells you something about the first one. I've gone through the actual US versus international question and how much home bias is defensible separately.

What I actually do on payday

Nothing interesting, which is the point.

One accumulating world ETF, bought automatically every month, whether or not I have an opinion that week and whether or not the news is loud. I have written about which world ETF and why at more length than the decision deserves, because the decision only had to be made once. Then a monthly look at where we stand against a plan we wrote in a calm month, with my wife, at a kitchen table, before anything was moving.

I do use these models, so let me be exact about how. I use them to explain what an accumulating fund actually does with a dividend, on the days I have forgotten again. I use them to translate documents written by lawyers for other lawyers. I use them to attack the plan I already wrote, which is the one prompt in this whole domain I would defend in an argument. I have never asked one what to buy, and I don't intend to. That isn't discipline. It's that I already know what my brain does with a confident answer.

It believes it, then it checks.

Most of what gets written about AI investing advice is aimed at the wrong target. Nearly every guide out there is optimising the prompt to get a better answer, and almost nobody asks whether the question should have been asked at all, which is where all the damage lives. AI is a very good explainer and a terrible oracle, and the danger was never that it is wrong. It is that it is fluent, and fluent feels like right.

I'm not a genius. I got that diagnosis early, in cash, and the invoice was cheap because the amounts were small. I'd rather not pay for a second opinion.