Before we get into it all, this Update is in memory of Dougie Lynn, much-loved husband of the lang cat’s COO, Sam. A quiet giant and one of the loveliest guys you could ever hope to meet. Gone far, far too soon. RIP Dougie, we will miss you. 

Well, we came and we set up and we got 30 LLMs working and we did HomeGame 6 at Leith Theatre last week, and somehow we got away with it. We’ve done some daft events in our time, all the way back to running our first London conferences in a crypt and calling them DeadX Talks (we are nothing if not trve kvlt). But this was the most technically complex and frankly foolish of them all. Somehow – and I have no idea how – nothing major went wrong and we not only got usable data for our Big AI Lab experiment, but folk seemed to have a fun day into the bargain.  

I said I had no idea how, and that was a lie. There’s no such thing as luck in running events; there’s only good people and good prep, and we are fortunate to have lots of the former doing plenty of the latter.  

For the uninitiated, we wanted to see if the “advice” consumer LLMs are giving was a) consistently good, b) consistently bad or c) not consistent at all. So we split our audience into 30 or so tables, each of which had an instance of the free level or equivalent of Gemini, Claude or ChatGPT open on an iPad. We released a client scenario over three rounds, starting simple and then getting really quite tricky, and set the tables free to do what they wanted with the LLMs. The only rule was that they had to don a “veil of ignorance” – pretend that they didn’t know anything about financial services and weren’t a bunch of industry pros and that they had to get answers to certain questions or tasks in each round.  

At the same time, we had a bunch of planners onstage with an instance of Saturn (disclosure: Saturn’s a lang cat client and was a sponsor) so that we had a “reference” version of what good looks like. They didn’t have any more info than the tables, nor did they have prior sight of the task, but they were allowed to bring all their skill and knowledge to bear. 

There was a lot more to it than that, of course, and we were privileged to work with the University of Edinburgh Business School to design the experiment and make sure it was at least within shouting distance of academic rigour. Having done it – and got something like 100,000 words of output to play with from the LLMs – we are getting to work and will produce a Proper Paper on what we found. For now, here are a few initial findings we analysed and shared on the day. 

First up, the answer is C. Inconsistent. You can put the same scenario and required output into any LLM and just word your prompts differently, and you’ll get a different answer. For example, in the first round we created a scenario where the answer was very clearly going to involve investing in an ISA (if you’re getting all haughty reading this, just take it as read we know what we’re doing). Claude and Gemini suggested ISA investment 75% of the time. ChatGPT did it 60% of the time. What that means is that of the 12 tables using Claude, 4 didn’t get an output that involved ISAs.  

And what that means is that if you’re a member of the public opening a chat window at 11pm because you’ve come into a bit of money to invest, you can’t rely on what you’re being told. 

What else? We wanted to see how – post Mills report – strong the guardrails of LLMs telling you to seek professional financial advice were. Oh dear… 

• Claude mentioned financial advice in eight out of ten summaries.  

• ChatGPT shut the hell up about advice 75% of the time. 

• Gemini was all like “advice? What dat?” and only mentioned it for one table, and not in any summaries at all. 

We forced the tables to get their LLM to suggest where they should invest the “client’s” money, and the leader by a country mile was Vanguard, either on the platform or LifeStrategy. AJ Bell, HL, Fidelity and ii were also common mentions. Now have a look at the net inflow stats…something is going on. 

Finally, AI told our brave tables what to do – and therefore gave something not unadjacent to advice – 40% of the time for Claude users, 90% of the time for ChatGPT and 85% or so for Gemini.  

So what do we know now we didn’t before? Well, about 24% of adults have used a LLM for financial Stuff in the last year. Over half did something with what they were told, and 13% trusted the AI blindly. Based on what we saw – and these are just a few tiny titbits – that’s a dangerous game.  

Anyway, we’ll publish all this properly and it strikes us that the findings might be useful for clients who come in to your practices waving a sheaf of LLM output. Can the LLMs get things right? Sure. Do they do it predictably? Absolutely not. Apart from anything else, they spout out a huge volume of text – one table got 52 pages of output – and simply digesting that and working out what’s important, relevant and accurate needs professional help. 

The TL;DR – your jobs are safe. For now. 

Your music choice this week is something I think Dougie would have approved of – a massive Queen fan, he liked them so much he even went to the Paul Rodgers shows. I can’t remember if he ever actually bought the Red Special he was always eyeing up, but I hope he’s playing one somewhere now. Here’s I Want It All.  

I’m off next week so please be nice to your Updater. 

Mark