Andrew Wiseman gives his own ‘f*** you, I shouldn’t’ve done what you told me’ after coming up short in this year’s Mustard Fantasy Premier League (FPL) mini-league.
You might remember a blog from this time last year, where I wrote about using gen-AI to help me win Mustard’s Fantasy Premier League over our serial winner, Richard Walker. If not, you can read it here.
Fast-forward 12 months, and the normal balance in our FPL world has returned, with Rich once again taking home the less than illustrious prize of the “Absolute Mustard” title. All this despite being bottom and adrift at Christmas.
This year, as last, I used gen-AI and followed it religiously through the year. As far as I know, Rich didn’t use AI. Maybe he did – a T-1000 version of the AI Terminator, compared to my T-800 legacy version? You’d have to ask him.
The limits of AI in FPL
So, what did I learn from this chastening defeat, having appeared to be home and hosed during the opening half of the campaign? First, it demonstrated the brilliance of AI and its ability to crunch the numbers and come up with recommendations based on what had gone before. What I found it didn’t do especially well was identify those ‘differentials’ (slightly left-field decisions to gain an advantage). These are so often the decisions that determine the winners from the also-rans.
Sure, the AI was steadfast in its support for Liverpool’s Mohamed Salah not only being in the team but being captain, but it was also painfully slow in spotting the emergence of the likes of Morgan Rogers and Bryan Mbeumo as ‘must-haves’ within the team, I presume based on the historic dataset from which it made its predictions. That’s without looking at the stoic Nottingham Forest defence, the emergence of Marc Cucurella as an attacking weapon or the various other players that you might not otherwise choose based on historic data only.
The limits of AI in insight – and what it means for market research professionals
This got me thinking further about two things that were already bothering me about AI.
First, using AI as a predictor of the future is purely based on what has happened in the past. After all, that’s the training set that the model is programmed against. So it probably shouldn’t come as a surprise that those players with no previous track record in the game were slow to rise to the surface, versus those with multiple seasons’ worth of data.
Second, it points to the dangerous ‘norming’ in the use of AI. As a data geek, I’ve always been a strong believer of the ‘shit in, shit out’ narrative, and the same can be said when it comes to how we construct prompts in our model of choice. The danger, as more and more of us start to use AI, whether as a co-pilot or otherwise, is that ‘vanilla prompts’ lead to ‘vanilla outputs’. For creative industries like ours, this is a massive risk. Do we all coalesce around the top of the normal distribution curve, peddling the same messages? Or do we maintain creativity by moving towards the tails of the curve to give us a differential viewpoint?
These for me are the questions that we should be thinking about. And given the FPL season that’s just gone, there’s definitely room to be less Erling Haaland and more Chris Wood. With that, I hand over the virtual crown to Rich, and in true football manager style, get ready to ‘go again’.
