Robots in disguise - we need to talk about AI

Unless you’ve been living under a rock, it can’t have escaped your notice that Artificial Intelligence has fully manifested itself in the mainstream. Yes, we’ve had it in our lives for years, but nobody really appreciates the AI function behind how Google operates, or the fact that Alexa can turn your lights off and ask how you’re doing. The release of ChatGPT gave the public their very own chatbot, and understandably millions of people spent an awful lot of time asking it questions, which Google was quite capable of answering, when they should have been in the pub with their friends. The growing awareness of ChatGPT brought about almost daily conversations with both the people around me and clients, as to how it was going to affect my business, and so began my relationship with AI.

THE EARLY YEARS

In the early years of our relationship, AI was the giggling teenager of ChatGPT2, full of promise and potential but still very much prone to making huge errors and lapses of judgment. You think you’re getting on fine and GPT2 is beginning to understand you, and then it flips personality,  starts talking rubbish, and even worse, talking about itself. The parallels between art and human life were terrifyingly obvious. I have to state here and now, I have never used AI to write any of my speeches, but carried out extensive research on what amounts to an existential threat to my business. 

So, I got stuck in, fed ChatGPT some info and waited to see what it produced. I watched the black dot cursor flash, and sat there wondering if I was about to witness the beginning of the end. So, imagine my relief when what it produced was complete and utter garbage. Even with precise instructions, it was completely incapable of writing anything that looked remotely like a decent best man speech. It simply listed the information in the order that it presented itself to the AI, despite constant requests not to, and it doesn’t understand what’s appropriate for a wedding speech. As far as off limits topics go, mulling over the smouldering ashes of your 20 year divorce in front of 150 wedding guests is about as bad as it gets. 

However, those flaws were nothing compared to the way it handles humour. Chat GPT is unbelievable at so many things: coding, programming, outlining business proposals, suggesting itineraries for vacations, but what it cannot do in any way shape or form is make people laugh. I was staggered at just what a low baseline it had when it came to humour. Wedding speeches, wherever they’re being presented, rely on a good dollop of humour; it should be original, not controversial and above all, genuinely funny. It wasn’t so much that GPT couldn’t quite achieve it, it was completely out of its skillset. Lame is the best word to describe it, and that’s being kind. 

GROWING OLDER TOGETHER

Throughout 2023 there have been various updates and incarnations to GPT, and with each one I think ‘has the time finally come?’. The capabilities of this technology is quite incredible, and I am a huge fan, but writing creative, funny speeches? Not so much. 

One of the most critical aspects of any wedding speech is length, and GPT has always struggled in that department. Even when supplied with stacks of content, anything above 800 words is a big ask for the tech, and it constantly needs reminding that it needs to do more. So, it apologises, and then produces something of exactly the same length. Don’t get me wrong, the best speeches in the world are the most efficient ones, but most speeches need to be longer than this, and you also find that it might delve too deeply into the best man, and dispatch the bride with a sentence. 

And that brings us to tone. Making sure a speech sounds authentic and relatable to the person giving it, is a huge part of my job, and I work really hard to look at who the client is and how to make it sound like them. AI doesn’t do this, and instead likes to write speeches in the style of a knight of the Round Table, which you can imagine someone delivering in a clipped BBC pre War accent, and then finishing with a flourish of their hat.

The speeches need to sound real. If you’re a plumber from Uttoexter, or indeed a recruitment consultant from New York, chances are you’re not going to be dropping the phrase ‘unambiguous and radiant love’. Exasperated with just how crap its understanding of the nuances of dialects, I did once ask it to make the speech sound more British…so it started with ‘Top of the mornin’ to ya!’ Enough said.

EARLY MIDLIFE CRISIS

In no way am I writing off the capabilities of what AI can, and will do, but I am a firm believer in that there are some things that humans will always do better, and for me the main one is humour. I come back to this again because with the advent of GPT 3.5 and GPT4, you might have expected the tech to begin to understand how to make people laugh, but no, it’s still woeful. Now, you could argue that humour is a very subjective thing, but I disagree: something is either funny or it’s not, and dropping unfunny lines into a wedding speech is, I believe, a crime against nature. Nothing will live longer in the memory than delivering something the robots told you was funny, only to be met by a sea of blank faces. 

In my spare time I have experimented endlessly with GPT trying to see if it could come up with anything funny, and it’’s failed every single time. If you have genuinely laughed at anything it’s created, then that says more about your sense of humour than it does about the failings of the technology. In years to come you might be able to train it so it begins to understand how you view things, but the dataset for that would have to be unbelievably vast, and that tech does not exist yet.

People not associated or interested in the capabilities of AI, have this belief that you can train it, and all it takes is a few boffins and a load of data to produce a digital version of you. That is of course, bollocks. 

Training a GPT model is referred to as ‘fine tuning’, and this capability has been available for each iteration of GPT. When a new version is introduced, it will then eventually be phased out. Even with fine tuning, the early models had limited scope for producing what you wanted, and by limited, I really mean they couldn't do it. Trying to teach tech how to programme HTML and input it into the header of a website is doable, ask it to be funnier, and it could be a long wait. The latest version, Chat GPT 4 has fine tuning, but it’s still only in the experimental stage and only available to select developers. The only way to ‘fine tune’ it at the moment, is by prompt engineering, which is such an inexact science, it makes Pooh Sticks look like Formula 1. It requires data guys to spend countless hours telling it how to behave, only for GPT4 to solemnly agree and promise to follow guidelines, and then do the complete opposite. Anyone telling you they have trained a model using GPT4 either doesn’t understand the science, or is being flexible with the truth. My belief is that there would be virtually no difference between a response that has been so called ‘trained’ and just sticking your information into CHAT GPT4. The end result is a largely unusable muddle of cliche.

THE END OF THE AFFAIR

So, after spending a lot of time in 2023 wondering whether OpenAi was spelling the end of my creative business, I end it feeling slightly euphoric. It can’t write in an original and creative way, it struggles to write an adequate length, and humour is such an abstract concept for it, you’re better off asking your dog. So, no I will not be joining the race to the bottom which will be substandard organizations offering to let their tech model write your speech. I’m not anti Ai, and I love the progress of technology, in fact the only people I feel sorry for are the ‘creatives’ who think that ChatGPT made them laugh. This Christmas I would like you to spare a thought for them, and we all hope they get the help and support they need.