Posted by Years ago studying AI at Aberdeen University, a professor – very respected, very senior – engaged me in conversation at a cheese and wine social. I remember the grapes artfully positioned at the edge of the cheeseboards.
This would have made the EPFL professor (le chef du labo) I met four years later, faint. He thought us PhD students were all peasants, albeit international, with our lack of appreciation for Swiss wine, so he set up regular tasting sessions to educate us. One time he saw me eating an apple just before we had our scheduled 11am degustation and almost rugby-tackled me to the ground to get the offending item out of my mouth. Then he declared a state of emergency until someone produced some bread and crackers for me to regulate my palate ready for wine-tasting. He would have seen the grapes on the table in Aberdeen as a symbol of our Northern savagery in a Swiss stereotype of British culinary habits.
Anyway, I digress! So, there I was in the common room at Aberdeen Uni trying to get to the Stilton and grapes (nom du pipe) when some auld gadgie made a beeline to chat to me, which happened a lot, being as I was, one of very few women in that world, and gorgeous young person. After asking me where I was from, he launched into some very small small talk (alas not in the programming language of Smalltalk) about high car theft rates in Middlesbrough town centre. Of all the things he could have said, he chose that.
Writing this now I want to describe him edging to the window to check on his car, in case he thought I might have nicked it on my way in, but that would be me taking some poetic licence with the story. It also reminds me of the scene in Smoggie Queens, Season 2, in which wholesome Stuart tries to comfort caustic Dickie, who shouts: Get off me you greasy pervert, which is so uncalled for and so funny, everytime I think of it, it makes me laugh.
What is true in the story I am telling, is that I just had to nod and smile like he wasn’t being really rude and stereotyping me, whilst seething inside and wishing he would get out of my way so I could get to the cheeseboard and place some grapes on my plate in an uncouth British culinary performance that would have shocked mon chef du futur!
Social identities and stereotypes
But, this is what us humans do, don’t we? We all have social identities and social labels and we judge each other on them, whether we think we do or not. To him, Middlesbrough meant car theft. To me, rude old man mansplaining who I was, when I was just going about my life, was an everyday occurrence. Something I have blogged about lots of times, especially in the Women series, (2017). The press in particular uses these labels all the time as a stereotype shorthand which is why I have always found the media to be so old-fashioned.
However, I am confident things are changing.
Last month, when I gave a public talk at Pints of Knowledge, a younger male journalist asked a question (there were a couple of older looking male ‘hacks’ in media spectacles there too, downing pints and looking sour – see that specific TV journo-hack stereotype I sketched for you) following up on Artificial General Intelligence (AGI). I had pointed out how ‘AI founder’ Marvin Minsky said in 1970 something like: AI will soon have general intelligence is just like today’s predictions in the press. I even showed an article with very similar quotes from 2025. This young journalist refreshingly and openly asked: What should the press be doing? I thought that was super cool and thoughtful.
I just went back to read the blog that I linked above [Prejudice: The social animal on social media (7), (2016)] about social labels and stereotypes. In it, I quote research that states that thanks to advertising, women are still taught to feel that they are less important than men, and behave in silly ways. A decade on and it still stands. Recently, I watched Ride or Die, with Hannah Waddingham and Octavia Spencer on Amazon Prime. The adverts, which were very frequent, were either sanitary products or cleaning and laundry products with women who looked ecstatic with their clean bowls, sheets and knickers. I don’t know, were they sniffing the chemicals? And in another blog, which I found when looking for the correct reference to social labels, Feeding the machine, (2015), I say again that we are trained from birth to give people the behaviour they want or the information they prefer, so that we get along! Like me and the auld prof gadgie.
Marvin Minsky, whom I have been very conflicted about referencing ever since he was named in Virginia Giuffre’s posthumously published memoir Nobody’s Girl (2025) (though the publisher made sure not to name the apparently well-known politician who raped and strangled her, since he is still alive), loved frames and stereotypes. Minsky suggested that we could solve AI knowledge modelling by having stereotypes and frames of answers.
In some ways that’s exactly what the LLM does. The LLM (large language model) which is the underlying technology of modern generative AI, the first public and most famous of which is ChatGPT, has read the whole WWW with lots of stereotypical, formulaic and socially trained responses, with the result that:
The LLM as the biggest stereotyping engine of all
And, now I have connected the two concepts together in my brain, I cannot disconnect them. At first I thought I must have been sniffing too much washing powder but the more I think about the more I am right:
The LLM actually produces a Minsky-style stereotypical representation of any given solution, based on what it has scraped and processed, and confidently presents it as fact. I mean don’t get me wrong, I think GenAI (generative AI) is fantastic but in the simplest of terms, to (ha) stereotype its stereotypical answer of the most probabilistic and statistically speaking ‘correct’ answer (I love it when ma blog goes meta and not in the Facebook way), it produces the average of what it has read, moreorless.
In GenAI speak, the result is the modal average or the prototypical solution, and in oldAI (pre-GenAI approaches mainly symbolic), it could be seen as an operational frame.
However, the difference is this:
With oldAI, we would know exactly what data, what representation and what reasoning it used.
With GenAI, we have no idea.
And I just don’t see why this has to be the way.
I mean it is petabytes of data we are talking about. But we could have a list – someone must have a list somewhere of where the LLM went scraping (aka reading everything on the WWW) at the most basic level. Or perhaps, could we have a list of where it didn’t go. Or a list of all the awful things it labelled up as ‘unspeakable’? It’s not unknowable, if it was we wouldn’t have reinforcement learning by human feedback (employees label up data so that the LLM knows not-what to say) nor the human-in-the-loop stage (employees check conversational output that the LLM has with users to double-check what it is saying). So, it follows that there are a lot of ways to make LLM more transparent.
This does not currently happen because companies fear being sued for copyright theft and losing the competitive advantage – something I said a while back referencing myself even earlier, saying: Don’t be fooled, it’s all just business! And in a blog, earlier this year, I talked about how Google and OpenAI wanting to sue for intellectual property theft of their weights.
If you can steal it, it’s obviously not untraceable!
Explainable ‘detective’ XAI and mechanistic myths keeping the mystery alive
Ways of trying to explain what knowledge and reasoning GenAI is doing currently focuses on doing so, after the fact. Explainable AI (XAI), is like a detective reconstructing what happened, to give a plausible account. We also have mechanistic interpretability, which tries to construct the path through the ‘neurons’ of an LLM to see how it reasoned, which puts me in mind of an article I read in the laidee Woman section of the Guardian not too long ago saying: Don’t go to the toilet in front of your husband, you have to keep the mystery alive. (FFS!)
Explainable AI to me feels like trying to keep the mystery alive which puts me in mind of the mathematicians wanting to use AI to explain mathematics. In the blog I’m linking here, I compare this approach to Ancient Greeks standing around the Oracle of Delphi, trying to interpret the messages.
And, this is not me trying to say it’s good or bad. I’m just saying it’s something else. It gives us something else, which is what AI does and it would do it even better if it represented us better. After all AI for me and it’s role should be as a collaborative tool.
Ironically, Minsky, later on after his frames and stereotypes, proposed in his book, Society of mind, that the mind is built from a vast network of tiny, non-intelligent, specialized sub-agents. This sounds rather like all the millions of neurons in the LLM. This is so cheeky really given that he was the biggest critic of the Rosenblatt’s Perceptron, the first neural network with its one neuron to rule them. Minsky even wrote a whole book on how right he was about how bad the perceptron was and that it didn’t work. That’s how certain he was about his opinion, which is a bit rich given that he went on to describe a similar – theoretical not practical – architecture, and outputs, anticipating current LLM outputs. Another time, he dismissed CYC, a 35-year project by Doug Lenat building common sense, as a failure, even though it too uses neurons and rules – frames, stereotypes and the whole shebang. It is easy to criticise, much harder to create! But, I guess the reason they are all converge at some point is that there are very few ways computationally to represent reasoning and knowledge.
AI digressions aside (or the boring techie bits of my blogs as some of my readers describe them, which are really here for me to flick my unfettered curls and say: God, I am good!) the question I am asking today in this brave new world of GenAI, as it is never a chat about AI unless I’m quoting Shakespeare, is this:
Who should say how truth, reasoning and information gets represented?
In my last two blogs I wrote about me being represented with a middle-aged makeover. The reason I won said makeover was that I gave media shorthand answers that satisfied social labels and stereotypes. At the makeover, I got turned into someone else’s idea of how a middle-aged woman should look to satisfy the clothes designer. And, six months after that, I had my voice represented when a journalist decided to write as if they were me, which wasn’t as bad as I thought it might be, though the journalist was really lovely about my, ahem, extensive feedback.
Even now, I feel slightly churlish saying anything other than that I was absolutely thrilled by both opportunities – again another expected pattern and our social psychology and conditioning kicking in. It was thrilling to get all that attention, I was just wary of the lossy compression, to coin an AI phrase – when everything gets reduced down into one take of who I am, I guess I was hoping it would a movie version of me, better looking with more articulate dialogue!
Like all humans, I am multi-faceted, something I blogged about in a Rhumba of Ruths and which Ruth I am on a given day depends on the people, places, things as John Betjeman would describe it, whom I have quoted before and a phrase that more and more, for me, sums up human-computer interaction (HCI) and AI, and now referred to as human-centred AI. Finally! It feels like that world is catching up with me.
It’s only now after having written about my desire to be styled like a Smoggie Queen and thought about where I am going with my Demystifying AI journey of public talks and courses, alongside my unending HCI obsession, that I now realise, that I have always felt very strongly about representation. And, let’s not get me started about being a woman in STEM and how technology is a mirror for society.
I have thought about it for years and years and years. I have lived it.
And seriously, why wouldn’t I? Every time I tell someone that I’m from Middlesbrough, they have something rude and stereotypical to say to me, based on something they have heard, instead of seeing me as a complete person who should be seen and heard for who I am. The only time this disappeared was when I lived in Switzerland and I got a stereotype upgrade. And, the time I represented Switzerland for technological innovation, everyone thought I spoke excellent English, which was very reassuring, and I was treated like a visiting dignitary.
Lego Ruth as a lossy compression
I chose for the picture at the top of this blog, the LEGO version of myself, which I got for my birthday this year. It was a gift and a lovely surprise. Lego Ruth is even wearing my CV t-shirt from my personal merch. She is great!
She is in AI terms, a lossy compression – lots of data and key facts about me, representing me, have been squashed down into three pieces, not as a stereotype more an archetype. By that I mean not some social label or shorthand like the Prima, but more recognisably me, the one and only, the archetype. The they broke the mould after they made me type. Or, a version of me in which I have dressed myself – I often wear a black t-shirt (and once the weather cools down I will be wearing my new CV t-shirt a lot) and jeans.
The Prima middle-aged make-over version of me which I talked about in the last Smoggie Queen blog is the prototypical representation (stereotype) of me – an LLM version of a middle-aged woman. I was made-over to fit in average clothes, average tidy hair – straight with a bit of a wave – and average woman in (sorry) forgettable clothing. It was not me. No personalisation, nor representation, no agency of me remains, it was all smoothed out to present a demographic and illustrates perfectly, the old design adage:
When you design for everyone, you design for no one.
And, that high resolution image of me does not represent high fidelity. It smoothed me right out to represent someone else completely – better than GenAI would!
When I used GenAI (in this case Google Gemini) to create an actual Smoggie Queen version of me, Drag Ruth, using what I think of now as my AI picture.
My AI picture is an actual photo of me before I left to give an in-person When computers were women talk in March and had my revelation about stepping into my digital heritage. I borrowed my daughter’s top as I wanted to wear the Prima photoshoot weightlifting belt and didn’t have anything to go with it. Now I use that photo each time to create AI pics of me to demonstrate all the GenAI principles I need to discuss from being an AIrrrrrrrr yourself, pirate to an Elven queen (Christ, that is a scary picture) and what the ‘average’ of AI pictures create complete with Hollywood Teeth. This is why I call it my AI picture.
So, using my AI picture to create Drag Ruth, created a representation which was more honest. It is not a likeness per say but it is a representation of my Smoggieness, aka Smoggie Queen persona or attitude of Style me! I’m game. The things the makeover system tried to smooth away.
So the versions have done this:
Makeover Ruth: You are this, middle-aged lady.
LEGO Ruth: These pieces were chosen as they represent you.
Drag Ruth: I’ll choose what to exaggerate.
My mission which I choose to accept
I often use Tom Cruise in my talks to illustrate why some AI ideas capture our imagination, and why the AI in Mission Impossible did not. In a nutshell it’s all about Jungian Psychology! So Mission Impossible style, I want to ask and answer:
What happens to those people who don’t understand AI but now become forced to use it? How can they make sure they are represented by AI makeover data? So that AI supports them instead of telling them who they are.
Because maybe that is the thing I have been circling around all these years.
Representation isn’t something that should simply be done to us.
It is something we should have a say in.
From being reduced to car thief and modal middle-aged woman to having my words turned into a version of recognisable not-me, leads straight to LEGO Ruth, a low resolution but somehow the truest version of me. Someone else may have chosen the pieces but because they know me, she feels like a better representation of me.
And Drag Ruth? Well, Drag Ruth is what happens when I take the Lego bricks and stick some glitter on them to create some glamour – the glamour I was hoping for at the makeover. I wanted to see a more stylish, magazine-ready version of me. Drag Ruth is Smoggie Queen me, and you know what I might have loved to have been made over like that rather than what I got. She is also a homage to Dolly Parton! I love Dolly. And, also a nod to the many Boro people I have known and loved all my life who were larger than life itself, probably because they never bent themselves out of shape to fit anyone else’s ideas of who they should have been.
I want us all of us to be like that. To be our full-on and fabulous selves. We are human, we are messy, we are contradictory and AI cannot possibly represent us in all our complicated glory.
And this is what we must be asking in our increasingly AI data driven world:
Who gets to decide what survives the compression? Who gets to say: “That’s not me”? And most importantly: Who gets to build the representation in the first place?
The big questions Dr Ruth Stalker-Firth and the AI Co-Creators are setting off to answer
Come join the AI Co-Creators and let’s get to work.
Glitter and big hair not compulsory!