Redefining Work in the Age of AI– Petter Ericson

Hi. Hi everyone. So good to see so many of you here. So I'm Christine and I just want to start by sharing a little story from my working life. Just before summer, I was having a meeting, an online meeting with around 10 people from different organizations from five countries. Uh some couldn't make the meeting, so they kindly sent their AI note-taking assistance to the meeting. So there we were uh in the online meeting, 10 boxes filled with actual humans and three boxes filled with AI agents and we don't really know each other. So I want to you know break the ice a bit. So I ask everyone a question. So it's soon summer. What's your favorite flavor of ice cream? And you know the mood changes, people light up. You have childhood memories popping up. There's some alliances forming between those who love pistio and those who hate it. And where can you find the best ice cream in the city that you're from? So each and everyone shares their favorite ice cream. And then we look at the three empty boxes kind of with the agents. Should we try to ask them? Uh maybe not. So here they are our agent our AI colleagues and they are excellent noteakers but they have never experienced the joy of ice cream. So how do we work with these guys? At AI Sweden where I run we uh where I uh work we run labor market union council where we check the trends what's happening at AI in the workplace in Sweden. And from our last report uh this fall we could see that AI is really actively in the workplace in mo almost every industry in every knowledge intensive sector from finance to legal to marketing. Everyone is working either alongside AI or with AI and AI is really transforming the workplace in that way. And we may hate AI, we may love AI, but deep down we also know that we actually need AI. I was at a big international forum in RIA this uh spring and there the OECD and the World Economic Forum pointed out that there are two quite big challenges for Europe as a continent. We are old in Europe. We're an aging population. So we're going to have a large of the part of the workforce are going to retire in the next 10 to 15 years. So the poor bastards that are left behind in the workplace, they're going to need to have smart and automated tools enabled just to maintain the welfare state for the rest of us. And then we have the productivity challenge. Europe has the potential to raise our GDP with 8% if we really can take advantage of the AI productivity boost. It's just that that productivity boost is not going to happen if we have lots of professionals out there who have excellent tools but they don't know how to use them. So that's also a huge challenge for us. So whether you like AI or not, AI skills are hot. It's projected to be one of the top three most wanted skills in the labor market in the coming four to five years. But now you work with AI, right? So you have a new type of vocabulary popping up and I want to see if you're familiar with it and have you experienced it. First of all, we have the term shadow AI. So shadow AI means that you're using unapproved AI tools at at the at your job under under your radar just to get the work done. Now, I don't want to expose any of you, but please raise your hand if you know of someone who has used shadow AI in the workplace. Ah, well, you see, good. And then there's another term, AI shame. AI shame is when you're nodding along in a meeting while silently panicking inside because you have no clue what AI tool your colleague is talking about. Raise of hand if you have experienced AI shame. I work with AI. I have AI shame every day. Still can't keep up. So the truth is everyone is scrambling a bit right now and employees are struggling to keep up. Some of them feel like they are lagging behind while others are this is great give us more tools and they're shouting this to their leadership. Meanwhile you have management teams that are stuck in some kind of virtual game of email tennis watching emails bounce back and forth between IT and legal fighting over compliance and security issues. So this shift is fast and at the same time it's so slow. And under underneath all of this is a very quiet individual anxiety. Is AI getting so good that it might actually take my job. Truths are being upended every month as AI advances. We used to say that leadership and management that's purely human domains. That's too complex. And AI cannot handle that. Until recently, there was an experiment right here in Sweden in Stockholm where the AI company and labs started a small cafe and tried to put a generative AI as a leader, a manager of this place. It turned out to be quite good at it actually, not too badly in terms of arranging permits and emailing with with authorities. Uh but as one of Andablab's uh employees told me when I was uh talking to them, uh this leader also did lack a bit of human touch and tact because it kept emailing its employees about we need to order more barista milk at 2:00 a.m. or 3:00 a.m. in the morning, which wasn't really good for for work morale. So dear all, we are in for quite a ride. working life is being redefined and today I am thrilled to introduce three exceptional speakers to share their different perspectives on this shift and together we'll explore how will AI reshape decision making human work and creativity first welcome up on stage Peter hello please join us >> hello and you can hear me lovely >> let me tell me let Let me tell if I can just get the page out the audience a little bit about you. So Petra is a staff scientist at the AI policy lab which is hosted by the department of computing science at UMIO University. He has been in other words on a very long train ride to get here. Um his research focuses on the interplay between technology, society and politics. Take it away Pepe. >> Thank you. Uh yes hello everyone and yeah thank you for the introduction uh and yeah my as mentioned I'm a staff scientist at AI polic lab up in northern Sweden um and thank you to media evolution and to the organizers for inviting me and to sort of giving the introduction to the topic let's say um and as mentioned I'm an academic I'm a researcher I work at the university and as a good academic I like definitions um So I'd like to start let's see here that's the one first and then next one is definitions. Yes. Uh so I'd like to start to dissect a little bit about what we are talking about because we're talking about redefining work and we're re and in the age of AI right so we're talking about definitions and definitions of work but it's also good to define what we're talking about when we talk about AI um because that is also a bit more of my field of expertise um another tricky thing about AI is that there isn't really a clear definition uh as such and definitely not if you look at sort of more general usage but even within the field itself it is very sort of up in the air let's say what is AI what isn't AI what technology we're talking about what's the research field that we're talking about um and instead sort of when you are in the field of AI and you're actually interested in something specific you're talking about those specific things you're talking about machine learning you're talking large language models, you're talking about expert systems or neural networks or agent-based modeling or what have you. Uh because then you at least are talking about something where you're more or less sure about what is the technical uh and sort of theoretical features that you're agreeing on here. Um and AI as such artificial intelligence it's what may be called a floating signifier or an empty signifier. Um, which is it is a term which is being used for all sorts of reasons to to indicate all sorts of different things and it takes on meaning depending on sort of its discursive function in the moment rather than being uh uh a referent that that that refers to something out in the world, right? Um, and that discursive function in general when it comes to AI is funding. uh it is a very good term for getting more research funding, getting more uh company funding and so on. Uh for being more specific about what you're actually doing, less good term, but we can talk about sort of the the let's see here uh different components of the word, right? So it is artificial intelligence. So what does it mean for AI to be artificial? Well, of course, any AI system is so technical. It involves both technical components and humans. But the key part about the technical components is that they are constructed by someone. They are understandable. They have uh a material basis. Uh someone has thought through what they're doing or at least maybe they've built it thoughtlessly, but there is some thought behind building building it. uh AI is not the sort of magical consciousness expanding thing that that it tends to again maybe get used as as a discursive function. Um and sort of it is also uh the artificial of it as I said is a social technical system. All AI systems involve humans in some way. Uh it can be more or less going in and out of the system like in the self-driving cars where you have a crew of always available uh human drivers who can sort of step in and help the car coming out of tricky situations. Uh and in others other systems it is more remote. So for uh large models for example in chatbots it is mostly in the embodied labor of the technical system. it is humans that have produced the data that have uh done the reinforcement learning and ended up with some system which you're then interacting with. Uh and then we have the sort of other type of system where the AI is the kind of background of the thing where you have for example uh warehouse workers or delivery drivers who are being managed by AI as in this this coffee shop. Um but there is always this always this combination of human uh and some kind of artificial system often computational. Um and again I want to stress this nothing about AI is magical. It is often well understood why they produce the kinds of specific outputs that you do. Even though that the in this specific case it might be hard to like it might be un unintuitive or obscure or even impossible to trace the particular the particulars of why a specific input gave a specific output. So that about artificial so what about intelligence? Well, while the fortunes of AI has kind of waxed and went several times in the year, uh there is kind of a core research subject of AI uh which can be summarized as the study of human intelligence using artificial means. So that is kind of without at its at its smallest uh core during the AI winters of of the 60s and 70s and then later in in the 90s and 2000s as well there was always researchers working on AI and that was their research subject that was looking at human intelligence uh using artificial means and that is we presume that there's some intelligence that we're studying right So what is this intelligence? It is an inherent quality of human minds or human brains which is generic. There is something which can do a lot of different things. It is involved in different faculties and different skills that we have. Uh it is somewhat stable. You can be more or less intelligent is the the underlying idea here. Uh and it is then expressed through all of these other things like language, like math, like spatial reasoning. uh chess for example um pro problem solving, music, everything like that. That is the the the the idea here is that there is some core feature of human brains uh which is expressed through all of these things. Um and by sort of looking at how well you're doing in language, in chess, in math, you get some information about this underlying quality of intelligence, right? Um, and just to sort of put myself out there a little bit, I am somewhat uh a skeptic of the usefulness of this term. I know a bit too many people that are really great at one of these skills and really garbage at one of the other ones. uh that I don't think that there is very much use in trying to find this one factor that kind of supposedly explains uh some underlying one-dimensional value of how well your brain works. That's not really my impression of how brains actually sort of function in society. Um but leaving that aside, there is at least this idea that most human brains can do a certain things to a greater or less extent and that there is a link between various different capabilities. Uh and that the link sort of goes that if you're good at language, you're also good at math, right? In general as a as a on a population level, right? Um and there is kind of a risk of re reversal here um where we mistake the kind of particular capacity that we claim that in humans uh signify a broader cap capability for thought or consciousness of reasoning and so on as something that does the same for computers. So we say uh if if a person is is good at skill X that means that they are likely to also be good at skill Y right. That doesn't mean that a machine that performs well at say chess or math is something that is even capable of that other skill. We have calculators which are really great at math but they don't play chess that well and they don't do language that well. Um and moreover there is a a a risk of uh confusion where we look at something performing a skill that we usually associate with an underlying human mind and expect that mind to be there also when we see the the the sort of behavior from something else. Uh humans are great at seeing minds. We have pet rocks. people pack bond with their uh wi with their uh like pet robots and their vacuum cleaners and whatnot. We are great at pack bonding. That also goes for AI, but that tricks us because we think that uh because the AI has this particular capability that brings with it all of these other things. So artificial intelligence are not just LLMs as well. That's also an important thing to remember that the field is much much bigger than LLMs. What we talk about when we talk about introducing AI in the workplace, that's not just LLM, not just chat bots, not just generative AI. Uh it is all kinds of algorithmic systems. Um that can have different features, right? So is all these different constructed systems. Um but an important feature or an important case right now is LLM and LLM have certain common features all of them uh they are built on massive uncurable uncurable data sets. The size of an LLM uh all in order to be sort of useful useful as an LLM requires the amount of data that is impossible for you in curation to sort of go through. it is simply they they need you need that much data and so you will need to get for example all of the internet or all of whatever uh and that means that all of that data is then compressed in your LLM so you have all of that there right um LLMs are also very inefficient as mechanisms uh calculating 1 plus 1 in an LM uses a lot more data lot more energy and and and resources is than just typing it on a calculator. Same for almost all tasks. I say almost because large language models are models of language. So for tasks where they act as language models, that is predicting the next token. It's very good. We're good at that. But for many other tasks, they are highly inefficient. Um and they are also inherently because of their sort of probabilistic nature uh quite unreliable for those tasks because they are predicting the next to the next token and there is no real way of constraining them to uh only keeping to uh some particular uh set of outputs, right? Um or rather you can do that but then you have built a system that could output the thing itself, right? There's some some some ways you can do that, but then you need more assistant than that. And they perform linguistic fluency. Um that is they seem to have linguistic fluency which again is this one of these core skills that we associate with underlying cognitive capacity. Um, and there is kind of a wrinkle in this, right? Because LM as models of language as predicting the next token, that's not the typical use case anymore. Instead, we are piggybacking on the the uh the language facility, right, of what is the next token uh to perform all of these other tasks like addition or like search or like summarization or uh translation or whatever, right? Uh we're piggybacking on the language next token prediction in order to perform all these other tasks. uh but they are still working. The the what what the language model itself is still doing is precisely uh predicting the next token. That is all it does, right? Um so uh so that is kind of what AI is. And does AI actually perform work? Well, what is work? Well, let's that that is also a fraud term, right? What is what is work? Uh and it is also a ter term that has multiple diff different and competing definitions that are I would say is more well- definfined than AI, but still some uh some work in there. Um but let's see here. Oh, you missed one. Um, three characteristics of work in general is that it something that results in output. It performs something. It does something that takes effort and what isn't there is because I sent the slides too late is that it's usually compensated is an economic activity is usually work. Um, and work is something that takes work. uh something that a human does by spending energy and time and other resources to achieve and it is compensated. So what does redefining work in the age of AI actually mean? Well, what when is the age of AI? Is that now? Are we talking about what AI is doing in the workplace right now? Uh well, in that case we have all of these different uh like current uh impacts of AI as in uh replacing managers for example. uh as in uh helping with some tasks and really really hindering others uh impacting with spam and so on. But then there is also the question of maybe we're not actually uh in the age of AI quite yet. It can improve the we can have growth of uh what you say 8% over however many years. Uh that's in the future, right? And how does that work? what what kind of work would that AI do, right? How does that fit into uh this this idea of work as taking effort for example and and doing uh uh things um so I'd like to sort of uh quote Corey doctor a little bit or at least reference him a bit uh as because he takes from automation the theory and talks about two groups of workers interacting with AI with AI those who really really like it in and sort of use it to speed up tasks and whatnot. And then you have uh the ones that are sort of subject to AI and are seeing their uh work cues being overloaded and so on. Uh what what he called the these two groups are Kentars and reverse Kentars. So the the Kentar is the human mind on a machine body. So someone who can use tools, decides over tools to run faster, longer for and better. Whereas the reverse caner is a machine head that drives the human because in a human machine system in that configuration the human is the weak link and the machine will try to drive the human as fast as possible as hard as possible. And that is something that uh goes for AI but also in general for for for automation over in factors and so on. So let's go back to sort of what is human work. Um because human work is never just completing tasks, right? You can get as many tasks as you want. Uh but just completing the tasks is not uh the work, right? Because you uh work is also a social activity where people interact with their uh colleagues and customers and other people. You build up these social relationships in the workplace and outside of it. Uh you develop skills by doing things. You learn more about the systems you interact with. Um and humans in the workplace also have some level maybe not a lot but some level of autonomy and moral accountability. You are responsible for the things you do at work and you're also responsible for keeping keeping an eye on your workmates to make to hold them accountable and their uh maybe keeping track of what is this decision. Does that really make sense? Okay. So for to to take one example uh on the from Goththingberg the placement of children that was a big scandal a number of years ago uh a human worker involved in or human workers involved in that process did raise the alarm and say that hello these kids in Maya maybe they shouldn't go to school on the other side of the river that doesn't make sense and so humans involved in work processes have these safeguards these independent minds that can work um through the uh through the work process. So how are we talking about redefining work then in the age of AI? Well uh I have sort of skipped a number of things so I'm sorry about that. Uh things took longer than I expected. uh but basically when you're talking about automation and AI in the work in the workplace uh it is always the case that the people who are actually doing the work are the one that knows the area best. They have minds, they have eyes, they have senses to know what is actually going on in the workplace that a manager uh might not have. Um when you're automating a task, that means that there's no longer a human involved in that task that can learn from it, right? Uh automation stops learning. Uh the workplace is not just a place where uh tasks get done. A workplace is much more than that. It is a social and relational space where uh people are learning from each other. people are communicating communicating with with each other and people are accountable to each other. Um work is always more than tasks which is important to keep in mind when you're introducing AI and AI lastly is again not magic. It is a specific constructed tool that we can understand and see how it was made and what it does. So I think that's the last one. Thank you.