On Wednesday Media Evolution in Malmö hosted a seminar on the role of AI in reshaping decision-making, work, and creativity, where AIPL Staff Scientist Petter Ericson was invited to participate as a panel member, alongside Ebba Lindgren and Isabela Bolotti, with Kristin Heinonen of AI Sweden moderating.
The three presentations were quite contrasting in tone and content, leading to an interesting discussion afterwards, with both panelists and audience members sharing practical and concrete examples of how AI impacts their working situation. The event was filmed, and recordings will be added to this post once they are available, but in the meantime, see below for the remarks given by Petter, and see the attached file for the slide deck.
Hi everyone, my name is Petter Ericson, and as mentioned I’m a Staff Scientist at the AI Policy up in Umeå. Thank you for inviting me, and thank you for giving me the honor of making, as it were, introductions to the topic of today’s seminar.
As a good academic, then, I’d like to start talking definitions, not just because shared understandings of what words mean is a necessity for productive discussions, but also because the topic of today in part rides on redefining work, so it’s a good idea to have an idea of what was and is meant by work previously, in order to arrive at a good redefinition.
Before going into work, however, I’d like to start with dissecting what is perhaps, honestly, more my area of expertise: AI.
Now, the tricky thing is that there isn’t really a clear definition of AI. Definitely not if you look at general usage, but even the field of AI itself struggles to find definitions that are useful even for specific subfields. Instead, when specifics and details are important, most discussions and discourse use other terms – machine learning, expert systems, large language models, neural networks – that at least can be agreed to share some underlying concepts and architectures. AI, in contrast, is what is called an empty or floating signifier, taking on meaning mostly in what the particular discursive function is of invoking the term in the moment.
That function tends to be funding, by the way.
So let’s talk a little about the component words of AI: Artificial Intelligence.
What does it mean for AI to be artificial? Well, any AI system is composed of some combination of artificial, automatic (usually computer-based) components, and human labour. In some cases, like in “self-driving” cars, the human labour is a crew of always-available drivers that can take over the instruments at any time and navigate the car out of tricky situations, in others it is more remote, say as the embodied labour in the construction and curation of the of the underlying datasets, and in some it is more front-loaded, as in cases of algorithmically managed delivery drivers or warehouse workers. The crucial thing to remember about the automatic and artificial parts of the system is that they are just that – artificial, constructed, and understandable. Nothing about AI is magical, and it is in general well understood why they produce the kinds of outputs that they do, even if it is occasionally unintuitive, obscure, or even impossible to trace in a particular case.
Next: intelligence. While the fortunes of AI has waxed and waned several times over the years, there is a core research subject in there, which can be summarized as “the study of human intelligence, using artificial means”. That is, we presume there is some general faculty of the human brain called ‘intelligence’, and what AI is, is making models of (aspects of) that faculty in order to better understand it. Now, as an aside, I am somewhat sceptical as to whether the term ‘intelligence’ is really very useful here, since it tends to bring to mind a generic, inherent and somewhat stable faculty of the brain which is then expressed through various skills, such as language skill, mathematics, music, problem solving, spatial navigation etc, and I’ve met too many people who are great at one thing and not great at others, or whose learning process has changed rapidly due to external circumstances, to really think that this kind of thinking is useful, but leaving that aside, there is at least this idea that most human brains can do certain things to a greater or lesser extent, and that there is a link between various different capabilities – being good at math means you’re likely to also be good at reading and writing.
There is the risk of a reversal, however, where we mistake the particular capacity that we claim (in humans) signify a broader capability for thought, consciousness, reasoning, and so on, as something that does the same for computers. Being good at chess supposedly means that a human is “intelligent”. Now computers are good or better at chess than humans are, thus it means that computers are intelligent! This is something that has happened several times over the lifetime of AI as a field, where some promising result in a particular area (e.g. reasoning, image recognition, etc.) is claimed to imply imminent progress in many other areas where AI research is active. This reversal is particularly powerful and convincing when the specific capacity that is being simulated is that of written language, since so much of what we do, especially in the modern, digital, internetified society, works through that particular medium.
Circling back to the term AI as actually used, let’s again check in with what we are actually talking about: What products, what services, and what technologies are we referring to when we are talking about AI? Well, for the most part, nowadays it refers to large language models and systems built from them, such as chatbots, various types of media generation systems, etc. Some key characteristics that I think is worth noting: 1) Massive copyright infringement and data collection 2) Massive environmental impact and inefficiency 3) Inherent unreliability 4) Linguistic fluency. To be clear, this is not all that is called AI (recall that AI is an empty signifier), but there is usually some measure of machine learning involved these days. And as mentioned, linguistic facility, in todays world, with the focus on the written word and the large repository of writing, counts for a lot.
Does it count for all of work, though?
Again, we should begin by talking definitions and usages of ‘work’ and ‘labour’. Again these are terms that have no single clear agreed definition. However, I think we can identify a few key concepts; First is output: we say that work results in something, whether that is a physical thing, or some service that results in some kind of change in things, healthcare, education, meetings, planning, sales, etc. etc. Second is effort: Work is something that takes work, something that a human does by spending energy, time, and other resources to achieve. Third is compensation: Work is specifically that type of effortful activity that generates outputs which is compensated by a wage or other remuneration.
So what does redefining work in the age of AI actually mean?
Well, we can take two somewhat distinct viewpoints here: On the one hand, we can look at the current landscape and how AI and automation has been introduced into the workplace, and how it has affected people’s working situations, i.e. we take the idea seriously that we are now in the age of AI. While it is hard to truly determine the impact, and many reported large-scale benefits and effects are highly questionable, we can identify a major divergence in how automation and the introduction of AI is written about by people subject to it: For some, there are concrete benefits to specific parts of their job, where automation of some particular part makes them moderately more productive at particular tasks. For others, the main impact of AI is an intensification and degradation of the working situation, where for example translators and designers might be delegated to “editing” AI outputs instead of producing their own work, and spending as much or more work on fixing those outputs, but under worse conditions and remuneration. Cory Doctorow, drawing from automation theory, identifies these two groups as “centaurs” and “reverse centaurs”, respectively: In short, is the human person driving the stronger, faster machine, or is the person being driven by the machine? Regardless of any wider impacts, it is clearly preferrable to be the one in the drivers’ seat in these situations. However, the question of ‘redefining’ work seems premature in these situations. For both centaurs and reverse centaurs, their work is clearly impacted by AI and automation, and what they do at work and how they do it (and how well they are paid) has changed, and will change further.
The other view is to look at the future of work. Here, it is instructive to look at various projections from companies such as McKinsey and Goldman Sachs, and look at what they have to say. The most well-known and well-report projections for the impact of AI on work have some.. interesting methodologies, let’s say. In particular, what tends to happen for these broader analyses is that workers are categorised into roles, and each role at work is defined as being composed out of a certain number of tasks. These tasks are then estimated in some way in how likely they are to be automatable using AI within the near (or medium, or far) future, and this feeds back into each role where, say, more than 50% or 70% of the tasks are likely to be automatable being classified as a role likely to disappear. As an aside, some of these estimations of automatability of tasks have been achieved by simply asking ChatGPT.
Now, here is where I would like to bring back what I said about intelligence earlier: A fundamental misunderstanding often made in and around AI is a confusion about some very particular output or display of presumed skill being a sign of a much broader underlying competence. A convincing linguistic output is presumed to indicate a whole person capable of all the other types of things we would typically associate with that type of behaviour. There is a similar sleight-of-hand at play with these categorisations of work as a collection of tasks: The claim is that work (or intelligence) can be completely, or at least sufficiently, broken down to a particular set of well-defined input-output mappings, and that a machine being able to automatically match inputs to outputs means that the rest of the process is also within its grasp. This is, however, not the case.
There is no workplace, no labour done by humans, which can be sufficiently described by a “very small shell script”, as some computer professionals used to put it back in the day, but realising why requires removing the abstractions and putting the task in its proper context. Humans doing work are never just doing the task, but are learning, growing, and reacting to all the other things happening at the same time. Writing code means recalling and learning about the codebase and the various libraries used, and seeing how they fit together, gaining a better understanding of the system. Researching and writing a report means internalising and evaluating all the research found, the texts written and their authors and networks, and relating them to each other, which not only means the researcher is learning about the subject of the text, but also of all sorts of tangential things that may be relevant for a different piece of writing and research months or years later. The moment a task or practise is automated, it is fixed into place, and no human learning and understanding will grow. This is as true for schoolchildren learning arithmetic with a calculator as it is for a researcher or writer using a text generator to synthesise ideas.
Moreover, at the same time, the question of ethics, accountability, and recourse gets immediately more complex. The more people are involved in making a decision, they more people are available to evaluate whether that decision is actually reasonable or not, and the more clear questions of accountability, responsibility, and power, have the potential to be. I say potential, because of course there are dysfunctional bureaucracies and muddy questions of accountability (and complete lack thereof) even in fully human systems.
To be clear, there are places where AI and automation is the correct approach, but it needs to be a step taken with open eyes and clear motivations, and with the understanding of what is being done.
So, to check back in with the topic of the day: Redefining work in the age of AI. Are we any clearer, if not on the answers, then at least the questions? Or at least some of the definitions? The need for definitions? Well, I suppose I will leave that up to you, and thank you all for listening.
