#259 What Microsoft’s Misunderstood Copilot Study Actually Means for the Language Industry

0:00Florian:This most misunderstood research paper titled Working with AI, Measuring the Occupational Implications of Generative AI.

0:14Florian:Hey everyone and welcome to SlaterPod today with a special episode on a piece of research, a research paper by Microsoft. And who better to talk about research papers than our very own head of research, Anna Wyndham. Hi Anna. Hi Florian. Hi everyone. All right. So today we want to talk about this most misunderstood research paper titled Working with AI, Measuring the Occupational Implications of Generative AI. And it was published by a group of researchers from Microsoft. It got a ton of attention. It got even more wrong hot takes. You got like, you know, X posts with like 5 ,000 reposts and things like that.

0:54Florian:Tons of traction on LinkedIn as well. And we covered it both on the website, but also in our Friday newsletter. And then we got some great feedback from readers. One was, let me quote here, Anna. One was just an email back. It was like, some engineer at Microsoft speaks and you guys dutifully amplify it? Question mark. Well, keep scanning the research papers for more self-serving prognostications. All right. So, yeah. Thank you. You can always unsubscribe. It's a free newsletter. So, you know, if you don't like it, feel free to unsubscribe. Well, point being, there was a lot of emotion. There was a lot of kind of also wrong views on this particular research paper.

1:41Florian:But it's still very, very interesting. So, let's unpack it. Anna, I'll hand over to you, unpack what's this paper about, and then let's try to go through it.

1:53Anna Wyndham:Yeah, as you said, it's got a really strong reaction. So the purpose of the paper was to measure the occupational implications of generative AI. That was the title. And just to start off, it's really interesting that Microsoft has access to all of this data about how users are actually using AI. So it's interesting, firstly, that they've chosen to study this and release it. So they're trying to understand the impact of AI on the economy and occupations generally. And they present this paper as just a small contribution to this conversation. So there's nothing kind of definitive about its conclusions.

2:35Anna Wyndham:They've just taken some data and done a few things with it. to see what they can find.

2:45Florian:The data is data that nobody else would have access to, right? Yeah. And so this is the interesting thing. Like there's lots of stuff that we can do, but we can't look at like what type of conversations people have with AI because, well, we're not running the AI.

3:00Anna Wyndham:Exactly. So they've looked at basically two sources of data. On the one hand, they're looking at how people are actually using Microsoft Copilot in the wild. And when they say in the wild, they mean for work purposes or for leisure purposes. They're not making any distinction. And they studied 200 ,000 conversations between people and the AI. And what they're trying to do is to understand what are people using AI for, how successful is the AI in helping them do the tasks, and what tasks the AI itself actually performs. So they make a distinction between the goals that the users have and the tasks that the AI performs.

3:41Anna Wyndham:And the way that they measure how successfully the tasks are executed is by looking at two things. They look at the user feedback. So they've got like a thumbs up, thumbs down option for users to indicate their satisfaction with how well the AI is performed. And their researchers also have gone through the data from the LLM itself to see if the AI task was executed successfully according to its own kind of standards. So that's the data on the one hand.

4:15Florian:And of course, when they talk about data, I think they went to great length and pointed out that it was privacy scrubbed and anonymized or something because, yeah, I mean, there's obviously a lot of very private conversations going on with the AI.

4:28Anna Wyndham:Yeah, they made that point, of course, and removed that identifying information. So that was kind of step one of the study. And the other step of the study was to look at this US jobs database called O-NET. And this is a database that lists all the possible occupations, all the possible work activities, and associates work activities with occupation. So they view every occupation as like a basket or a bundle of work activities. And then they looked at to what extent the activities that AI is being used for and is good at overlapped with the work activities that are relevant for any particular occupation.

5:15Florian:So key findings in general, what were some of the key findings that they got from looking at these 200 ,000 conversations between people and Microsoft Copilot?

5:27Anna Wyndham:Yeah, well, perhaps unsurprisingly, it was knowledge work or knowledge-related work activities that had the highest overlap. So specifically information gathering, writing and communicating with others are the most common user goals in Copilot conversations. And they also happen to be the most successfully completed. So they're getting the most positive feedback from users and the inspections show that the AI task is being executed completely. and as you mentioned the one of the outputs of the study was this list of occupations and this is the kind of list that we're seeing on on x and on linkedin and it's got interpreters and translators right at the top at number one followed by historians also in the top 10 writers and authors customer service representatives and if you go all the way down to the bottom of the list you're getting more into the physical and manual occupations and right at the bottom of the list You've got rail track laying, motorboat operators and dredge operators.

6:24Anna Wyndham:So these are the occupations that have the least overlap with these successful and common AI tasks.

6:32Florian:All right. So let me, for those of you who are looking at this on YouTube, let me just spin up the paper here. So you see, yeah, that's what, of course, kind of triggered a lot of people in the language industry. At the top here are interpreters and translators, historians, passenger attendants. We want to speak about that in a second. Sales reps, writers and authors, customer service reps, CNC tool programmers. And they go all the way down to your dredge operators and bridge and lock tenders and water treatment plant and system operators. So yeah, you go from the information world to the physical world.

7:11Florian:So they calculated this AI applicability score. What does this actually mean? An AI applicability score. So is it a replacement score?

7:20Anna Wyndham:So the researchers actually are very careful in adding caveats to how the study should be interpreted. They say it would be a mistake to conclude that occupations that have high overlap with activities AI performs will be automated and thus experience job or wage loss. So they're explicitly saying that this result that puts interpreters and translators at the top of this overlap does not in any way indicate that one of the implications is that these jobs will be replaced. So I think some of the headlines that we've seen spinning around are about translators and interpreters being made obsolete or jobs are going to be lost.

8:04Anna Wyndham:And this is really not the case at all. And we spent quite a bit of time working on the Seda headline for covering this. And what we came up with was, co-pilot shines on tasks that language professionals know best. So this is really what the study says about translators and interpreters. It doesn't say any more than that. And the researchers are very open and explicit that it doesn't say, it doesn't imply any more than that with regards to the economy.

8:34Florian:I mean, they probably should have expected that putting out a list where, you know, like that the hot take was going to be top of the list means most likely to get replaced or wiped, you know, wiped off the face of the earth immediately. I mean, and that was the hot take. Like literally every single tweet, every single post was like, oh, okay, top of the list means bad. You're going to get replaced. Bottom of the list. Hey, you're going to, you know, dredge your, what is it? Dredgers? dredge operators are going to stay for the next 20 years. So that was the kind of obvious hot take here. So do you think that their caveat was basically knowing that this was going to be the take and they wanted to have that caveat?

9:18Or is it like a genuine limitation of the study?

9:23Anna Wyndham:I think they most probably expected it to be interpreted in this way. But I think that it would be a shame if every single attempt to understand the impact of AI on the economy kind of elicits these responses. Because there's a lot to understand and it's in our best interests, obviously, in everyone's best interest to dig into the data that we have and see what it can tell us about the impact that AI is going to have on occupations.

9:53Florian:And so, yeah, let's talk about maybe one of the first points to understand this here relating to translation interpreting and other language work. So AI is relevant and useful in translation interpreting. Yeah. What does that overlap actually tell us?

10:11Anna Wyndham:You can see this nice list on one side of the occupations and of the tasks on the other side and lines joining the tasks with the work activities. I should stop saying tasks because they made an important point, a distinction in the paper between work activities and tasks. So we're talking about work activities. And the work activities, some of the reasons that translators and interpreters rose to the top is that AI is very good at things that translators and interpreters do a lot, such as interpreting language and cultural information, presenting research or technical information, gathering info from various sources, examining material for accuracy.

10:50Anna Wyndham:So AI is good at doing those tasks or assisting with those tasks. And in a lot of cases, the majority of cases, it is assistance rather than full execution of the task.

11:03Florian:So the job does not equal a set of work activities, right? So there's this, they acknowledge that kind of decomposing a job into its work activities does not provide a full representation of every job, right? And then they had this really interesting anecdote. Maybe you want to talk a bit about that. Metaphor. Metaphor. Sorry, not anecdote.

11:27Anna Wyndham:Yeah, so the researchers acknowledged that this O-Net database, which has a lot of, it's very useful in some ways, but what it basically is saying is that if you do a job, it's basically a basket of disconnected work activities. And they acknowledge that looking at occupations in this way is of course inaccurate because it does not give a full representation of every job and the analogy they use is that the connecting glue between tasks also contributes to the value of the work so in that case what they're saying is that it's one thing to present research and technical information it's another thing to interpret language and cultural information it's another thing to be able to do both of those things and to apply that to a specific translation job at hand or a specific interpretation situation at hand.

12:19Anna Wyndham:So they acknowledge that seeing jobs in this way is useful to an extent, but it doesn't give a full picture of what any one person does, what any professional does in that occupation. The other thing as well is that if you look at that list with interpreters and translators at the top, you'll see an impact score. And I think it's about 50 or 57 for interpreters and translators. And what that indicates is that the work activities that AI can perform

12:56Anna Wyndham:only covers a moderate amount of the work activities that translators and interpreters need to perform. So even when we're looking at occupations at the very top, the data does not show that the AI can perform all of the work activities of any occupation. In fact, it shows that AI can perform some of those activities, a moderate amount. In no cases did they find that AI can perform a significant proportion or total coverage of the work activities for one occupation.

13:32Florian:I mean, trying to summarize this, like, so that is probably where the misunderstanding came from, right? That people naturally would look at the list, would look at the one line where it says, you know, interpreters and translators, historians, writers and authors, CNC tool programmers, that's one line, right? And they see the coverage score. That's how it's, is it though? Yeah. I mean, or kind of the composite score. And they say, well, this entire activity, interpreting and translation, is the one that has the highest score, i.e. it's the one that's most basically at risk of replacement. But what the researchers actually looked like, and please push back if I get this wrong, were these individual components, right?

14:15Florian:These individual work activities. And if you kind of sum all this up, no job is more than 50 % or has, what is it called? That score? The impact. The impact score, exactly. So in no job does it have more than 50 % of this impact score. So these jobs aren't single work activities. It's a composite. The job itself is a composite of individual work activities where AI is useful or does take some kind of replaces certain human activities.

14:51Anna Wyndham:Exactly. And connected in a way that the AI is not doing as well.

14:55Florian:Maybe that's kind of conceptually where all this hype around the agents is coming from, right? Exactly, yeah. Maybe in different research or just in reality, people are noticing, well, the AI is great at XYZ, but not at these things that humans do because they're the glue. They're kind of gluing together all these activities using their powerful non-artificial brains. And that's why we need these agents. and these agents is kind of the theme of 2025. And there's pros and cons and people are saying it's a hype and other people are saying it's the next greatest thing, right? All right. So let's talk about business application, the business, I don't know, applicability or the business relevance of this.

15:44Florian:Do you have any thoughts on that?

15:45Anna Wyndham:The researchers mentioned that, of course, that just because we can see that AI is useful in certain contexts does not imply anything about how businesses will choose to use AI to either automate or augment work activities. They say it's tempting to conclude that occupations that have high overlap with activities AI performs will be automated, experience job or wage loss, we mentioned that. And they say that occupations with activities AI assist with will be augmented and raise wages. This would be a mistake as our data do not include the downstream business impacts of new technology. and I think I mean what we look at at SATA every day is the downstream business impacts of this new technology and that's what we've been looking at for the last say decade the impact of machine translation on the translation industry so we already know quite a lot about the business impacts of machine translation on the on the language industry and this is something that is not necessarily new and it's not something that we can take anything from the study.

16:51Florian:Yeah, any kind of extra limitations, considerations mentioned in the paper or not mentioned in the paper rather?

16:58Anna Wyndham:Something that I thought that the researchers didn't really consider is that the extent to which AI is embedded in the work and work activity environment of knowledge workers compared to, say, non-knowledge workers. So MS Copilot is embedded into Word, Excel. So, of course, these are tools that knowledge workers use every day. It's not embedded into tools that dredge operators and boat operators and chefs use every day. Maybe one day it will be. But, of course, the usage is going to be higher and the success is going to be higher in cases where it's embedded. And the other thing I was wondering is, What about the design and user interface of Copilot?

17:42Anna Wyndham:Is that influencing usage? So when you're in Word, for example, to what extent is Copilot proactively suggesting certain tasks be completed with AI? Would you like me to translate this? Would you like me to summarize this? It's not clear from the paper if that was influencing how, if that was taken into account in terms of how users are using.

18:06Florian:No, we wouldn't know because we're on Google Workspace. We were, yeah. I'm not sure if I ever used MS Copilot. Yeah. So, yeah, the conclusion there would be.

18:17Anna Wyndham:For anyone who doesn't like this idea of this AI applicability score being applied to translators and interpreters, you could think of it as it's more of an MS Copilot applicability score. If you want to be really literal about it, this is the limitation. It tells you how useful and relevant is MS Copilot to the activities that interpreters and translators do.

18:40Florian:So, yeah, so we started with, this is a very misunderstood paper, but also, I mean, if you really boil it down, like it doesn't tell us anything massively new. It just kind of confirms a lot of the kind of existing narrative, I guess.