Transcripts are auto-generated and may contain errors.
0:00
Hey, I’m Rose.
And I’m Angie.
And this is Our Lives with Bots, the show where we ask important, timely questions about what it means to live with our bot counterparts.
And from time to time, we also dive deep into what an AI future might look like for us.
0:15
Sometimes we agree, sometimes we spiral, but we always go deep.
Welcome back, listeners, to another series of our lives with bots.
We’ve covered companion chat bots and the impact of AI on children and young people.
And this time around for Series 3, we’re going to be talking about the impact of AI on the workforce and on our relationship with work.
0:38
We have a lot to cover in this series.
You get the sense when you just feel like the hype that loads of people are adopting, it’s going to displace everyone’s jobs.
We’re going to be talking about whether or not it’s actually going to be taking everyone’s jobs.
But I think we’re going to start off by seeing what is the actual adoption of AI, who’s using it?
0:54
What are they using it for?
Then talking about displacement.
So is AI really out to take all of our jobs and whose jobs are really at risk?
And what does that mean?
In later episodes, we might hear from of our chat box friends again.
We’ve got a whole lot of interesting stuff in terms of some fun news that we’re going to share.
1:10
We’re going to get to that today actually.
And there’s a lot of different perspectives about this AI in the workplace.
So are we losing our skills?
So are we deskilling?
Are we not building skills?
And what does that even mean?
And so of course, within the series, we’re going to be talking about the psychology of our interactions with AI and the ethical implications of it.
1:28
So thinking about things like deskilling, automation bias, cognitive offloading, all of these processes that are naturally human, but with AI, it might have some additional things that we need to consider to start off before we go into some metrics.
We have some fun facts for you.
1:44
As promised, in the inter episode of our last two series we added some fun facts about us just to give you a better sense of who we are.
So, Angie, do you have any fun facts to share with the audience?
So I’m from South Africa, as people will know from my accent, but the first time I got to travel internationally was actually to the USI was very excited.
2:03
I was 15 and I got to travel to Minneapolis, which may seem strange except that the Minneapolis is very.
Exotic.
Well.
If you come from South Africa, I’ll tell you one thing about Minneapolis.
It has the Mall of America.
And when I was 15, did you not the mall, the Mall of America mall?
2:24
What?
I’m from the Midwest, maybe I should know this, but I do not know what this is.
It’s a shopping mall.
Oh, Oh yeah, even.
Translated for people that don’t understand my accent.
So it’s the shopping.
Mall.
2:40
There you go of America.
But this was such a big shopping mall that if you went to each shop for one minute and you didn’t get lost, it would have taken you 8 hours to get around the entire mall.
And for a 15 year old that came from South Africa, that was huge.
It was three stories and they had a Camp Snoopy.
2:58
They had a Snoopy that was as big as the entire mall anyway.
But more importantly, the reason I got to travel at 10 to age for the first time was because I was attending and competing in an international figure skating competition.
So that’s my fun fact.
Oh my goodness.
So what were you skating for?
3:14
What was the What was the theme or the show?
So this is a bit weird so we did precision ice skating, which is 16 people on ice at the same time.
So imagine like cheerleading or something like that, but we don’t get thrown up into the air.
Nothing weird like that, but 16 people and you’re doing like wheels and formations and stuff to music for about 5 minutes.
3:34
And so we we were competing with other teams all doing the same.
It was called the That’s Amazing Competition.
Snowflake competition.
How apt.
I can’t even imagine the amount of coordination and the added layer of what if someone falls down or is in slightly the wrong position.
3:50
It’s more just like there’s a lot of speed that goes on in this and a lot of precision.
That’s why they call it precision, because it’s very much the formation, because you’ve got 16 people all needing to coordinate and do things and there’s splashes and it’s very, very cool.
I admire figure skater so much.
4:06
My favorite parts of the Olympics is figure skating and gymnastics, and that’s partially because I can’t even imagine the amount of coordination and skill needed for those sports specifically that I simply do not have.
So my admiration is in part due to my lack of skill in that area.
4:23
I’ll take.
Any other sports but?
I will take you skating one day.
Yeah, I’ve never been able to, like, glide, but maybe it’s just because I’ve I’ve always skated on very uneven lake surfaces in Indiana, which is where I grew up.
Fun fact for viewers, yeah.
4:40
Tell us about yours.
So I don’t know if you’ve noticed with me flailing my hands around in every episode, but I always have long nails and I actually make my nails.
So you might be wondering, Oh my God, Rose has a new nail set every now and then.
4:56
That must be very expensive.
It is not.
I make them myself.
And so it’s very cheap and it’s also very fun because I like art.
And so that’s one of the things that I do.
Another fun fact, which is not so creative but more so work, is that I recently went to Washington, DC to talk to mostly staff members of Congress in the House and the Senate on AI and mental health.
5:18
And so this was with Princeton and their Center for Information Technology program, and they had this partnership to do briefings on particular topics.
It went well, and it came right after the executive order that was officially released by the White House on the AI Litigation Task Force.
5:35
And so one of us panelists talked about the implications of that and what policy can and can’t do with respect to mental health and AI.
Hopefully, we gave enough information for things to move forward in a positive direction.
One of the things that I mentioned were all of the lawsuits over teens committing suicide or young people committing suicide and talking about the psychological implications of our interactions or social interactions with AI.
5:59
So one person stood up to ask a question at the end of the House Office briefing and said from this, I feel terrible.
Obviously this is something that we need to do.
Something about it definitely hit.
And I know from prior workshops and things like that, that when you’re speaking to people who are in charge of policy, they are obviously inundated with so much information.
6:22
And So what they can take away from it is really the emotion of the message.
So I hope that came across.
I ended it by saying what big Tech does is move fast and break things.
But what if the things you break or people’s lives?
I hit that at the.
End well done and great to have that opportunity to speak to people.
6:38
Yeah, I was really honored to be invited to be panelists for that, and I hope that there are many more that we can do in this area to hopefully move things forward in a positive direction.
With that said, welcome to Episode 1.
Let’s get into the intro.
6:53
What we wanted to do is really give our listeners a sense of what is happening in the workforce.
But we wanted to start with is really grounding that in some statistics in some.
And even that is quite difficult to do because it’s a lot of research being done.
In fact, by the end of today, we’re going to talk to you about some that we’re going to do, which is very much desk research.
7:12
But let’s talk about an article that came out.
So a study was conducted earlier this year between July and August and they actually had a decent sample size, whereas about 7000 people.
From what 6 countries to?
The sample, six countries, yeah, Yeah.
So it was Australia, Canada, Germany, New Zealand, United Kingdom and United States.
7:30
So a really good robust sample.
I mean, we can talk, but it didn’t include the entire globe, but I mean a good, a good rates of countries and also quite nice in terms of different types of working groups.
So they had workers and managers and executives and then obviously people aged 18 and above.
7:46
And this was this was done by Dayforce, which is kind of like a human resources research group.
And so they put out this survey this year to ask how it is that these different individuals across these countries are using AI at work.
8:02
And they have, this is really interesting headline for the take away.
The headline that I was seeing was the discrepancy between AI use on the executive level versus on the worker level.
Yeah.
Big differences, right?
So 87% of executives using AI versus 27%, I mean, it almost feels like they clipped the numbers round, right?
8:25
So 87% of executives versus 27% of workers using AIA, Huge discrepancy. 58% of employees see ethical challenges and AI, but then only 26% of them will say things that there’s a leader overseeing it, 63% saying that we know we need to develop AI skills, but then we’re not actually having the AI skills being developed.
8:50
Only 17% of the employees saying that the organizations are actually doing anything about developing these skills.
And yet 80% of executives say that employers should re skill workers whose jobs are going to be displaced.
So there’s a lot of do this, but this is what’s actually happening.
9:06
There’s a big divide.
And I think we’re going to see that a lot in today’s conversation around what executives are doing versus what workers are doing.
And I wonder to what extent do you think that there’s a pushback from workers?
Do you think it’s a training or do you think workers are like, I’m not using that because it’ll it’ll replace my job?
9:24
Oh, that’s a good question.
I do think there’s a little bit of hesitation with, OK, AI might replace my job, maybe I don’t want to use it.
But I do think there’s a lot of employees distrusting AI and maybe using it initially, trying it out and recognizing that it doesn’t do their job well.
9:42
I think workers are so much more invested in terms of knowing what their work is and what it takes to get it done, whereas executives and managers are more so overseeing the work that’s being done, not necessarily in the day-to-day putting it together.
9:59
So the output that executives and managers see might be something that looks quite polished and a little bit like what an AI system might output in the sense of just how it looks at baseline.
Whereas the workers recognize that when they put into a prompt and an AI system gives them something, it’s nowhere near where it should be, even though it looks like a final product.
10:20
And I’m kind of seeing this anecdotes totally with friends who are in the workforce and their managers are OK with using AI and think it’s great.
But then the people were actually putting in the work to make these reports, look at the AI output and say no, this isn’t cutting it.
10:36
So everything you’re saying makes absolute sense to me and I can relate to that entirely.
I think this is a question we should hold on to as we look at a number of the articles that have come up and we talk about today.
We can almost revisit this question as we go through it.
And to your point, is it that what the executives are using it for means that that it kind of works for them, but it’s different when you go into deep functional use?
10:58
There’s also some research done by Goldman Sachs about AI adoption and job displacement.
So with AI adopt, workers are concerned about AI potentially replacing their jobs.
And so Goldman Sachs did some research, which the research is a little bit unclear in terms of how they actually conducted it, but this is research done by people who are familiar with the landscape of things as technologies come into the workforce.
11:22
And they said 9.3% of companies reported that they use generative AI in production during the last two weeks.
And Goldman Sachs was saying this was kind of low.
And this was from January 2024 all the way to July 2025.
Five bi weekly surveys about the use of AI in the workforce.
11:40
And it seems as though most companies are operating underneath the 15% mark.
So only 15% of companies of various sizes are using AI in production.
They have this graph and I’ll put it on the screen.
11:55
It shows the adoption of AI from January 2024 to July 2025 across different company sizes.
So companies that are only one to four people or maybe adopting it from 5% to a 10% rate across these months.
12:11
And bigger companies are adopting AI on a wider scale, but it’s still Goldman Sachs says is at a lower level than expected in terms of adoption.
They also talk about AI replacing jobs and so they have an estimate of like 67% of jobs will be displaced.
12:29
C7 or 627?
627.
There’s a Madagascar movie that would have made that hysterical.
Isn’t 6-7 the meme?
That is another one at the moment. 670.
No, we are just so sass.
Goldman Sachs was like, I think 6 to 7% of jobs will be displaced by AII.
12:48
Think that number makes sense when in reality the way they’re coming up with that is just because 6-7 is so on.
Exactly, even if I’m not aware of it.
Anyway, take that with a grain of salt.
They mentioned something called frictional unemployment, which is the period between technology being introduced and replacing certain jobs and then more jobs that are related to the technology come about.
13:10
So there’s this period where unemployment increases, but then it’ll decrease over time.
And they mention how technologies, especially technological disruptions, so technologies that have a large impact on the workforce tend to have perhaps a longer frictional unemployment period and one that people are more concerned about, but that actually these technologies bring about more jobs.
13:31
And I’ve heard this narrative a lot, and I think it makes sense.
But there’s also something interesting about the types of jobs that are being displaced by AI.
So there are certain areas of work that are more at risk for displacement.
13:47
So these sectors are things like marketing, consulting, graphic design, office administration, telephone call centers, computer systems design, and software publishers.
They notice that there is a decrease in employment among these sectors and they’re attributing it to AI.
14:03
But I mean, it makes sense in terms of the roles, let’s say that are kind of at risk or the roles that are kind of contracting.
When you think about where big tech companies are focusing their efforts, I mean, I’m surprised they didn’t just talk about engineering roles.
If you think about marketing and graphic design, I mean, look at Nano Banana, you know, we spoke about that before.
14:25
You know, all of the going into content development is huge.
And then you think about call centers and agents, all of that work is really going straight after agential work.
And how do we automate, how do we replace those kinds of workforces, right.
So the Economist actually wrote a piece postulating new jobs that will be seen in the future.
14:46
They spoke in a similar way.
You know, it’s not all this doom that we’re going to have job displacement in a mass job displacement and we won’t have work in the future.
There’ll be new roles going forward and it really is about roles that’ll be tangential or adjacent to the AI economy and kind of supporting AI industries and a lot of it about kind of supporting the AI agents, AI systems.
15:06
There’s a job description going around at the moment about a kill switch engineer.
What are they calling it a kill switch engineer for?
Well, it’s a joke.
It is a joke.
Listen to this.
A kill switch engineer for open AI.
The description requires a successful applicant to stand by servers all day and unplug them if this thing turns on us.
15:26
Useful skills include the ability to throw a bucket of water on the servers too, just in case.
I love that.
You got it.
You don’t need a big bucket of water for that.
So I don’t think that’s a real job, but the kinds of real jobs that they are, they give 2 examples.
15:41
So the one is these forward deployed engineers and they talk about the kind of roles that are around, you know, you build this fantastic agent, you do a lot of work to develop these agents for these companies and there’s a lot of investment that goes into training agents.
They gave examples like Palantir having come up with these kinds of roles.
16:01
And then what’s needed is to go and embed those agents and organizations to get the full return on having developed that kind of agent.
So if that’s one of the roles, and FDE, if you will, it’s already got an acronym, so it must be a role.
And then of course, there’s the chief AI officer, which is already a new executive role in businesses.
16:18
This role would oversee how AI risk and deployment are managed across the business.
And you know, typically these kinds of roles would be people that bring technical for tease deep industry knowledge as well as having a record of overhauling corporate processes, let’s say in their previous roles.
16:33
So that’s the economist view, with some specific examples of new roles that might occur in the future.
Speaking of roles specific to AI announced in September, Open AI has two AI labour market related initiatives.
There was a CNBC report about this and then you can also find this announcement on Open AI’s blog and it was announced by the CEO of applications.
16:55
They have two main things that they’re bringing about.
They are, I guess, in direct competition with LinkedIn in this respect.
One is the Open AI Jobs platform.
So they say if you’re a business looking to hire an AI savvy employee or you just need help with a specific task, finding the right person can be hit or miss.
17:12
The open AI Jobs platform will have knowledgeable, experienced candidates at every level and opportunities for anyone looking to put their skills to use.
And we’ll use AI to help find the perfect matches between what companies need and what workers can offer.
So if we’re talking about AI and hiring and, and the fact that a big tech company is offering this platform, I don’t know what your thoughts are, but I’m sure they’re similar to mine in terms of all of the risks associated with this, all of the biases.
17:35
I mean hiring and getting jobs is already so difficult in the age of AI and I.
Read these articles and I thought I get what they’re trying to solve.
They claim that they’re getting ahead of AI kind of having created a problem within hiring so people can go and fabricate their CD and CDs look amazing.
17:54
Like literally for those of you that don’t know you can put in a job spec put into ChatGPT, give me the best CV that matches this job spec, job done right.
And so very bloated CVS M dashes all over the show, wonderful credentials, LinkedIn profiles that match everyone just looks wonderful and no one is real.
18:11
So I get that it’s trying to overcome that, but what concerns me is the way it’s overcoming it.
So if I understood correctly, and please correct me where I’m wrong, open AI with this platform want to create a database like by the year 2030, where whoever’s been on this database and they’re partnering with Indeed.
18:29
So everyone that’s been on this database will be profiled, lovely word will be profiled.
And it’s competency based, you know, not even experience based, not skills.
So it’s not your degree, which I have some sympathy for to some extent to be honest, having hired extensively, but it’s competency based and and then you will do some sort of competency assessment with Open AI, which just sounds like a data collection tool quite frankly.
18:52
Oh yeah, so there are other pieces of this that they are offering are open AI certifications.
Again, quote from the blog by the CEO of applications.
Studies show that AI savvy workers are more valuable, more productive, and are paid more than workers without AI skills.
19:09
First of all, there’s so many assumptions there.
That’s why earlier this year we launched the Open AI Academy, a free online learning platform that helps connect more than 2 million people with resources, workshop and communities they need to pass their AI tools.
Now we’re going to expand the Academy by offering certifications for different levels of AI fluency from the basis of using AI at work all the way up to AI custom jobs and prompt engineering.
19:30
We’ll obviously use AI to teach AI.
That’s so.
Funny, obviously.
Use AI to teach.
AI.
Anyone will be able to prepare for the certification in Chachi PT study mode and become certified without leaving the app.
You can stay on the app to get this certification.
19:47
Well, I mean, that’s to consider it.
Yeah.
Oh my God.
But anyway, they’re committed to certifying 10 million Americans by 20-30, and they’ll be doing it with our launch partners, including the biggest private employer in the world, Walmart.
On top of all of that, the blog post itself says we’re launching these new initiatives as part of our commitment to the White House’s efforts.
20:06
We’re expanding AI literacy.
As we continue to build these programs, we’ll remain focused on serving the needs of both workers and employers.
It’s just OK, this is like a monopoly, right?
Like this is what a trust Buster would go after, where it’s like, oh, we’re offering you the input and the output, they’re monitoring you.
20:25
It’s like it’s all under our terms.
I.
Mean, what’s really difficult about this is that if you take a step back and you say the premise AI is coming, it’s here, AI could be causing some skill reduction.
We’re going to talk about that later.
And it’s certainly making roles and jobs change and it would be super helpful if people that are in work learn more about AI.
20:47
That’s all true.
I mean, I don’t think we’re going to get to a place by the end of this podcast that says that’s not true at some form.
The extent to which that that’s like big or small is a debate.
But there’s a part of that that’s true, in which case it would be really helpful if people had more AI literacy, had more skills to use AI that could support them in their work.
21:07
Like that would be really good.
We also know AI in hiring is not helpful.
It’s often prejudicial to the candidates.
It is difficult to hire people and matching people for the right roles is really difficult.
So that is all true.
If you could get a solution that matched the right people, sound like a dating app that matched the right people for the role that was competency based where we wouldn’t just make it about what elite universities people went to.
21:32
That all sounds like a really great solution.
The problem is this that if you keep it all, as you say, within one economy where we are profiting big tech companies by staying on the platform, to your point, I think that’s where it starts feeling a little bit like who’s benefiting more from this?
21:49
How much of this is just good press because it’s big tech that are seen to be displacing people’s jobs?
And so this is them getting ahead of that and going, no, no, look what we’re doing to help.
And we’re doing this as part of the White House’s aim, who have also said AI literacy.
And I think that’s the part that gets a little bit concerning.
22:06
Yeah.
Who’s profiting here?
That’s a big concern of mine as well.
And also big concern is just that how much AI is integrated into the process when we know that AI is extremely biased and we know the amount of oversight it requires for good outcomes.
So if you are just off the loading this entire process onto AI and you’re expecting to bring so many people into this platform, you’re not going to be able to monitor that and make sure that it is ethical and responsible.
22:35
I just worry that this is going to keep pushing jobs and the labor market into a place that’s so far beyond what we can really reach.
And it will once again be how do you trick the system in order to get something out of it?
Like how do you trick the people who are using AI to filter through a bunch of of CBS that they received for one job?
22:55
Oh, you use AI to make your CB to try to get on top of that pile?
It’s all an algorithm.
I mean, you will have seen recently the LinkedIn stuff that was doing the rounds.
They just changed someone’s gender from female to male.
Yeah.
Have you seen that so far?
23:11
Listeners that haven’t heard this, literally what somebody did was they changed their gender in the background.
So it’s not, it’s not their name, it’s not their profile picture.
It’s nothing about them.
So it’s nothing that anyone would see.
They have their LinkedIn profile and all they do is tweak their gender from.
23:27
Which is in the back end, man.
Yeah, and it’s in the back end.
So none of their followers would see this at all.
And their posts got like 9 times or some ridiculous number.
The amount of what do you call it?
Like impression likes impression.
Yeah, 9 times more something ridiculous, maybe even more, but a significant number more impressions just because they changed their gender, which, remember, no one can see except the algorithm.
23:50
Yep.
And so how much else we’ll get tweets in that way.
And so I worry about that.
If it’s being linked to competencies, how much is that going to become a proxy for intelligence?
Does it become static?
I worry about all sorts of questions like that.
Yeah.
And I’ll just point to some research that was done on the bias and algorithms against certain genders.
24:10
So researchers made two identical CVS.
The only difference between the CVS was the person’s name.
And so one person’s name was typically a man’s name and the other person saying was typically a woman name.
And they saw how many callbacks or interviews these different CVS would receive for the same types of jobs.
24:30
And they found that the CV with the man’s name got more interviews and got more offers.
But even when you tweak an algorithm so that you try to avoid this gender bias, the algorithms will find other information relevant to that person’s identity that then gets weighted as the reason why someone might get a callback.
24:51
So I’m thinking also of the application of AI in whether or not people receive mortgages.
And that was seen there as well.
So people’s zip codes where they were currently living were weighted heavily, given that that was supposedly an indicator of people’s, for example, socioeconomic status and whether or not they would qualify for a lower rate or even qualify for mortgage.
25:13
And so even when people who were applying this algorithm wanted to avoid the influence of socioeconomic factors, well, people’s zip codes were still somewhere in the application.
And the algorithm recognized that plus that plus that equals this person is historically not given loans or not given a low rate mortgage, therefore they’re not going to get it in this case.
25:37
So once again, it’s how the algorithm works and the types of information it picks up as relevant based on biased data, right.
So yeah.
These things become proxies for other things that we’re trying not to bias, and that is exactly the risk.
So an interesting story and other interesting things that we found that we wanted to share with our listeners in this workspace.
25:58
Should we talk about Glean?
Yeah.
OK.
So there’s this thing where Glean, this company, launches the Work AI Institute and unveils autonomous agents built on Glean Enterprise context to operationalize AI at work.
So they’re trying to figure out what makes AI at work work.
26:16
And in order to attain that, they are launching autonomous AI assistance to help AI work at work with AI working at work.
Angie I.
You want to say that again?
That sounds like we’re doing a pantomime and you’ve got to say it faster and faster, that’s where.
26:35
I need the impact of the irony in there.
Without work doing at work at work.
At work, at work, at work.
So I’m I’m confused by this.
I just I don’t quite understand what it is, but it seems very relevant to talk about.
26:52
I do get what they’re trying to do and I don’t think it’s a bad idea.
I just think there’s a whole lot of concerns with it.
So let me, let me give you like the meta view on this lean as I understood it, they’re bringing in, there’s some interesting marketing language going on here.
There’s a whole lot of language.
27:07
As you said, there’s lots of work, work, work.
And I think what’s confusing to me is who’s bringing it.
So it’s the work AI Institute and leading researchers from Stanford, Harvard, Berkeley, not UCL, Emory, UNC, Charlotte.
OK, so to be comprehensive, so the institute and like you say, trying to understand what’s really working with AI at work.
27:29
So on one hand, you feel like what you’re going to get is research.
Cool.
And they do have a report that comes out and they talk about the kinds of things that you do need in order for AI agents to work in the workspace, in order for AI to be valuable at work.
27:44
Let’s just say that.
And the report is actually quite useful.
And that does so you’re like how AI needs to be deployed in an organization and you expect that from a piece of research.
I think what then becomes confusing is it does feel like they have also launched these agents for an enterprise system.
28:04
So Glean has an agent builder that anyone can then use to create autonomous agents in a matter of minutes.
And that’s where it becomes a little bit confusing.
So is it this AI institute for work Glean that is a Research Institute that then has also created an actual product and now that’s a for profit company.
28:24
Here they’re talking about agents that complete end to end tasks across some very specific tools that most companies, most big enterprise companies would use like Salesforce, JIRA, Confluence, GitHub.
So these are particular tools that many enterprise companies would have.
28:40
And so it seems like if Glean integrates across those, then it would work as an agent kind of pulling a workflow across those.
Apparently, I guess they’re using data from the report of what works with AI at work to implement into these autonomous agents to make them better at performing in various workplaces.
29:02
So from the report, some interesting insight into what works with AI at work.
You say in this report that AI amplifies.
So point the megaphone carefully so AI doesn’t fix broken system.
It amplifies whatever it touches, for better or worse.
And don’t automate the soul out of work.
29:19
So AI can take on the grunt work.
But if you automate the human craft in judgment, the work becomes hollow and alienating.
These are, again, these are points I think are very useful and on the mark.
They say also leaders can’t phone it in.
AI adoption spreads when leaders actively use the tools themselves, signaling that experimentation is safe and expected.
29:38
OK, that I don’t fully agree with.
What don’t you agree with about that?
Because I’ll tell you why I agree with it.
OK.
Well, I guess it makes sense that for uptake to occur, leaders need to initiate the uptake.
Yeah, but the whole signaling that experimentation is safe and expected, that’s more so a, OK, if you want people to use it, you have to act as though it’s good to use that same.
30:03
It just doesn’t have responsible AI ingrained in it.
So that’s just a little bit.
Concerning, yeah.
So I agree it’s irresponsible from a responsible AI perspective 100%.
But that’s more of an innovation mindset statement.
That’s more like fail fast mindset, yes, that’s kind of like try something, build it, it doesn’t matter if it doesn’t work, try it again.
30:23
And that is very much the messaging that is coming about how you use AI in corporates.
And unfortunately, this is, I think how it gets used internally in corporates is that somebody else needs to think about the ethics.
Somebody else needs to think about the safety.
30:39
Because the reality is not everybody can, not everyone can think about that because you can’t solve for the ethics of it or the data privacy for your company at an individual user level.
So somebody else does need to have thought about that if they’re now going to roll it out to 10,000 people.
30:55
And so by the time it gets to roll out across a corporate, actually somebody needs have thought about that and the best message you can tell everyone is use it, try it, see what happens.
I think when I read this article about Glean and what they were doing, I can completely see from a corporate perspective, you know, obviously there’s always a concern about job displacement, but that concern aside for a moment, this potentially could be some sort of utopia.
31:20
You’ve got automation that’s increased productivity, that gives you more speed, increased margins, great for the shareholders.
You’ve got greater business insights potentially if you’re syncing up Confluence, Jira, Access, sales force.
I mean this thing’s on fire.
31:36
So it’s better for your own business, better for your customers and your clients.
So it gives you a more competitive edge winning and therefore great in innovation in theory, what’s not to love?
So if it was that good, why is everyone not doing it?
Why would they need to try and sell themselves?
31:51
Because if it can do all of that, why is everyone not just jumping on the glean wagon?
Yeah, obvious, right.
So I think some of the concerns and and it does link a little bit to what do they say, be careful where you point it.
There’s obvious kind of concerns that many companies are still feeling, I think which is around data privacy.
32:08
They worry about that that’s really taken care of both for their own data, but also their clients data.
So in some more financial services, military, health, there are some really sensitive data out there and many companies are concerned about how that’s protected.
So even though you have enterprise solutions where they say your data is projected, I think that’s a real concern still for many companies.
32:30
But there’s also, and I think this to me is the biggest barrier to these products being truly earth shattering and groundbreaking, wonderful mixed metaphors is the quality of your internal data.
And that goes a little bit to be careful.
Where you pointed is that if you already have problems, AI will either amplify them or worse, it just can’t stitch together data that doesn’t talk to each other or really bad quality data.
32:53
That’s probably the biggest issue with many corporates in not being able to fully go through this digital transformation is that their data isn’t yet optimal in that way.
Something related to that, which might be the next step forward is what’s called a RAG model.
So large language models, when you use ChatGPT, Gemini, Claude, they have a giant set of data, right?
33:15
Their context window in that sense is a giant set of data, which means that there’s a lot of other information maybe not relevant to your particular prompt that’s being filtered in to the response.
So let’s say you drop APDF into chat GP, that is some sort of research article.
33:33
And you say, give me an overview of this research article, what it is they found?
Well, the context window from this large language model is giant.
So it’s not just this PDF that it’s focusing on, but it’s also getting other stuff from other information.
And So what it gives you is predictive and it’s weighted by, yes, this PDF, but also so much other data that’s unrelated.
33:53
And so that’s where you see hallucinations, but there’s something called a RAG model.
So RAG and what it does is it limits the AI agents context window to a specific set of things that you want it to focus on.
34:08
So it’s not muddied by all of this excess data.
Maybe your context window is 3 PDFs of research articles.
So it’s just operating right there.
I think some sort of application like that might be more useful for companies, but there’s a lot of money required to make those.
34:25
I don’t know.
What do you think?
No, I think you’re right and I think so these narrow applications are really useful, but you’ve got a balance between a narrow application and just how much you can leverage stitching together big data, which is kind of where these solutions are.
34:40
Like when you want an automated agent, you’re wanting lots of data, you’re wanting it to have integrated solutions across multiple data sources, and you want it to be able to operate in the real world.
If you’re wanting to offer automated agents, a RAG format is better in terms of being more narrow for lots of reasons.
34:58
Yeah, I think there’s somewhere in between.
Yeah.
And I mean, we’re seeing so many companies like Lean offering certain solutions for AI.
And one of the things I’m thinking about is I think it was some Microsoft team, but there was a Ted talk by one of their employees, one of the people heading a project.
35:15
And the product that they are providing was an AI agent that helps you think rather than thinks for you.
So with all of this concern around loss of cognition, loss of skill, loss of creativity, loss of autonomy, like I own this work, there are independent teams and firms creating AI agents that are designed not to just give you information, but help you be better processors of information and help you retain your, you know, human value of thinking creatively and thinking autonomously.
35:50
So I think we’re going to see more and more of that kind of thing come out.
We’ll have to keep tabs on the development of Glean and their autonomous AI agents.
At the very least, I think these reports that they put out will be illuminating.
36:06
We’ll let you know more about that as we go about this series.
With this series, we’re not just going to be talking about news and what’s happening with AI at work, but also talking about the psychology and ethics of people’s interactions with AI and how that might change and relationships at work and the relationships with work.
36:25
And one of those relationships is the level of skill that you build.
So what’s all of this about deskilling?
Well, it’s the idea that you might be offloading too much of your work and your cognition onto AI without engaging with the process critically.
36:42
And so then you lose those skills.
And this is reflected in things like automation bias and cognitive offloading.
So automation bias is people’s tendency to adopts an autonomous suggestion by a machine because they believe that the suggestion is accurate.
36:59
Automation bias has been very heavily researched within aviation.
So airline pilots might receive an automated notification from the system of I’m not going to be accurate with this, but let’s say the automation suggestion is turn right and it’s based on some amount of data.
37:15
And so the airline pilot turns right or does whatever.
And as a result, if the system goes down, the airline pilot might not be able to act in certain dire situations without the aid of the machine.
So this idea is digital amnesia, where because you’re relying on a system to provide you with suggestions on how to act, you don’t actually know how to act without the help of that machine.
37:40
So there can be issues if the technology is no longer available.
We talked about also the Google effect with AI in education.
So the idea that people’s memory deteriorates when they the fact that they can find information or search for information on a search engine later.
37:56
So if they want to know something about the Roman Empire, they don’t memorize anything about it because they can just search it on Google later.
Then this is once again the digital amnesia or digital dementia is also what it’s called.
When they look at the research around deskilling, when they look at the extent to which this is actually happening, there is a real risk.
38:14
It’s happening in a few places.
I think when we look at deskilling, we can talk about always like not getting skilled.
So there’s that first level level of entry level jobs, and that’s not like deskilling.
It’s always like not being skilled under skilling, let’s say.
So when you think about entry level jobs, imagine engineering jobs that you’re junior engineer.
38:33
Imagine a junior lawyer, you know, kind of that that entry level, which would have typically been your grad student or can somebody out of school and that’s a really important level for your intern level.
I know you pointed to a conversation with the young people Reliance where they were talking about this risk of these intern roles being displaced by chachi PT, by large language models, which if they can go and do those jobs literally for free or like for pro subscription or enterprise subscription, as opposed to paying someone for a role then or having a whole group of people that are not getting an opportunity to build these skills.
39:10
So I almost don’t want to use the word deskilling because they haven’t yet had the opportunity to build the skills.
But the problem is they’re the people that we need to be doing the higher level jobs in 10 years time.
But how do they leapfrog over that to kind of build those mid to senior level skills in 10 years time when they don’t have that foundational skill?
39:28
That’s scaffolding.
And then we talk about deskilling at the more senior jobs.
There was an interesting article that came out.
I love this guy’s name.
His name is Will Wilson.
Will Wilson.
I literally think that’s his name.
Yes, it’s Will Wilson.
He’s the CEO of Antithesis and he was sharing how both his and other elite teams, he refers to it.
39:50
They had had an early adoption of AI for coding and they’ve now moved away from using AI for coding or to the extent that they had.
And they spoke about a number of reasons for it, one of which was like accountability gap.
So when you have AI generated code, you kind of lack a human owner.
40:07
So who you’re going to blame if there’s something wrong with it.
But the other thing that he did speak about was coding becoming a perishable skill.
So if you’re not constantly using it, then there is this risk of senior engineers no longer having the skill, no longer kind of.
Flexing that muscle and using that muscle.
40:24
And almost like there’s this paradox that these highly skilled developers are losing this skill because they’re outsourcing the coding to AI.
And so they found that that was not working for them.
And they actually spoke about a loss of tacit knowledge.
40:40
And that when you’re having somebody else build the code for you or write the code for you, I should say, you know, it’s almost like when you write your own thesis or you write your own article, you know exactly how you would tweak it because you know, like what that paragraph means, you know what those words mean.
But when somebody else has written the code for you, you no longer have that visceral sense of what’s in there and you can’t kind of get into it.
41:00
And so AI authored code almost functions like someone elses work, but there is no someone.
It’s just a void.
Yeah.
And they speak about it like developers lose the context window needed to make quick informed updates, and it forced them to relearn the code almost from scratch every time a change is needed.
41:18
So there there’s a kind of like a different kind of deskilling and like a dumbing dumbing down, but like a loss of knowledge, shall we say?
And it’s also just really interesting what’s been highlighted from the rapid adoption of AI in different sectors and at work, where it works really well upfront for filling in tasks, getting things done quickly.
41:39
But then when it comes to, for example, bug fixes, the time that it takes to rectify it maybe is even longer than before across the entire process because you have to relearn what the code is and understand what it is that the agent created.
It’s like this whole move fast and then things are broken and then you have to move real slow to fix them.
42:00
And then there’s a completely different perspective, which it would be remiss because we did promise to give different perspectives.
So to give a different perspective, Anthropic put out a piece where they shared their experience of their early adoption of I’m using AI, you know, obviously code for code.
42:18
Now, bearing in mind when we talk about early adopters, I don’t think you’re going to get earlier than adopting your own large language model for code.
So you know.
Also the irony of using coding agents to create agents.
Yeah, I mean, really.
But they’re all doing it.
42:34
They’re all doing it.
I mean, how else do you scale, right?
But that’s so crazy, it’s like asking an entity create yourself.
And I think this is an important point though, because it is always good to know the provenance of these studies.
So this is quite positive towards towards this.
42:51
Not surprising, but interestingly, they surveyed 132 of their own engineers and researchers, and they also conducted 53 qualitative interviews.
And this was conducted in about August in 2025.
And they wanted to understand how AI is changing things that are anthropic.
43:07
And they were looking at it over like longitudinally.
So over the last, let’s say, a year and just in the survey data and going back to bug fixes, they actually found that the majority of the use was for fixing code errors.
So debugging and learning about the code base.
43:23
So people were using Claude to learn about the code base itself.
And you can see how it would be a good teacher.
And you hear this.
I mean, we will hear this in our own work on chat bots, how they’re non judgmental.
Can you imagine how much better it’d be to go and ask a dumb question to claw right than to go and ask your boss where you feel like you should have known it?
43:43
Like, I’m not surprised.
But man, to put a pause on that, talking about the change in work relationships.
I mean, I know that there are issues with not being able to go to your boss with stupid questions, but what if all questions then are just kind of too AI and those who are managers maybe know a lot less about what workers are having issues with.
44:07
And so it kind of sets this expectation that everyone kind of knows what they’re doing when that’s not necessarily the case.
And there could be an opportunity, for example, for a workshop or some sort of programming to help people get their bearings.
44:22
But there’s not going to be any opportunity for that because there will be no recognition that across the board workers are struggling with this.
Therefore, let’s create some infrastructure internally to help people feel that they can do it.
I don’t know.
I don’t.
Worry about, I think that’s the exact point.
And they picked it up in the qualitative research.
44:39
So they actually found that the workplace social dynamics are changing.
So instead of people talking to each other, Claude is becoming the first stop for questions that they would ordinarily have gone to their colleagues.
And so now there’s less mentorship going on and less collaboration.
44:55
And surely that’s not the intended outcome of using Claude.
So I think that is disappointing.
And to go right back to where we started when we spoke about executives using the AI versus workers, one of the things they found in the survey data is 27% of Claude assisted were consists of tasks that wouldn’t have been done otherwise.
45:17
So it was scaling projects and making nice to have tools.
So interactive dashboards, but basically kind of like nice to have.
And so it made me think, going back to that stat, like how much of this is just, it’s not core to the job.
45:32
So to your point, like people that are at the working level, they are so busy already.
They’re doing what has to be done, where maybe the executives are tweaking and doing like that extra chart, they can now finesse something or do a little bit more analysis.
They would have had to do it in Excel or PowerPoint, and now they can use Chloe to make their charts.
45:53
Yeah.
And I wonder if it’s being used there, which I think is interesting.
So I don’t know.
What do you think?
I just see this gap widening between executives and workers, where executives are like under the impression that they can do a lot of the work that their workers do by just using AI without recognizing how much human in the loop labor is required for things to work.
46:20
I don’t know what happened with, for example, the Deloitte thing of who it was.
In case any of you missed it, Deloitte, one of the big consulting firms, put out a report to a client full of AI errors, so hallucinations of citations, all this, whatever.
46:38
It was something that was for the Australian government.
Deloitte agreed to partially refund $290,000 paid by the Australian government for this document.
But I got to think it was also partly with the Australian government.
So it was like a Co produced document because unclear exactly where the liability is.
46:56
But yes, it was for the Australian government.
Yeah.
And fabricated quotes from Federal Court judgements, multiple academic citations of non existent academic research papers, which we’ve seen happen before.
Not a great outcome for them.
There’s a great quote by Brian Lapidus who is an FPNA practice director for the Association of Financial Professionals, obviously in Australia, who says this situation underscores a critical lesson for financial professionals.
47:23
AI isn’t a truth teller, it’s a tool meant to provide answers that fit your questions.
Oh yeah, that’s good.
Which I thought is useful, kind of.
Let’s just remind ourselves what we’re using here.
There are obviously going to be more and more cases where companies or individuals are putting out reports for work that are written with AI injected errors.
47:46
I think we’re just going to be continuing to see this.
I think it’ll roll back maybe a little bit as people start to recognize that AI is not always accurate.
But something that came out also recently that surprised me.
I saw it on LinkedIn today.
So Springer Nature is one of the top academic publishers and it is considered very reputable.
48:05
However, someone created a Mastering Machine Learning from Basics to Advanced book from Springer Nature, published by Springer Nature and Retraction Watch, which is the platform run by academics where they look at whether or not articles or what not have been retracted due to false data, malpractice, anything of that sort.
48:27
Retraction Watch found that when you pay $169 for an introductory e-book on machine learning with citations that appear to be made-up, what do you do?
So it turns out the author of this book used AI, at least to the extent at which there were hallucinated citations.
48:46
Authors were receiving notifications that they had a paper cited that didn’t even exist.
Et cetera, et cetera, it’s.
So ironic when you have something about AI helped made by AI with hallucinations and false information.
49:04
I think you don’t get it burnt, no.
And I just, I’m I it shocked.
Me and it.
Shocked me.
Yeah.
But I am, I am shocked and I think it should shock you because I think Springer, that’s not on brand for them.
But also something to note as well is there’s something called a peer review in, yeah, academia, where if you write up a study and you want to publish it as an article, it has to go through a peer review process where the editor of the journal sends out emails to experts on that topic to read your work and vet it and provide a peer review.
49:41
And so peer reviewers might look at it and say it needs some revisions, it needs major revisions, this shouldn’t be published or yes, publish it.
So there are all sorts of options for the sort of suggestion that the peer reviewer gives to the editor of what to do.
And the peer reviews hold a lot of weight, right?
49:58
Those are or what determine whether or not something gets published.
There are machine learning and AI conferences that also publish papers when people submit for a conference.
So they submit a paper for a conference, they go present it and then their paper gets published in the conference proceedings.
50:16
Now these conference proceedings go through the same peer review process.
There is an online platform called Open Review that a lot of machine learning conferences take advantage of or they do for transparency.
And what this does is it allows you to see all of the peer reviews online.
50:34
Everything is anonymized, but you can see the peer reviews, see the back and forth.
So the process is a little bit more transparent now.
One of the conferences recently released the reviews and an independent company that looks at AI generated texts and has an algorithm to detect AI generated texts found that a large percentage of the peer reviews for this conference were fully AI generated.
50:59
And then even more on top of that were AI assisted authors who submitted papers to this conference publicly withdrew their paper in protest.
So there’s also AI and peer review quite a bit.
Let’s just make sure our listeners get this on me.
51:15
So you mean the actual reviews have been done by AI, just to be really clear, So not the article itself that’s been submitted, the review that’s been the review itself.
Yeah, now.
There’s also a lot of AI generated publications and papers.
Of course that’s becoming a different view.
51:31
Well, but yes, peer reviews are now written by AI.
And when I recently did a peer review for a journal, they were rolling out an AI review assistant demo that you could use to conduct your peer review.
51:46
I did not use it because I, I mean, like if other people want to use it in a way that’s responsible, go for it.
But I personally do not want to use it, at least at this stage.
And three peer reviewers, including me, review this paper.
And the paper got accepted.
52:02
I said it, I, I liked the paper, but you can see what the other reviewers say if you’re one of the reviewers of the paper.
So typically it’s closed to the public unless you are open, open review.
But the other reviewers see the other reviewers reviews and I looked at it and you could indicate whether or not you use the AI demo assistant in your review.
52:21
The other two reviewers both used it and I did not, and the other two reviewers didn’t really have great reviews in terms of not liking the paper as much as I did.
So I do worry.
Also the AI is over critical and a lot of people have found that as well with sort of AI assisted reviewing where it’s like you ask the AI to nitpick someone’s paper and it’ll go further than what a real human would in terms of its criticism.
52:50
But the problem is we’re creating flattening, right?
We’re just flattening everything.
So it’s a regression to the mean.
So for the rest of the series, we’re going to be focusing a lot more on the psychology of things and of course, bringing in recent news.
But something to share with you is that we want to hear your experience.
53:06
So we of course have anecdotes from our colleagues, our friends and our family with their experience about using AI in their jobs.
And we want to hear from you.
So we’ve put together a little questionnaire that asks you about your experience with AI at work.
53:22
So couple of multiple choice questions and then some free response.
Now this is going to be available and open to you if you wish to partake in it.
We’d really appreciate it if you do, but also know that everything stays anonymous and we prefer that you stay anonymous when you do it as well.
And then we plan to share that in later episodes and let you know what it is we found.
53:42
And in future episodes, as we mentioned earlier on, we are going to, as always, have some guest speakers on, but we are also going to be talking to some of our AI Co workers.
So we plan to hear from Gemini and we plan to hear from ChatGPT again and we’re going to hear from notebook.
53:58
LM I meant to tell you, Rose, I had a pre chat.
You know how we set up pre chats with our guest speakers.
Yeah.
So when I was in London last week I thought I would have a pre chat to ChatGPT and with Gemini just to tell them that they would be coming to talk to us about this podcast and they were both interested with a different degree of enthusiasm I must tell you.
54:18
Just more enthusiastic.
Check.
GPT was more like, I want to say emotional and flourishing but I didn’t feel took the job as seriously.
Gem and I felt like a more conscientious Co worker and would be interesting to talk to.
54:35
So it’ll be interesting to have them on in our future episodes.
Yeah, I’m curious about notebook LM, because notebook LM can generate podcasts, right?
Yes.
Do you think they’ll put us out?
I don’t think so.
I don’t know, listeners tell us you think I would do better than us.
I don’t want to say better because whether or not you want a podcast to be better, maybe you want it to be more accurate, more concise, whatever.
54:56
So it’s different from whether or not you would enjoy the podcast as much as human and our errors.
We make it interesting.
But the notebook LM also doesn’t have avatars yet, so if you’re watching the YouTube version then at least we’ll have that going for us.
We have faces.
55:11
Okay, I kind of want to see that.
If it had avatars and could mimic us, I would be fascinated to see what it comes up.
It would be interesting.
Yeah, Speaking of recently I did some sort of interview and the podcast thumbnail was AI generated and it looks like my face, but my eyes are pointing in two different directions.
55:31
Oh.
That’s that’s sweet.
Well done and.
I was like can.
We can we change that?
I’m partially worried.
I was like during the podcast recording.
Were my eyes actually pointing in two different directions at one point?
Was this really just a screen grab of my face?
55:47
Yeah, after I do that, maybe it’s feedback.
But in any case, we want to hear from you the.
Listeners about your own experience of AI in your job.
So we have a questionnaire that we put together that asks you how to use AI at work.
56:03
How do others at work use AI in your team or your organization?
And we also really want to know whether or not you’re given any sort of formal instruction on responsible AI use or best practices for AI use and any stories that you might have that you want to.
56:20
I think this is going to be a really interesting.
Series that I know a lot of people are curious and concerned about.
So we hope you enjoy this series and after.
Listening, get a sense of the research in news in this space and to understand how to use AI in work responsibly and maybe communicate it to both your team and organization for a better future with AI at work.
56:42
Well, that’s all from us.
Get back from Angie, goodbye from.
Rose, we’ll see you next time.