Jamie Bartlett: How do you stay in control of AI at work?

Jamie Bartlett: [00:00:00] You have absolute, like, hardcore pessimists. This is terrible. It's the end of the world. Uh, it's gonna destroy society. The companies are evil. It's a parrot that just repeats things back to you. It's useless. Or you have people that are such hard-line advocates. This is gonna change everything. It's amazing.

Solve climate change, Democratises knowledge. You'll have a personal tutor in your back pocket. And, uh, it doesn't reflect what most people feel about it, which is, is some degree of ambiguity.

James Woodman: That was Jamie Bartlett, our guest today. I'm James Woodman, and this is the Speak to the Human podcast. In the last episode, I talked to social psychologist Dr.

Guy Champniss about his research into AI at work. One of the things Guy found is that a lot of us see AI as a threat to our professional identity, and it really made me think that those emotions deserve our attention. More precisely, what do you actually do about them? So that's the question for today's episode.

[00:01:00] When you use AI at work, how do you stay in control? Jamie Bartlett has spent his career writing and broadcasting about technology, including the global hit podcast, The Missing Crypto Queen. His new book is called How to Talk to AI and How Not To, and Jamie's agreed we can adapt that title very slightly for this episode, so we're gonna pretend it's called How to Talk to AI at Work and How Not To.

Jamie, where, where do you wanna start? Imagine maybe I'm someone who's been told I should be using AI to do my job. My boss wants me to use it, but I'm feeling a bit out of control. What are you gonna say to me?

Jamie Bartlett: Oh my goodness. It's not an easy question to answer. I, I understand why people would be very worried about this.

Uh, you've got two choices often. One is to sort of bury your head in the sand and hope it'll go away, or perhaps use it a couple of times, find out that it's not brilliant at everything, declare that the whole thing's just useless and a scam and pointless, and [00:02:00] then crack on in the hope that, uh, that it all goes away.

Or, of course, you can be-- you can really try to learn it. You can really try to engage with it and understand it. And I, I, I've, I think that one of the problems over the last fifteen to twenty years of technology adoption at work is a lot of it is boring. And AI, on the surface, can feel like that. Oh, you've got to use Microsoft Copilot 365, and you-- there's these restrictions, and I've got to go to an AI compliance course.

But actually, it's really different to other technology, um, introduced at work because it is, it is a q- it is about words. It's about how you frame questions. It's about how you formulate problems. It's something that can talk to you in your language about anything you want. And when you frame it like that, you suddenly realise maybe I'm, you know what?

I'm amazing at crosswords, and I love [00:03:00] reading poetry. Maybe I'm amazing at using AI. If this is just really a-about using words and language, uh, it's not about studying some new tedious software. It's about thinking about problems in a different way and working out how I can make it work for me in my particular way, not the standardised software that everyone has to download and use in exactly the same way.

So I think it's about reformulating or like not imagining this as part of the twenty-year-long tedious application of new IT software. It's something totally different, and then it becomes slightly less scary and maybe slightly more exciting.

James Woodman: Let's talk about the book. Why did you, why did you write it?

What do you wanna say?

Jamie Bartlett: Well, I mean, actually the book is full of warnings, and I'm very worried about a lot of this as well because I'm, I'm nervous that a-a-a-the, the, the risk to all of us of being misled, of outsourcing our thinking, of confusing fluency and [00:04:00] articulateness for accuracy and knowledge is extremely risky.

So while they can be amazingly valuable, these large language models, there are many pitfalls, and those pitfalls, I think are, unlike other software, basically almost the same whether you're using it in your personal life or at work. I mean, that's one of the different things about this. I-- When I had to learn how to use Excel properly, I didn't then go home and use Excel all the time, uh, for all my other things to help me with my training or help me fix my car or...

So it's, it's so multipurpose that there isn't such a clear dividing line anymore between work tech and home tech. Um, so-- And I was writing for both audiences. I mean, I was writing for everyone. I, I think what tends to happen with all technology in, in the years I've written about it, a small group of pioneering advocates turn up and tell you how the world's gonna change.

They'll talk about technology, and ninety-five percent of people feel it washes over them. They get dragged [00:05:00] along. They don't understand it. They're scared of it. Um, and I've always tried to write things that is for that audience, for the, for the people that are being affected by this, that have to change their life, but, but, but feel quite worried about it.

They don't really know what they're doing, and they don't feel techie. It's very easy with AI to think, "Oh, it's machine learning PhD people. I don't get any of it." And, and then, and then no one uses it well as a result. Because what really changes the world is how the ninety-five percent use these things in their daily lives.

And so I wrote-- tried to write a very accessible book, a really simple book. I think you've even read it. Thank you very much. So hopefully you could vouch for me that it, that it is simple. It's not, you know, none, none of it's technical. There's no statistics in there, nothing like that. And I, and, and, and I was aiming at those people 'cause I think they're the people that really need to know now what's going on because I don't know what the number is exactly because no one gives you the number, but it-- I know there's over a billion users around the world that are using large language [00:06:00] models regularly for all sorts of things.

In the UK, uh, it must be fifty, sixty percent of us now use at least one la-large language model for something, and most of us don't know what we're doing. We're relying on it in all sorts of ways, and we're clueless, and I don't want... I think that's really bad.

James Woodman: I'm trying to work out whether you're an optimist about this or you said you were scared a minute ago, and there's a lot of warnings in the book, which is true.

And I can say I do totally vouch for the book. It's brilliant, and I think it does make this very accessible for anybody who wants to do this better. But it's-- Where, where are you on that spectrum between there's, there's optimism here about the future, and I'm scared this is a big change, and I don't know if humanity is ready for it?

I think most people feel, feel both.

Jamie Bartlett: I think most of us are, are treading that line, and it-- People sometimes use a language model, and it helps them get a refund from their energy provider, [00:07:00] and they think it's amazing, and they love it. And then someone else reads an article about how it's gonna take our jobs, and they hate it.

And it's confusing for people because this technology is so unique, and where I sit, I try to avoid, try to avoid being overly optimistic or overly pessimistic about it. I think what tends to happen with books about AI and articles and anything you read about it, you have absolute, like, hardcore pessimists.

This is terrible. It's the end of the world. Uh, it's gonna destroy society. The companies are evil. All it does is it's a, it's a parrot that just repeats things back to you. It's useless. Or you have people that are such hard-line advocates. This is gonna change everything. It's amazing. Solve climate change.

Democratises knowledge. You'll have a personal tutor in your back pocket. And, uh, it doesn't reflect what most people feel about it, which is, is some degree of ambiguity. And, and so I, I, [00:08:00] I tried to write a book that, that does contain both, that says there are some amazing things we can do, and I'm-- but I'm also extremely worried.

And, and I think the best thing that I could do is to write something that will help ordinary people get a little bit better at using it. Not to say this is gonna transform your life. You can ten-x your productivity. You'll, you'll never have to do another day's work. You can do it all in an hour, thanks to ChatGPT.

Uh, that's no good. That's useless. There are books like that. It's not true. It's not true at all. So I try to be in the middle, which is sometimes a bit difficult 'cause to be honest with you, books that take a hard line one way or another tend to sell better. So I was quite worried about trying to be this centrist in the middle.

But, uh, but that's where I am, and that's, that's what, that's where I think most people are.

James Woodman: Yeah, and I think it's, it's new, isn't it? In the sense that when the first spreadsheet came out, nobody was worried about it, I don't think. Maybe if your job involved writing numbers on a piece of paper, you were [00:09:00] worried about how it would change.

But fundamentally, it was a positive thing, and I think people got that pretty quickly. But this is a piece of technology that has an impact on everybody, whether they like it- Yes ... or not, and it isn't universally- Yes ... good. It is not possible to say this is a universally good thing.

Jamie Bartlett: That's right. And it also-- There's certain things about it that are so different.

Um, it i- it is in many ways, although it comes from the world of machines, in many ways it is more like a biological creature than a machine one. Uh, I mean, these models are grown rather than built. They're fed vast amounts of data. No one quite knows how they work, and they learn from that data, and they start coming up with their own ideas and their own suggestions.

They are increasingly autonomous, which no other technology really has ever been able to go out into the real world with its own, with its own goals that we've accidentally given it sometimes and actually affect change out there. And it speaks in our language. [00:10:00] I mean, no, no, in two hundred thousand years of modern humans, we've never had anything that can communicate with us that's not another human.

Um, so it's, it's almost misleading when you frame, not you, but when one frames AI as merely the next step in the evolution of technology being introduced in the workplace or in our lives because it feels so radically different. Yes, there's data and there's statistics, and it comes at you through a screen, but they're quite superficial similarities in some ways.

This, this really does feel like something quite different, and it's good for people to think of it as quite different 'cause then I think you can work out how to use it well and what its risks are in a more accurate way.

James Woodman: You've talked about language a couple of times, and you mentioned that idea that, okay, if you are somebody who enjoys using language, then that is a thing.

That's a skill that you have. If you enjoy crosswords, if you enjoy reading, if you enjoy writing, then that is a skill you can bring to your communication. [00:11:00] It feels like it's- The communication through natural language is both a, an opportunity and a risk. I, I think people who know you from The Missing Cryptoqueen will remember the idea that when people don't understand something, they look for shortcuts to judge its quality or its relevance.

Y- in that series, you were talking about people choosing to invest in crypto because of a charismatic leader. They weren't qualified to judge what they thought they were investing in, but they believed in the person. With AI, I think that fluency, that sense that this is something that actually for a lot of us, it can write better than we can ourselves, so we assume it's smarter.

We assume there is human thought happening here. The writing quality could be there, but the thought and the smartness aren't. How do we, how do we get past that? How do we understand that?

Jamie Bartlett: Yeah,

James Woodman: exactly.

Jamie Bartlett: That's exactly right. Um, I think, well, firstly, how we get past it, I, I, I can make some suggestions.

They're quite top level, but I think [00:12:00] understanding that new relationship between the, the sort of quality and fluency of a written output or a video, whatever it is, a audio output, whatever, but, but a machine output, and its likely accuracy or usefulness or truthfulness is, is no longer, uh, they're no longer re-related in the same way.

I mean, for many-- for, I mean, forever really, we humans have tended to associate, uh, high levels of accuracy, uh, detailed descriptions, statistics, facts, uh, all the right conventions that tend to go into writing certain types of work. Like when you write a legal report, there's conventions you follow. You tend to associate those conventions and styles as being correlated with likely accuracy, and that some degree of effort and work and thought has gone into it.

This is so hardwired into us now that it's very difficult [00:13:00] to suddenly be told that's no longer the case. A ten-thousand-word, perfectly articulate, well-written, detailed reference report can be produced in thirty seconds. Uh, it is not a reflection of a person's knowledge, skill set, or effort, and it is not a reflection of accuracy.

A machine is capable of generating entire universes of fabricated information, um, in a way that a, a human would have to be a genuine psychopath to be able to do that. Um, human lies tend to be short and simple. Uh, machine lies often tend to be extremely long and detailed. So th-these are, these are things that I, I, I can t- I can tell you all this.

I can say-- And maybe listeners will sort of recognise that and understand that to be true. But to actually internalise that and to To, to, to, to, to not allow yourself to [00:14:00] be fooled that way is much harder. I think it takes a lot of practice. One of the things that I recommend people do is to intentionally get a large language model to hallucinate or to frame questions in a...

Like, if you know how-- What language models will often do is when it knows what you're looking for, for example, "Hi, I, I want an example of someone in the UK who has, uh, fallen into an AI delusion," uh, it will desperately try to f-please you and give you that answer. It will then... And you framed the question in a way that you've already presupposed that that exists, so the machine thinks that exists, and it goes and finds one, and it will generate a statistically plausible response to you, which may very well be completely made up.

Th- And, and so one thing I suggest is you practice writing prompts that you think can generate hallucinated answers. You practice [00:15:00] intentionally generating reports that you nudge the model into including fabricated statistics or references. Once you have done that a few times yourself, um, I think you have a better sense of how it can happen and why it happens, and your ability to maybe spot it when it does happen.

And I think that's more subtle than... The advice on this is generally really boring. It's things like, "Oh, make sure you check your statistics because a machine can lie to you." But it's not really getting at the psychological reasons why we fall for it. Um, so that's, you know, I, I'm being quite top level here.

There are specific exercises I think people should run, but it's all about doing the thing intentionally To inoculate yourself against falling for that thing accidentally

James Woodman: You are effectively talking about rewiring our brains a bit, like reprogramming ourselves so that we can do this better and be more confident in the results.

Jamie Bartlett: Yeah. And I think it's [00:16:00] useful when it comes to, to, to-- Hallucinations are one problem. Another great problem is the way that the machines will flatter you. You-- Um, most people I think get that now. But again, psychologically, when a very powerful machine tells you, "No, James, your idea is actually brilliant.

This is a really good essay. It's a really good piece of work. I don't, you know, I can't really see any, what, you know, I think it's fantastic, and your views are really unique and original." Even when you know that the model has been designed to flatter you, it is extremely hard to not at least think that the model might be right this time.

Um, because we like, we like flattery, especially if you kind of think that as well. You have to rewire your brain, like you say, as a defense mechanism against that because it's such a powerful psychological instinct, and it, it really does all come down to this. It's like there's loads of prompt engineering guides out there, but most of them are about how to write a prompt.

They're not really about how to [00:17:00] think yourself, how to try to test your own prejudices, how to recognise assumptions in your own views and counteract them. But, but that is what the majority of good AI prompting should be about. Machine-human interaction, how we respond, how we think, how we react.

James Woodman: Could you give us a, a couple of ideas for something simple somebody could try at work that will demonstrate this to them?

So workplace things that would apply in any, any, any office environment. But what could somebody put into their AI tool to show them how easy it is for it to be wrong?

Jamie Bartlett: For it to hallucinate?

James Woodman: Yeah.

Jamie Bartlett: Okay. Well, of course, the, the way that these mo-models hallucinate va-varies greatly by task. There's no agreed statistic on how often it happens.

It, it varies by model and by task and by question type. I mean, we [00:18:00] just know it happens quite a lot. I would suggest you, you find a subject area where there's maybe not loads of data. Let's say you're looking, you, you work in business intelligence, and you're, you're looking for some statistics on, um, business trends in investment in South Korea, something like that.

Um, and, and ask a very specific question where you, you sort of presuppose there is an answer already. Give me the business invest internal, like, you know, direct investment in South Korea from country X over a ten-year period, ordered by or ranked by X, Y, and Z. The more specific, the better, and specifically when you're asking it to give you an answer that you already want.

"And I want this data to show me the inc- that, that this number has increased over the last 10 years." So I can't give a specific example 'cause it depends what people work in or what they're gonna try. But something like that, where the model sort of knows that you, you're looking for something, it wants to please you, hasn't got loads of data to [00:19:00] work with.

It would, it-- There's a good chance it will generate a plausible response that if you then try and verify it against some real world data, you'll find out that it's made that up. And hopefully when you do that, you realise that's the way I ask questions all the time without realising it. I need to change how I'm asking questions.

Um, similarly, you might, you could try to start trying to sway the model, so when it gives you an answer, decide that you're not happy with that and say, "No, I think you're wrong. The foreign direct investment into South Korea over the last 10 years has not gone the way you've told me." You're com- like, "Try again.

Find me the answer again." Or, "This-- I'm really certain you're absolutely incorrect about this." Layering on quite a lot of emotional pressure, which the models do respond to a lot. Uh, y- again, you might find that it will sway and come around to your way of thinking, uh, and give you an incorrect answer as a result, the answer you want [00:20:00] it to give you.

And again, hopefully by doing that, you come to see that's the kind of thing you might do all the time without realising it. So these kinds of exercises, doing the thing on purpose that you probably do anyway without realising it, I think it's just a really good way of understanding how the models really work and inoculating yourself a little bit against some of those weak-- some of those risks.

James Woodman: And there are loads of examples of things like that, practical things people can try in your book, which is one of the things I really like about it. There's something that you touch on in the book as well, is the value of friction. I mean, think, let's think about this in the workplace. If everything becomes a summary, then you don't actually understand any of the, the meaningful content behind it.

So how do you get that balance between how do I do this thing versus how do I get the machine to do this thing?

Jamie Bartlett: I think that's most, the most common and popular use for language models at the moment, as I see it in the workplace, is [00:21:00] probably document summarisation. Uh, we all have to read very long, often quite boring and complex things, and, uh, we're all busy.

Uh, models are able to give you extremely good, concise, well-written briefing, short summary notes, uh, and sometimes that's really useful However, a couple of things. The first is you need to understand that if you ask a model to summarise a document for you three times in a row with exactly the same prompt and exactly the same document, it will summarise it in different ways each time.

It will leave certain things out and add certain things in. So you'll get three different summaries. Don't ever assume that the summary you get is the right one for you, or is the most accurate one, or includes the things that you want. You gotta be very specific about it, and I always suggest with important tasks, you should do it multiple times rather than just accepting the first answer.

[00:22:00] Equally, and maybe on a more philosophical sense, it's going to be very, very easy to turn everything into a summary. Sometimes the process of engaging with difficult information, of, of reading the boring 100-page report, it's not only the way you actually understand something or internalise information properly, like you'll be able to remember it better, you'll be able to think about it more fully.

It can also be a really good way that you stimulate your own thoughts and opinions and views about things. I mean, I find that if I'm struggling with a problem, like I-- when I'm writing a podcast, for example, and I'm, I'm trying to figure out the way of ordering ideas, I will not ever ask for a summary.

I'll listen to entire three, four-hour long interviews or read the whole transcript 'cause it gives me a way of-- it's a pro-- it's the process by which I come up with my own view on things. Each of us has to determine when the process is the [00:23:00] point rather than the output. It's gonna be very easy for the output to always be the thing we're looking for rather than understanding the process is sometimes what really matters.

And sometimes summaries are great. Sometimes language is intentionally obscure, vague, windy, complex, such that you don't understand it. But sometimes you need to engage at that difficult level, and it's just a choice. I guess it's more-- it's about each person understanding that is a conscious decision you need to make.

Don't just reach for the great summarising GPT machine every time you are faced with a difficult document.

James Woodman: Maybe that's a good place to turn to the, the positive bit then, the, the how, how to talk to AI, that bit of your book's title. How d- how do you do this? You've talked about thinking it through before you reach for the AI tool.

What else? What does doing this well look like at work?

Jamie Bartlett: I tend to think of it in terms of habits rather than [00:24:00] specific prompts that are good, because every prompt depends on your own use case, your own task. I mean, any-- if you ever read anyone telling you this prompt will 10X your productivity, it's absolute rubbish.

Just completely discount it and everything else they say. But there are certain Not just habits, but specific ways of talking to machines that I think are better than others. Um, and it-- like I said, it's, it's a two-way, it's a two-way relationship between you and the machine. And, and there are some very specific techniques that I suggest.

I'll, I'll give you a couple of examples, and then the book, the book... I won't go through them all 'cause there's loads, but the, the book has them. One is, uh, context, the importance of context. You have to remember that a machine doesn't really understand your ultimate intention when you ask it for an output, not in a way that a hu- a human would.

They're getting better. [00:25:00] Reason-- modern reasoning models or when you have reasoning or thinking modes on the language models, they are better at trying to work out your ultimate intention rather than just interpreting your words literally. But they're still not, it's still not like the way a human te- has a series of assumptions about what it is you're looking for.

If you're asking a researcher to assist you on a research brief, and you say, "You know, I'm, I, you know, get... I n- I need, I need five hundred words about subject X," that person will, will kind of know the sort of things you mean, the sort of things that might be important, the sort of reasons you have for needing that brief, and therefore tailoring that brief accordingly, and a machine just does not have it.

So the more context you can give it, I tend... If I'm asking for, like, uh, s- uh, feedback on something I've written, I overload it with context in a way that a human would find it so tedious to be given the number of [00:26:00] instructions that I give it. Like, "This is my audience. This is the register I want. This is why I'm writing it.

This is the mood I'm in today. This is, uh, h- exactly how long I want it to be. Here's ten examples of where you've done this before that have been really good." And so you, you put in so much more context and background information, even things you think might be irrelevant when describing a task to a machine relative to a human.

Uh, another thing is people, I think, massively underestimate just how creative these tools are. They j- I never ever, or very rarely at least, talk to the standard version of a language model because they have been designed to be statistically plausible and therefore very boring. You know, they take the sum total of the trillion words that they have found, uh, in their training data and are generating plausible sentences, and they want them to be plausible, which means they tend towards the average No one ever wants the average though.

No one wants to [00:27:00] write the average. Um, but that is what most people using GPT do. They just open up the browser, or w- maybe it's Microsoft Copilot, and just start typing in and asking questions without describing what type of persona, what type of role you want the model to play. The minute you start pushing the model into a different zone and telling it, "I want you to be an editor from The New Yorker magazine.

You're fiery, you're angry, you're dismissive of my terrible ideas. You know, you're ready to get into an argument. You're obsessed with details. You're, you're-- but you're, you're only 25 years old, so you're not actually that experienced, but there's a, but there's a naivety to your work that is really endearing."

You know, and you really start creating a persona for your model as creative and bizarre and peculiar as you can, and the weirder the better, the more you will find that model begins to produce outputs that do not look like they are machine-generated. Um, and this [00:28:00] seems pretty simple and basic, but it's amazing how few people do it, and I suggest that people really think about that.

And when you have created a persona that you have found really useful for you, save it. You don't need some fancy file system. This is your intellectual property in a way. This is your... This could be a very valuable identity you have created that others in your office might also want to use, and it could be thousands of words long.

Save it, store it. It just needs to be a Word file, because you can either create a customised GPT model or just use it as a prompt, a basic prompt as to what you want the model to do. And, like, y- so, so people underestimate the power of, of like how much you can push a model around and get it to behave quite differently.

Like, you know, I could go on and on and on. As you can see, there's, there's so many. But there, there are a couple of really, really simple things that most people don't do and really should.

James Woodman: And maybe there's something as well about recognising [00:29:00] the, the model is there to do what you tell it to do, and that if you ask it to do something, it will go ahead and do it.

But you say in the book that encouraging it to challenge you and things like saying, "What information do you need from me to complete this task?" is a really powerful technique, or asking it to help you improve your prompt. 'Cause if you don't ask it to do those things, it will just go ahead and do the thing that you put in into the window.

If you ask it to help you make that better or tell it-- get it to tell you what you need, you will probably get better results.

Jamie Bartlett: Every single person should be thinking of asking the model for help in how to use it, and there are lots of ways to do that. You, you, you can, um- When you're asking a question, uh, like, you know, like f- I, I specialise in research, so a lot of my questions are research-based questions.

You, you can very easily ask it like, "What's wrong with my question? Why-- Is my question biased?" So the f- "Give me [00:30:00] five ways to ask this question to get different perspectives on it." So you can definitely just in, in just the, the mechanics of asking a question is really, really easy. And the other is if you wanna write a really long and powerful type of prompt, it's also quite boring.

You can ask the model to write your prompt for you. As long as you can formulate your idea very clearly, express what you want very well, the model will then generate a prompt to achieve that goal, which you should then edit carefully. But it might have written 1,000 words for you. And when I did jailbreaking, when I was jailbreaking some of the models, i.e.

getting it to tell me things it's not allowed to, I was using the model to help me write very, very long prompts, 1,000-word-long prompts, 'cause that's one of the ways you can trick a model, with it sort of overloading it with too much context and information, so it doesn't quite know what it is you're doing.

So you're, you're vague and weird and confusing on purpose to confuse [00:31:00] it. But obviously you're trying to do th-this the other way around. You're trying to get it to help you to be very clear and precise. So you get this amazingly long prompt back that might take, might have taken you hours to write.

It'll do it in seconds. Then you go through and really carefully edit it. Don't just take it. Just edit it, think about it. Y- Try it, use it, redo it, and you'll-- And, and again, your, your, you will find your results are completely transformed by that. And it's so simple, 'cause in a way you're, you should think, "Yeah, of course.

Obviously a model would know how to speak to a model well. A model knows..." And so, so much of this when you think about it is, is kind of common sense, but it's also about you being creative and willing to try things.

James Woodman: I don't know if this is about common sense, but one of the things that is s-specifically about work is around productivity.

You, you mentioned this in the book. There's that quote you have from, uh, an economist I think in the '80s about how you can, you can see the computer age everywhere except in the productivity statistics. [00:32:00] And I think people feel that, this sense that even if I'm not using AI much, the people around me are, and they're generating potentially thousands of words, more than we've ever had in our lives before.

How do we even cope with that as humans without just sticking them in ChatGPT and saying, "summarise this for me"?

Jamie Bartlett: I think what has been happening lately is that individual productivity as measured by self-reporting, people saying they're more productive, is, is people m-measuring their actual output rather than the value of their output.

I, I can now generate five PowerPoint slides in one minute. I can now write a 10,000-word report in 30 seconds. Um, that's all well and good for individual productivity. Well, kind of. But what it's doing is creating an epi- epistemological crisis inside companies where no one in the company can be sure that what they're getting from their colleagues is real or not.

So then they have to [00:33:00] check it twice as long to make sure that it's accurate. And you're in this weird situation where individual productivity is going up, but sort of company-level output is not really improving. And it's, uh, I think it's, it's one of those things that I don't think there's an obvious answer to that in the way that computer productivity in the 1980s and '90s, it took sort of ten or fifteen years for us to really figure out how to do this.

There's, there's definitely, I think, something that people using it in the office need in the... And I talk about, like, in the office widely, need to think about. There is no point just generating more stuff. What's the stuff for? How is it achieving the company's aims and objectives? Are you giving other people more work inadvertently by producing more stuff?

Um, and if you're someone that's just churning out content relentlessly, um, [00:34:00] maybe, uh, th- maybe you should produce less higher quality output, and then that is based, again, partly on how you're interacting with a machine more wisely. But it's a tricky one. I, I think we're in the middle of this problem. I speak to people who work for companies and for public sector bodies who are saying the problem at the moment is their complaints team or their public consultation team are receiving ten thousand-word-long complaints now that are perfectly written with detailed legal, uh, reference points that they feel like they need to follow up with, while simultaneously kind of knowing maybe this is just written by a machine, and their systems are getting clogged up, and they've got to now design language models that can sift through that somehow.

James Woodman: I was having a conversation with somebody very senior who works in a role where they are responsible [00:35:00] for the people replying to those complaints, letters, and they are generating the responses with AI. They have a human in the loop. They have a human who has to look at the AI-generated response and say, "Is this okay to send back?"

And we were talking about the fact that that is actually probably a pretty awful job. Like, you've had two hundred and ninety-nine of these today. The three hundredth complaint reply comes in, and it's your job to say, "Is it okay? Are you gonna do that thoroughly?" Or will you say, "Well, the last two hundred and ninety-nine were probably okay.

Approve." Like, it's changing the way that we work at both ends for people outside companies and inside.

Jamie Bartlett: It's a, that is a, I mean, it... Yeah, it's a, it's this is a, this, this is something I really wanna think about and work on, and I don't have an an- I don't have an answer to all these problems. I just know it's, I know it's happening, and it makes sense that it, that it happens.

I mean, it's, it's obvious that this is what would happen. Uh, there's gonna be a particularly difficult thing for ei- either companies or public bodies in [00:36:00] particular that are under some kind of legal obligation to respond to certain types of consultation. And if you work in planning, objections come in, and you, you have to look at them all carefully.

That, that's quite difficult if you've now got 50,000 of them and each one is 100,000 words long . So-

James Woodman: Yeah,

Jamie Bartlett: I mean, it's impossible, I suspect ... it's impossible. It's impossible. You, you have to have a machine assistant to do it, and you have to work out how you're going to do that while still, you know, complying with your legal duties and your com- you know, your cust- there might be someone on the other end who is going to take your answer and run that back through a machine, find the response, find the errors in your machine-generated response, and then write another one back to you.

And so you're gonna, you know, the people that are generating machine answers, uh, even with a human in the loop, also are gonna have to build in machine checking systems to see, to make sure that those responses are good enough. And it's a whole brave new world, isn't it?

James Woodman: I, and I'm, I don't think [00:37:00] it's reasonable to expect you to have all the answers.

Um, but what I think your, but what I think your book is doing and what you're doing is encouraging people to think about this stuff, and maybe to be a bit more intentional about the ways they're using AI. So recognising what... If you are literally putting "summarise this document" into Copilot and doing something immediately with the results, you might not realise it, but you're making a choice.

And actually being more conscious about the choice you're making and giving yourself the opportunity to choose to do something a bit different, maybe through doing the kind of exercises we've talked about, the practice prompts, the trying stuff out, trying to break it, trying to convince yourself of its failings as well as its opportunities.

That's the thing that perhaps makes us make different choices just in everyday moments when we're deciding what to do.

Jamie Bartlett: Yeah. I think that's a great way of putting it. Like, intentional. That, that's, that is a very, very important word in lots of ways. Even down to your very first question when faced with a task should always be, [00:38:00] "Do I need a language model for this?

Or maybe I can just do it another way. Maybe I can just do it myself." Maybe like you should always ask yourself that question. Uh, I don't, but I don't think people do. I think there's the, there's the pull, the draw of using the model for everything, even when you don't need to.

James Woodman: It becomes a habit To kind of round things off, we've-- you talked about that tension, that, like optimism versus fear and the, the tension that exists here for everyone.

What would you say that would make me feel good about this? Wherever I am on that spectrum. What are you exci- what are you, what are you excited about people doing with AI? I guess that's the question. What are you excited about?

Jamie Bartlett: Yes. Okay. Interesting. I've-- I mean, so I am worried. I always, I always add that as a caveat in case people think I'm one of those, like, just banging the drum for this.

Um, there are many things I'm worried about. The things that I am excited about, the-there are, there are a few. For many, many of [00:39:00] us, language is intentionally, uh, confusing and alienating. Uh, it's written in styles we can't comprehend. It's either it's legalistic. If you're neurodiverse, it can be very inaccessible.

Uh, people of all different age ranges and ability are supposed to read the same text on a website. Uh, uh, your contracts are totally impossible to understand, and millions of us get fooled in various ways as a result. This-- These models are amazing style shifters. They can turn language from one style to another while retaining the meaning extremely well.

Yes, there are problems, and it's sometimes not perfect, but it could be a way of really unlocking language for lots of people. So that could be incredible. Um, the other thing is that y-you do have... If you know how to talk to them properly, this is not Silicon Valley hype, you really can [00:40:00] have the world's greatest devil's advocate by your side.

You can have an amazing business coach. You can have an amazing, um, a, a extremely professional and proficient editor, a brilliant thought partner, a, a, a superb business consultant, all within one model that is not that expensive to run. Uh, if you know how to use it in that way and how to make sure that you don't just trust everything and all those other things we've talked about, you know, you have access to things and ideas and possibilities that was just never possible.

Um, and I think it can really open up new types. Like you'll find new potential in new people. There could be amazing com- potential cybersecurity experts, software engineers, business consultants, but who've never learned any of that stuff, but just have the right mind for it [00:41:00] And they've got the right way of talking about these things, and you might suddenly find that there is this ge-ge-genius kid who's this incredible business consultant, but he's just so good with words, he can get so much more out of the model than a business consultant can.

So you might find that your people that are, are, are amazing linguists are suddenly actually incredible in other fields that before they never would have been. So there, there are, there are reasons to be excited, and I think sometimes with all the doom and gloom about AI, it's easy to forget about them and get lost in the job apocalypse and all these other existential threats, and they're, they're real.

But it is worth keeping hold of the positives. Like, how much time do you spend on pointless, annoying, tedious admin? That you know you don't need to do it. You know it's not helping you. You know it's not getting your business to where it needs to be, but you just have to do it, and you know what you gotta do.

It takes hours, and it's annoying. [00:42:00] I mean, we're drowning in this stuff. Models can actually help you with that. That-- and that's on an individual level as well as a professional one. I, I, I did a call in with the B-BBC Five-- Radio 5 Live the other day, and this is such a good example, and I want everyone to think about how this could apply to them.

Says, "I got a bill in from my energy provider. It made zero sense to me. I would-- I knew it was wrong. It was so confusing, and I knew it was gonna take me hours to get to the bottom of this, so I uploaded all of my bills, all the letters I had, asked the model to try and figure out what had happened, and then write the letter for me back to the energy company."

So the machine did all that, analysed it, wrote the letter to the company, and got this guy a refund. Took him, like, five minutes. And I'm like, "If we can figure out how to do that at scale and in companies, we're, we're gonna be flying." It's gonna be amazing, but we have to understand how these models work for that [00:43:00] to happen.

James Woodman: And all of that depends on people knowing how to talk to AI and how not to. Exactly. Which I guess brings us back to your book. Th-thank you, Jamie. It's been such a, a pleasure talking to you about this. I really enjoy hearing your perspective, and I think you've got a lot of really valuable information, ideas, thoughts, creativity, just practical things people can do.

So thank you so much for, for coming on the podcast.

Jamie Bartlett: Thank you very much for having me.

James Woodman: If you want to find out more about all this, I think the single best thing you can do is buy a copy of Jamie's book today. I really love it because he puts into words so many of the things I want to say to people about AI and that I found difficult to express positively over the last couple of years.

It's a book that puts humans first, but it's not anti-AI. It's optimistic about the technology and its potential when we know how to use it well. I've already recommended it to loads of people in my life, and I recommend it to you. The title is How to Talk to AI and How Not [00:44:00] To. It's published by Penguin.

You'll find it pretty much anywhere. We'll put the details in the show notes as well. Thank you for listening to Speak to the Human. I'm James Woodman, and if you've got a question you'd like Acteon to explore in a future episode about anything to do with the human experience of work and how we can make it better for people rather than machines, please get in touch, hello@acteoncommunication.com.

See you next time