Leah Krauss: Responsible AI and Content Design at Microsoft – Episode 31

Podcast: Play in new window | Download

Subscribe: Apple Podcasts | Spotify | Android | TuneIn | RSS

New AI products like Microsoft’s Copilot can be powerful productivity enhancers, but if designers aren’t careful they can inadvertently introduce into the product the bias and other hazards that can come with large language models.

As a content designer working on Microsoft’s Copilot for Sales product, Leah Krauss helps her colleagues understand and follow the responsible-AI principles that the company has developed.

Leah’s advocacy helps her design and product teams create a product that balances the need for transparency about the use of AI with the prerogative to keep customers in flow as they use the product.

We talked about:

• her work as a content designer on Copilot for Sales at Microsoft and her advocacy there for responsible AI
• how she collaborates with her data science team, which had established a relationship with the content team even before Copilot on other products
• the evolution of their AI product-development process
• how their design system supports the implementation of responsible AI
• the six principles that guide responsible AI at Microsoft:
  1. fairness
  2. reliability and safety
  3. privacy and security
  4. inclusiveness
  5. transparency
  6. accountability
• how she advocates for responsible AI on the Copilot for Sales product team
• the balance between keeping customers in their flow and being transparent about AI features
• the concept of the “human in the loop” and how they apply it in the Copilot for Sales product
• the importance in AI product design of always being aware of edge cases and possible misuses of the product
• her encouragement to anyone working on AI products to stay curious, ask a lot of questions, and to bear in mind the importance and relevance of our language expertise

Leah’s bio

Leah Krauss is a senior UX content designer at Microsoft. She works on Copilot for Sales, Microsoft’s AI software for salespeople, where she also collaborates closely with the data science team. She champions responsible AI to anyone and everyone who’ll listen, including inside Microsoft and at various UX conferences. Outside of work, you can usually find her reading, or spending time outside with her family – hiking, exploring cities, and hanging out on the beach.

Connect with Leah online

• LinkedIn

Video

Here’s the video version of our conversation:

Podcast intro transcript

This is the Content and AI podcast, episode number 31. The introduction of AI tools like Microsoft’s Copilot creates new opportunities for content designers. But as with any innovation, the new technology can be a two-edged sword. For every customer workflow that is streamlined there may also be an opportunity for bias or hazard to get into the product. As a content designer and champion for responsible AI, Leah Krauss helps her colleagues at Microsoft understand and apply responsible AI principles in their product design work.

Interview transcript

Larry:
Hi everyone. Welcome to episode number 31 of the Content and AI podcast. I’m really delighted today to welcome to the show Leah Krauss. Leah is a senior UX content designer at Microsoft where she works on Copilot, which many of you may have heard of. So welcome Leah. Tell the folks a little bit more about what you’re up to these days.

Leah:
Hi Larry. It’s so nice to be here. So yeah, as you mentioned, I’m working on Copilot for Sales, which is a flavor of Microsoft Copilot, and that’s been really exciting to be in on kind of the ground floor of AI at Microsoft. And responsible AI, which is what we’re going to talk about today, is one of my most favorite topics to talk about. I’ve done some conference talks about it and my coworkers are really tired of hearing me go on about it. I actually serve, no, that’s not true. It’s only half true. I serve as actually a responsible AI champion, one of the responsible AI champions on my team. So it’s sort of my thing and I think it’s so exciting the moment we’re at here and how big a part content designers can play in it.

Larry:
Yeah, I think when you’re working with language models, you think the content people would have a leg up on some, but that’s really, first thing I want to follow up on is that Copilot, I think it’s sort of like multiple products then. It’s a suite, well I guess Microsoft knows about suites with the Office suite, so there’s integrations of copilot with each of the Office elements, Word and Excel and all that. And then there’s also specific tools like Copilot for sales. How big is that little budding empire at Microsoft?

Leah:
Well, so there yeah, like there’s one Copilot and then the different flavors depending on the user’s needs. So as you said, if people are using Office, they’ll see Copilot in Word and they’ll see it in PowerPoint and Outlook of course. And if they have more specific needs like a salesperson, then they can use our more specific flavor, which has the kind of email summary that a seller might need. Did my customer mention the budget? Or things like that. And it can also pull from other sources so that the seller can have everything they need right in that one place. Copilot is a big thing at Microsoft, as you know, and I think it’s only going to get bigger in the next couple of years.

Larry:
Yeah. As you’re talking about it too, I’m reminded of how this thing is coming together. It’s like there’s this one, I assume it’s based on one of the GPTs from OpenAI given a Microsoft relationship there. But it sounds like then that most of the applications, each of these flavors of Copilot, is a lot of your work around fine-tuning the model then for that specific task?

Leah:
Absolutely. I work really closely with our data science team and the first thing you sort of have to, there’s the GPT prompt, but then we also write our own prompts on top of that. So we tell the GPT that this is a selling audience, for instance. And I’m really involved with working with the data science team to define what a good output looks like because while they’re the experts in the model and how to create an output, I’m the expert in what makes it good and what makes it human and what makes it useful to a seller and scannable and valuable and things like that. So the great thing about this project has been that we get started early on working together. As we content designers know, sometimes content people are brought in too late. And that is definitely a danger here too, like for people who are listening, you can always do something with the data science team, but if you’re there from the beginning and you’re talking about the prompt together, then you can move forward and really have an impact.

Larry:
That’s really interesting because there’s all these new collaborations that are emerging along with these AI tools. So you’re working, is it mostly data scientists? I’ve talked to other folks who’ve worked with the machine learning engineers and other new collaborators. What does the team you’re working with look like these days?

Leah:
So at Microsoft it’s called data science. At other places it may be called machine learning, but it’s the same group of people.

Larry:
Yeah.

Leah:
There’s the people who care about the algorithms, basically we can call them. And then we have the people who care about the words, which by the way is not only content designers, it’s also product managers and interaction designers, but content is really leading the way.

Larry:
Yeah, that’s fantastic to hear. And what you just said too is it sounds like there’s more opportunities or is there more opportunity or have you just made it happen to get in earlier or do data scientists see the need to involve word people earlier on?

Leah:
Well, we’ve been working on AI features and machine learning features even before there was Copilot and even before I joined the team actually. So my manager who has been on the team longer than I have was working with the data scientists when we had a feature called Conversation Intelligence. And what that did was when you would record a meeting, a seller would record a meeting with their customer, then Conversation Intelligence would analyze afterward and give sort of a recap that has the main action items and a bulleted list of the highlights of the meeting. So my manager whose name is Erga Herzog, and she was the one who in the sales organization really built that relationship with the data scientists. And also I work with a lot of other content designers too. So basically I was lucky to come in about two years ago with this really strong base that was already very far along.

Leah:
We had a good relationship with the data science team and there was already that conversation was starting. So what we’re doing now is we’re sort of trying to formalize that process because sometimes it relies on the individual content design or the individual data scientist to decide when we start talking about a certain feature. And we’d rather have it be more formalized into a process that like, okay, at this point we start talking and then at this point we start looking at sample outputs and maybe we first decide together along with the PMs and the rest of the product squad and the leadership of course, what we want out of this particular AI feature. So that’s something that we’re working on right now, which is really exciting.

Larry:
That is exciting because this is also new and at some point you have to, like it must be, I can only imagine the pace of work there, but so being far enough into it to kind of step back and go like, oh, hey, this is how sort of the routine way that we, or not routine, but the common way that we do things. Is that sort of, do you think that’s common across the other flavors of Copilot or, like because you mentioned you were talking to your peers as well as your immediate colleagues. Is there sort of patterns emerging around how those, like you just said, when to start the conversation, even when to formalize things I guess.

Leah:
I don’t know of specific processes, but the way things at Microsoft often work is sometimes it bubbles up organically and sometimes it comes from top down. So I think in the end we will have a much more unified process, but it’s very scrappy right now. So it’s hard to say exactly what the other teams are doing. But we do have a design system and the content designers are also involved in the design system. And in a lot of ways we’re sort of bringing up those things that made, like for instance, a pillar of responsible AI, which I know we’re going to talk about in a minute. But one of them is the transparency.

Leah:
So we want the user to know what’s going on. We want the user to know that the AI is being used and how it works. So our design system work often involves saying we need to be able to have a tool tip, for instance, that can have a link to the documentation because we want to use progressive disclosure and we don’t want to put everything in the tool tip. We want to have this. And so it’s like technical conversations like that, but they start from a responsive AI need for instance, or a content need.

Larry:
Yeah. That notion of, this is a good time to pivot to talking specifically about responsible AI because I want to talk about the principles that underlie it, but it sounds like once you’ve got some guidance around responsible development in these things, that you have places for it in the design system. So that seems like a really good formalization of things. But let’s back up a little bit and if you can tell me, because you mentioned before that you’re sort of an internal champion for responsible AI. And tell me, well, first tell me about the principles and then about your role as a champion for them.

Leah:
So your listeners I’m sure are very smart and very tuned into what’s going on with AI and content and all that. So they probably, I think we probably all read a lot about what’s going on and think a lot about, like hear these stories about what can go wrong with AI. AI these days can affect whether you get a job, whether you get a mortgage, whether you get a date, and even whether you go to jail. There was a story a couple years ago about a man who was wrongly arrested and he spent a few days in jail before he was finally released, but he was wrongly and matched by an image recognition software and he happens to be black and image recognition is less accurate for black people still. So there’s things like that. AI has a lot of potential to do good, but it also has a lot of potential for harm.

Leah:
And we hear about deep fakes or voice cloning and all that kind of stuff. And so there’s real value in thinking all along the way about how we can prevent that harm as much as we can and make sure that we’re building models that don’t perpetuate the kind of biases that we already have in society. So Microsoft, I’m lucky to be part of an organization that not only has already thought about the principles of responsible AI but is very open about sharing them. So I’ll share the link to the Microsoft responsible AI page, has a lot of resources that anybody can use in their jobs. And there’s six principles that have been identified as being the most important for keeping in mind. And I have them written down, so I’m going to read them.

Larry:
Excellent.

Leah:
There’s fairness. Yeah, I don’t have a tattoo of it yet. Reliability and safety, privacy and security, inclusiveness, transparency, and accountability. So when we think about all those things together and there’s more involved definitions of each, and when we try to bring those into our daily practices, that’s how we create the most responsible AI that we can. And so what I do is I’m sort of the voice that is saying like, hey, what about this edge case or what about that edge case? And I’ve gotten in the habit of trying to think of what those could be and bringing those up when we can.

Leah:
And another thing that I do a lot is I go through a lot of output. Before we release things, I’m reading a lot of tags, or generated emails, or generated summaries, or things like that. I’m reading a lot of output, and it’s kind of like at some stage I think that we’ll be able to rely on AI checking itself and filtering itself, but I don’t think we’re there yet. I think we really still need people to go in and sample and really look at these things. And I just bring things up. And so there are other responsible AI champions on the team who have also a very formal process of when a feature is released, we have to go through a checklist and make sure that we’ve discussed and defended different things to make sure that they meet the criteria of responsible AI.

Larry:
The way you just said that reminds me of accessibility, that like I know that there’s a checklist at the end, but I’m assuming it’s not just a checklist at the end, that you’re trying to bake it in at the beginning of the process as well. And I guess tell me as a champion, is this mostly around your fellow content designers or with the data scientists you’re working with, or how big is your championing soapbox, I guess.

Leah:
My personal one is within the Copilot for Sales product squad, but we have official champions who are in the entire business unit and that are working on all different products and they get to see across what’s going on with AI in the business unit. So it’s definitely a position that’s taken seriously within Microsoft. And it’s actually, it’s good that you brought up accessibility because it is similar in that we have to get used to thinking of it and get used to remembering it and get used to baking it in along the way.

Larry:
Now I’m wondering, I don’t think I’ve ever seen it, but I think I have it in my head of what that list of principles for accessibility looks like. It’s probably very similar actually. You’ve probably internalized these principles at this point, but how do they manifest in the product? For example, accountability. That was an interesting one. That’s the sixth principle you mentioned. Is there a mechanism for accountability? I’m just thinking like if Copilot gave bad advice, is anybody accountable for that or do you just because of the nature of the tooling, do you just have to say, look, we don’t always get it a 100% perfect? I guess how do you balance the principles?

Leah:
Well, I mean, we don’t promise that Copilot will never mess up, honestly. And we have an AI disclaimer on our generated text just like GPT does by the way. If you’ve seen chat GPT, it has a disclaimer saying, just so you know, this ChatGPT might make something up. And so on our output, which in Copilot for Sales, it could be a summary of an email or it could be a sales email that is generated by Copilot or a recap of a sales meeting. So on all of those, we have the disclaimer, which is that Copilot is just a helper and you still need to check and make sure that it’s really the information that you want to be sending to your customers or to your boss. We also have feedback like thumbs up, thumbs down, and send us feedback, which we read and which we act on.

Leah:
And as far as accountability, our main premise there along with transparency is that we want our users to be empowered with as much information as possible. So we have a team in my business unit that’s called the Human AI Interaction Team, and it’s led by a researcher named Jennifer Wang and there’s a bunch of researchers there, and they are in my organization, they’re the ones who are very focused on transparency and accountability. And when I joined the team, I was very pleasantly surprised to see that we were telling the user exactly when we were using the AI, exactly how it works and what they can and can’t expect from it. So one of my first projects in Copilot for Sales was to write a pretty detailed set of questions and answers. What is this? When do I use it? How does it work? What can I expect from it? All that sort of stuff. So I think that that is empowering users to be savvy consumers of AI I think is very important.

Larry:
No, the way you just said that, you reminded me of a couple of things that have come up in other conversations. Noz Urbina and many others have made the point that these GPTs are like interns. They’re really smart and clever and can figure stuff out, but they’re not as experienced as you. And that’s sort of, I’m assuming that’s part of that disclaimer is like, hey, look, you’re still the salesperson here and it’s up to you to take this information and tailor it to your professional style and things like that. What was the other thing I was just thinking of? It’ll come back, but that’s really that point of sort of empowering the users because I guess this is all so new. I guess, do you feel like that’s the other thing that’s come up that I was just thinking of, is that to this point before the arrival of GPTs, you could pretty much count on computers to always do the same thing. They were pretty reliable computing machines.

Larry:
And now the very nature of Transformers and GPTs is that they’re predictive models and that they’re not always based in fact. There must be different levels of championing, I’m guessing because there’s both in any one product making sure that people understand what’s going on at the time. But do you feel like an obligation as a champion to inform people like, hey, this is a different kind of computing we’re working with now and you need to develop things differently or?

Leah:
That’s an interesting question. I think that there is definitely a place for that. It’s hard though because you want to balance the user being in their flow and getting their work done with empowering them with this knowledge. When someone just wants to write an email and get it over with, they want to make sure it’s correct, obviously, especially if they’re sending it to a customer, which is the case for our users, but they also just want to get the email done. So one of the things we think about a lot is what’s the balance between education and keeping people in their flow? And that I mentioned earlier, we use Progressive Disclosure a lot for that.

Leah:
So we may have a line of text in the UI, some microcopy that says, use Copilot to craft your emails faster or use Copilot to get a detailed recap of this sales conversation. And then the next level of disclosure is a tool tip with more information, and then it may have a link inside to the documentation which will give you really a detailed description. So what I found is that people who are using Copilot for work are interested in how it works, and it’s good to have that explanation there for them when they’re ready for it. We just have to make it accessible but not interrupting their flow of work.

Larry:
Yeah. And you just said it’s like, yeah, the whole point is to get stuff done. And that notion of both progressive disclosure and just sort of not being an intrusive force in their life, that’s an interesting balance and I think it’s good, but it also reminds me that’s sort of like there’s this notion that has, I think a fairly specific meaning in the machine learning world of the human in the loop, like in reinforcement learning with human feedback and other kinds of human stuff. And you’ve mentioned also that one of the primary things that drives the need for responsible AI is people, biases and assumptions and sometimes not good motives. Can you talk a little bit about just that notion of, is that something you think about, like that notion of the human-in-the-loop and how does that manifest in your work?

Leah:
Yeah, absolutely. So we definitely use human-in-the-loop for some of our features because as I said, these are people who are using AI for work. The stakes are really high and we need to be able to give people a way to correct if the AI is wrong or fine tune things. So we definitely use the principles of human-in-the-loop for the users. I’ll give an example. We have a project that I worked on was a sort of enriched summary of a seller’s email where we also pull in information from their CRM, so like Salesforce or Dynamics 365, which is where they keep all their customer information and they don’t normally have automatic access to it when they’re in email. But it was helpful, we found in research that it was helpful for people to have a summary that was rich and had pulled things from both.

Leah:
But we had to be able to let the user choose a different opportunity, let’s say from the CRM in case the one that we were proposing is not right. So in that case, the user is in the loop of creating the summary. And in other cases we have, for instance, if the user is generating an email, they always have the option to edit it in Outlook before they send it. They always have the option to start over or change it. So there’s that. As far as human-in-the-loop in the overall product development, that’s also something that we think about. As I said earlier, I go through a lot of output and sometimes it can feel like a lot of grunt work at the time because the data science team will say, “Here’s 50 emails that we generated because we’re testing variants of the prompt.” And so what I’ve been doing, and which I really also warmly suggest to anyone who’s involved in AI features is to roll up your sleeves and to sit down and spend an afternoon just reading through them.

Leah:
Because you can, in the best case, you say, okay, we can tweak the prompt and make an even better output, but other times you’ll catch things. And so that’s definitely, like that’s how we do human in the loop in the product development process. But as for you were also talking about bad actors and thinking about what kind of things can go wrong with our AI. So this is something that I think is really, really important for people who are designing AI products is to get in the habit of thinking of these edge cases. Like, okay, I’m a well-behaved content designer. I don’t want to do something bad with this product, but if I were a troll, what would I do and how would I try to break it and how could I cause harm to someone? Or how could I use this product in a way that is not what Microsoft wants them to use the product for?

Leah:
So it’s something that didn’t come naturally to me, but basically I started by reading as much as I could about things that went wrong with AI. And that sort of gave me the basis for like, okay, this is how people have misused AI in the past, or this is how the model went wrong. And then you just decide to do it. So just like when you need to write something, you sit down and you decide, and then you start brainstorming and it comes. So for instance, if you are, well, I’ll give you an example. So I work on Copilot for Sales. And at the beginning when I was saying, okay, people are at work, like are they really going to behave that badly? Are they really going to try to misuse this? But then I did a little more research and I found out that, so people who are doing call center sales often experience a lot of abuse from their customers. It’s terrible. They call these people and the people are rude or they’re abusive.

Leah:
And I was thinking, well, we’re telling sellers that they need to record all our phone calls and then read the transcripts later or read the recap. We don’t want them to read this abusive tirade that someone has said to them. So that’s just one example of how a person can misuse it. Another example that I came across was, I was interested in Midjourney and DALL-E and all of the image generation tools, and this was I guess about a year ago now. So I was using Midjourney to just generate some images like my kids were home and we were playing around with it and stuff. And the interesting thing about Midjourney is you can, like I am sure by now lots of people who are listening have used it, but for anybody who hasn’t, you are in a chat room and not only can you see the prompt that you have asked for for Midjourney, but you can also see what other people have asked for and the images that are being generated.

Leah:
And so that was how I came across an image of a man who was sitting and typing on a computer in front of a pyramid. It was like a drawing, and he was supposed to look like an ancient Jewish person. And so I wanted to see what the prompt was, and it was like an ancient Israelite sitting typing in front of a computer in front of the pyramid, something innocuous. It caught my eye because I’m Jewish. And I thought about how does this make me feel? Maybe it’s fine, but maybe whatever, it’s nothing. But because I had started in this habit of like, okay, let’s think about how this could go wrong. So I figured if there are people who are asking for innocuous pictures of Jewish people, then maybe there are people who are asking for antisemitic depictions of Jewish people, and you can do a search in Midjourney, which I did, and I very quickly found terrible things.

Leah:
So all of that is to say, once you get in the habit of looking for what could go wrong, it gets easier and you start to say, oh, I can see this and I can see that. So nothing like that thankfully can happen in Copilot. I’m glad I don’t have to deal with that in my work life. But we do have to sort of decide. We have to decide, okay, someone’s going to ask for, is going to type an inappropriate prompt for their email or something that we think is inappropriate, and we have to decide what is the experience going to be.

Leah:
And in the end, what happens is we show them an error message and it says, this prompt may be inappropriate or something like that because first of all, we can’t be 100% sure that what they’ve typed is really inappropriate. We do err on the side of caution and there’s actually a bunch of fail safes that scan, stuff like that. But this way, if the person has typed something that’s, we don’t want the Microsoft name on, then we know that we’re not going to have a part. And Copilot’s not going to have a part in generating that inappropriate email.

Larry:
Yeah. And I love the way you articulated your role earlier as the well-behaved content designer versus the trolls, and thanks for all your help on that because those examples you gave too are just, that’s just typical human stuff.

Larry:
But hey, Leah, I can’t believe it. We’re coming up close to time, but before we wrap up, is there anything last, anything you want to revisit from the conversation or just want to make sure you share before we wrap up?

Leah:
I would just encourage everybody who’s working on AI stuff to get really curious and ask a lot of questions and never be afraid to ask questions of your team members and data science and all of your colleagues. Because as we know from working with, in our history of content design, we’ve sometimes been told that it’s too technical or it’s not ready for content to be involved yet. And we know that we can handle the technical stuff and we can handle it and we can ask the questions. And so just do those things that we’ve been doing this whole time and that tool that we have in our toolbox. Because the thing about AI that’s so exciting is that so much of the input is words and so much of the output is words, and that’s our thing. So we really have the chance to make a big impact here.

Larry:
Nice. I love that, reinforcing that point that this is, a lot of people are worried right now, but reinforcing the point, like no, lots of opportunities for words people here, so thanks. Oh, and Leah, one very last thing. If folks want to connect with you online or follow you, what’s the best way to find you online?

Leah:
I think LinkedIn is best. I’m happy to meet people and talk to people there.

Larry:
Great. Well, thank you so much. Really enjoyed the conversation, Leah. Thank you.

Leah:
Thank you so much. I did too.