I'm Harshita Lal. I'm gonna be talking today about AI systems and trust. I built software products.I've been a student of psychology literature, and I I think some of these concepts are very pertinent in today's world as we are building AI products. And I wanna talk to you about how how I'm a big believer in AI. I think this is one of the biggest changes that our generation will see and perhaps one of the biggest, changes in the history of software. And I wanna talk about as we build the next decade of software, how should we incorporate trust into it. Let's get into it. Alright. So just couple of thoughts before we dive in and understand the framework of what I'm mentioning over here.

So this is, you know, a very typical scenario we see when we're shipping products. We're in a rush to ship. We ship something. It's the same product. Engagement is high. Our leaders are happy. We're seeing good results. Six months later, we see that trust tanks. Churn happens. Our support tickets talk about things like the product was pushy. It was creepy. It was wrong, but sure of itself. And net promoter score is down. So what what happens when when this happens? Why did this happen? The model was set up correctly. The system did what it was designed to do, but it was not optimized for trust. It was optimized to, it was set up to optimize outside of trust. And that's the core of the problem that I'm going to address today, which is when we're optimizing, why aren't we optimizing for trust, and what happens when. Alright.

So the framework of today's topic is in three portions. So the capability of the model, which is, you know, where your tech tech teams live, how smart the, system is, how technically sound it is. Then second, business incentives. This is where product and growth teams live on what we optimize for. Right? It's engagement. It's conversions. It's the top line metrics that we report on to our leadership. And then the third one is human behavior. This is where all of us live. Real people respond over time to how we we interact with software. And the funny thing is as we're using AI more and more, no team or organization instantly owns this in in an enterprise. So we need to design for this third force.

And when we build for behavior and how human behavior revolves around these other two forces, that's when we'll build for trust. Alright. So getting into the meat of the concepts here, like I mentioned, psychological concepts that have trained, human behavior over time and has been studied deeply. So the first concept here is variable rewards. So this is about unpredictability in the reward you're gonna get, and this is one of the most powerful behavioral hooks we know of. So in nineteen fifties, a psychologist called BR skin Skinner had, come up with an experiment where he had a lever which gave out food to some pigeons in a box. So these pigeons would peck sort of comfortably when there was a predicted time, say, PM when the lever would open and they would get the food.

But when this lever opened at unpredictable times, as in no pattern, then these pigeons would want all of it. They would go nuts over when to peck. So Skinner realized that unpredictability is actually leading to compulsive pecking, And this at its heart is how unpredictable reward has become the most powerful human hook that we know. So think about all the things you get addicted to, slot machines, your social media, and today, AI products. Things like regenerate recommendations, things that give you right results sometimes but sometimes don't. So this is a great tool for engagement and and retaining engagement. But where it backfires is when compulsion replaces preference. So the reason this happens is because users don't come back because they value the product. They're coming back because that they can't predict it. That predictability is building for the wrong thing.

So it's the same pattern, but in in this interaction of this graph I'm showing here, you have to build for that interconnected red and green line where engagement and trust overlap. So when you don't do that, what happens? In commerce, we see users churning after several bad recommendations. In chat, you keep asking the LLM the same question and you see the same wrong output, and you're realizing that, okay, it's just not giving me the right thing. In retail media, which is my world, we see advertisers not trust the auction process. Trust gets lost. Alright. Concept two, authority bias. Very, very common where we're looking at how confidence and correctness are under uncorrelated. A lot of times, we it has been studied rather that, humans react really positively to, to authority.

So think of a very natural reaction to, you know, uniforms, to titles, to a confident tone and tenure. These things inspire confidence. The problem is that, because of the uncorrelatedness of fluency and correctness, we don't see always that LLMs are authority machines. Just because the tone is crisp, the wording is very professional, and the cadence is like an expert doesn't mean that it is an expert. However, every signal that humans get in this condition is that it is. So what happens when this correlation goes wrong? Right? You ask it a question and it hallucinates, and this is where hallucination can be quite dangerous. You get a completely fabricated answer because it's pulling from the wrong place. And if you're not an expert on what you're searching for, you may trust it because it's talking with such authority.

So in commerce, this showed up shows up as a bad recommendation delivered with conviction. In finance, this could be a pretty bad wrong tax read that could have those implications, and it scales up from there. Right? In health care, you're looking at a completely wrong diagnosis. In an enterprise, it could change your entire code base. So where it backfires is that trust fields here, and then it has a cascading effect. You have a verifiable domain where the user can check something. So see that in this example I was giving about health care, I'm a doctor, and I do realize that the LLM is giving me the wrong answer. At this point, the user can check, and I am now not only going to not trust the answer I'm getting right now, I'm also not going to trust any of the other answers I've gotten thus far.

So this loss of confidence in one answer is gonna lead to a loss of confidence in all the prior answers and the answers henceforth. And this is a very dangerous thing because, you know, I wanna be clear. AI is very powerful. You want to be able to use its capability, but you need to trust it. And that's the problem with authority bias. Once trust is gone, it's very hard to get it back. Alright. Third concept, choice overload. I love this picture because I think it conveys a lot. A lot of times, options feel very generous, and it feels like we have more autonomy, more satisfaction. But the reality is that's not always a good thing. It can lead to a lot of decision fatigue and lower conversion. Think about you walking into a grocery store and seeing all these options for serial.

And sometimes it's good. Initially, it feels good, but when it actually comes to picking things, you're like, there's too much going on over here. That is a design fix that can be made when you think about software development. So, again, going back to next experiment, these were two psychologists that made went to a grocery store for an experiment. So they took they set up a jam table, and they took several varieties. A few days, they set up a jam table with 24 varieties of jam, and another few days, they set up only with six. So they assume that, you know, the 24 varieties of jam meant larger tables, more footfall, and they did get a lot of footfall.

But it turned out at the end of the experiment that the people who converted more, who actually bought the jam, were actually the ones at the six variety table. So what does this indicate? Right? It indicates that more options aren't always, the the reasons why conversion increases. And AI does this mistake constantly. Recommendations give you several personalized picks. Everything is personalized, and yet it feels like you're being overburdened with choice. Generative tools will give you several variants, and ask you to choose. Enterprise tools can have several different suggested next actions. And at the end of it, you feel like you have actually worked a lot more than you initially did had you not used these tools. And that's my point. It goes against the whole reason AI and these tools exist to help us to help us be more productive, and that's not happening because of choice overload.

So let generosity not create friction. Right? And when you do overload overload a user with options, what happens? You're a user. You're getting all these choices. You're either going to abandon and just close the tab and not not deal with it. Second, you might actually take the first option you see and think that all the other personalized picks are too many. Or you actually pick something and then you second guess your choice. You don't trust the system. The point here is personalization is not a buffet. It should be a thoughtful suggestion, and that is one of the key takeaways I want you to take from this session. So all these three cal concepts of variable rewards, authority bias, choice overload are summed up, I guess, the the heart of it is a concept called reactance, which a scientist called Brem talked about in the late sixties.

What this is is when you have, a human being who understands or realizes that their agency is or autonomy is being threatened, they will push back. And sometimes they don't push back proportionately. And that reactance is the heart of all these three types of concepts. So variable rewards, trigger reactance when the user realizes that the system is trapping them, hooking them. It's a fight or flight response. Authority bias is when, a user caught catches a confident system being wrong. Choice overload is when the user realizes that the right choice is hidden somewhere, and it's too much work to get to it. So now how to make sure that you are building for reactants because that's the heart of the problem? The answer in one line to that is trust.

Trust is built across hundreds of interactions, but it only takes one bad one to break it. And that's why trust is not a metric that you can AB test your way into. So as promised, a framework to help you figure out how to design for trust. So for against variable, rewards, you wanna build in constraint layers. And this is a fairly, pointed, note for product managers out there, designers out there. When you're optimizing for task completion, you need to optimize not for session length, but task completion only. That ensures that while your top line metric, your north star is engagement, you're looking at that. And if you see that compulsion is staying flat, you're building compulsion.

The completion is staying flat. You're building compulsion, not value. And that's happening because trust is not ingrained into this method of, pushing against variable rewards. Incentivize for trust. Measure yourself for both completion and trust. Second, authority bias. So, obviously, you can change the way that AI talks, at least not to a very large extent. But you can design for how your UI looks, meaning that you should get intent into your UI. What that means is that when you calibrate uncertainty into the UI itself, not just the model, if the model is unsure, your UI will reflect that. And when you have that, the user clearly can see that this may not be the right answer or may need to do extra checks, which is sometimes the recommendation that the model is making, but the UI is not reflecting. And in this way, you can hedge yourself against the risk of these high stakes actions getting friction and making sure that the user feels more control.

Three, against choice overload. Calibrate your choices. Maybe we don't need all these many personalized choices. Perhaps defaulting to a three choice option with one very clear primary recommendation is the right way to go. State the reason why you think this is the primary recommendation on the UI and let the users expand for more. Less is more. You know, age or adage, I guess, it applies here as well. Just keep in mind as you're designing that personalization is a thoughtful suggestion. It's not a buffet. That's not the point. So in summary, just to pull it all together, if you've not heard anything I've said so far, just remember, trust is the only mode in AI that's gonna compound.

You're all working with AI. You have teams that are designing, with AI, designed for trust deliberately. As you do that, you're going to be balancing the this capability incentive and behavior triangle. And if that is done correctly, we will define the next decade of software with trust at the center. And that's the good news. These these design moves aren't complicated. They're a choice we are supposed to make, and we are all here to make that choice together today. Thank you for listening. I hope I'm doing good on time.