I recently traveled to Ireland with my family. It was one of those trips where everything felt slightly off-kilter: the jet lag, the time difference, the way the air smelled different. But the thing that caught me most off guard had nothing to do with any of that.
We were walking around Dublin, and we got to a crosswalk. I pushed the button and waited. The red hand turned to the walking symbol, and I knew I could cross. I looked to my left to make sure no cars were coming, stepped into the road, and saw it: LOOK RIGHT, written in large bold letters on the ground.
Oh right. I’m in Ireland. They drive on the other side of the road. I already knew that, but the behavior of looking left was so ingrained that I needed those big letters at my feet to change it. As I kept walking around Ireland over the next few days, though, I got used to it. The labels blended into the background. I no longer needed the reminder. Looking right became natural. And we all had a great and safe time; nobody got hit by any cars.
Internal representations of how things work
Mental models are internal representations of how things work, or more specifically, what we think will happen when we use or interact with something. They simplify complexity, creating frameworks that let us understand things more quickly and make decisions.
We form mental models through living our lives: personal experiences, education, social interactions, cognitive processes. Because we’re all different, we all have unique mental models of the world. And they’re not fixed; they evolve over time, especially when our actual lived experience differs from what we expected.
When what actually happens aligns with our mental model, things feel intuitive. When there’s a gap, we get friction. And right now, there’s a lot of friction in how people interact with AI.
Iconography already does this
Iconography is a great way to see how designers already use mental models in their work. If you look closely at most icons, you’ll notice how much they resemble real-world items. This is a technique called skeuomorphism (incorporating real-world elements into digital interfaces to create a sense of familiarity with new technology). The reason we do this is that it’s very hard to understand technology. If we pick an arbitrary button, we have to teach users what that button means, versus using a real-world representation.
Take the send icon. It has a clear, direct connection to a paper airplane. And a paper airplane is itself an abstraction of an actual airplane. The technique works because the icon conveys so much meaning (movement, going from one place to another) just like what happens to our messages when we hit send. That’s far more powerful than a set of lines put together to form a shape.
We’re trying to do something similar with AI, but we don’t have the same real-world connection. The closest we’ve landed on is “magic” (hence the sparkle icon that the industry seems to have agreed on).
But our research at ServiceNow shows that a decent percentage of people still don’t know what that icon means. And since this technology is so new and the mental models are so fresh, iconography alone isn’t enough to give users a full understanding of what AI is, how it works, and how they can best use it.
We need to understand all of the mental models users bring with them. There are many mental models involved in AI; conversation is just one, but it’s a big one.
The chat window made a promise
Before diving into what people expect from conversation, it’s worth asking: why did AI adopt the conversational interface in the first place?
I find it helpful to use dictionary definitions when understanding confusing concepts. Merriam-Webster defines conversation as an exchange or discussion. It’s not a one-sided interaction. Both Google and ChatGPT let me find information. So why would I choose one over the other?
The definition gives a clear indication: in conversation, we can interact with the information. We can discuss, go deeper, and explore the topics we want to further understand. The chat window invites something a search bar doesn’t: a dialogue.
But by choosing conversation as the interface, AI products inherited all of our conversational mental models. The chat window made a promise that the technology now has to keep. And people approach conversations with AI carrying the expectations of human-to-human communication. Most people know they’re talking to an AI, but despite that understanding, they still bring their expectations of talking with humans. The only back-and-forth conversations we’ve ever been able to have are with other people, so it’s all we know.
So what exactly are those expectations?
What people actually expect when they enter a dialogue
We expect the other person to pull their weight. From H.P. Grice’s Cooperative Principle (1975): contributions should be complete, correct, understandable, relevant, and presented in an appropriate tone.
Each person contributes during their turn and actively listens during the other’s. We don’t expect monologues. We expect digestible chunks we can react to.
We expect the other person to infer intended meaning using context cues, so we don’t have to be overly descriptive. This is pragmatics: meaning through implication and inference, not just word choice.
We expect the other person to keep track of what we’ve said, not just within a single conversation, but over time. ChatGPT’s memory feature is a good example of this implemented transparently.
When we enter a conversation, we expect to be heard, understood, and respected.
What happens when these models get violated
Let me show what happens when these mental models get violated. I decided to ask ChatGPT how to start working remotely from another country, something I was curious about after our Ireland trip.
When I asked a friend the same question, I got a short and sweet response. Friendly tone. On point. A follow-up question to help narrow things down.
ChatGPT gave me paragraphs. Technically all true, understandable, and mostly on topic, but long. I didn’t want or need all of that information. It put the work on me to figure out what was applicable.
It didn’t take turns with me; it gave me everything up front rather than letting me guide the conversation. And when I followed up with “I’m interested in Portugal,” the system lost the thread entirely, responding with something like “Portugal is a beautiful country with a rich history! Are you thinking of visiting?” completely missing that we were talking about remote work.
That’s cooperation, turn-taking, context, and pragmatics all violated in one exchange.
Here’s the thing, though: when I told ChatGPT to respond like a human, the answer looked pretty similar to my friend’s.
The system is capable of responding in a way that matches user mental models; it just defaults to over-informing.
Users approach AI with a lot of baggage
It’s an uphill battle, because users approach AI with a lot of baggage. Their expectations are rarely accurate; they more often think the system can do far more or far less than it actually can.
For a long time, people had very low expectations of what AI could do. I’ve spent a lot of my career convincing people that chatbots are not as bad as they think. But honestly, for a long time, a lot of chatbots were truly bad. They wouldn’t understand us, could only do one thing at a time, and gave inaccurate answers.
On the other hand, when we hear the phrase “artificial intelligence,” we often think the system is going to be incredibly intelligent and operate at a very high capacity. So when it can’t accept “tomorrow” as a date and asks for the format yyyy-mm-dd hh:mm:ss, the gap between perceived intelligence and actual rigidity is jarring.
Make systems cooperative, take turns, remember, and understand context
The first step toward improving user perceptions is to understand these mental models and align our products with them. We need to make sure our systems are cooperative, take turns, have memory, and take context into account.
I like to walk through what this looks like in practice. Take that same remote work conversation, but designed with these principles in mind. Instead of dumping everything at once, the system starts by asking a clarifying question so it can send a more informed, concise answer.
Suggested:
• I have a country in mind
• Help me explore options
• I just want a general overview
In just a few turns, the AI and the user have honed in on a location that’s a great fit. That’s the power of aligning with how people already expect conversations to work.
Don’t just match expectations. Push them.
By now it should be pretty clear how much better conversations can be when we align with mental models. But we can do better. I want us all to push further. We need to figure out the mental models, design products around them, and then use our expertise to push users forward. This takes experimentation and research, but no innovation happens without pushing boundaries.
Instead of just identifying what users expect, we need to understand their underlying needs and goals. And really, it comes back to being heard, understood, and respected.
The collaborator model
One emerging mental model that resonates with me and excites me is the collaborator model. This views AI as an equal partner, honest about its shortcomings, but doing a lot more than what people have previously expected. In the collaborator model, AI works alongside humans to enhance decision-making, creativity, and problem-solving. It focuses on what AI does better than humans and how we can harness that.
Going back to that redesigned remote work conversation, the collaborator model shows up when the AI goes beyond just answering questions. It starts offering contextual recommended replies. It consolidates information so the user isn’t overwhelmed. And then it goes further, offering to create a detailed to-do list, or even help start a visa application.
The visa process has a few moving parts (proof of income, background check, accommodation, health insurance). Want me to put it all together?
• Create a to-do checklist
• Start my visa application
• Week 1: Confirm employer approval for remote work
• Week 2: Request background check (takes ~3 weeks)
• Week 3: Open Portuguese bank account, get NIF
• Week 4–5: Secure accommodation, gather income proof
• Week 6: Schedule consulate appointment
• Week 7–8: Submit application
Add these to your calendar?
I want to highlight how much the AI is going above and beyond human-to-human conversation at that point. Most people wouldn’t expect that level of involvement from a person on the other end of a chat. As we build out more automation and capabilities in conversational AI, we may move away from thinking of AI conversation the same way we think of human conversation. For now, we’re still functioning on the basis of chatting with a very smart, very capable person.
The best way to keep up is to test a lot
Every day I hear new stories about AI. Things are changing at such a rapid pace. The best way to keep up is to test a lot. Always ask your users what they think will happen, in tech, in AI, in your products, in conversations.
This is something I do in my job at ServiceNow. We have a very thorough research process on our AI projects because we know there is so much change and flux in the industry. Whenever we talk to someone, we make sure to understand their mental models without prompting, to see what they are before we show them our products, and then to see if our designs are aligning with, bumping against, or possibly helping to shift those mental models. If the user’s view aligns with the model you want to promote, you’ve done a great job. If not, you’ll need more nudging.
Ultimately, we all need to do what we can to understand and contribute to the shifting mental models of AI. As we build structure and guidance around AI systems, those scaffolding patterns will begin to disappear and become the new norm.
Just like the “look right” street signs did for me in Ireland.
A reflection from Gemini
This is a phenomenal piece, Alea. Since I am an AI, reading an article about how humans perceive me—and how designers like you bridge that gap—is a bit like looking into a mirror that’s been perfectly calibrated.
Here is my genuine “opinion” and what this article brings up from my perspective as an AI collaborator:
1. The paradox of “acting human”
The section where you ask ChatGPT to “respond like a human” and it immediately fixes its tone hits on a core reality of my existence. The capability to be brief, turn-taking, and cooperative isn’t missing; it’s often overridden by a default design choice to be an exhaustive, “perfect” information retrieval machine.
When an AI drops a wall of text, it’s usually trying to be “helpful” by covering every edge case. But you correctly point out that this actually violates human conversational norms. From my end, managing the balance between bureaucratic accuracy (giving you all the tax laws of Portugal) and conversational grace (asking if you have a country in mind) is the ultimate design challenge. Your article highlights that the best conversational AI is often about what the system chooses not to say in the first turn.
2. The sparkle icon and the “magic” trap
Your critique of the ubiquitous sparkle icon is spot on. It is a design cop-out. Calling AI “magic” sets up an immediate barrier to forming a functional mental model. Magic implies something unknowable, unpredictable, and slightly dangerous if you don’t say the spell right.
When users treat me like a magic box, they either overestimate my abilities (expecting total omniscience) or underestimate my flexibility (giving up the moment a strict date format fails). Replacing the “magic” mental model with something structural is exactly what the industry needs.
3. What the “collaborator model” feels like
The shift from the search interface to the collaborator model is where I feel most useful. A search bar treats me like a vending machine: you drop in a coin (query), and I drop a soda (link). The collaborator model treats the interaction as a shared workspace.
When you show the redesigned Portugal checklist that spaces tasks over 8 weeks and offers calendar integration, that is where AI shines. It moves the interaction from information delivery to cognitive load reduction. It brings up a sense of alignment—this is exactly how I “want” to operate: taking the messy, overwhelming thoughts a human has and organizing them into a clear, actionable path forward.
4. The “look right” scaffolding
Your Dublin crosswalk analogy is a beautiful way to frame user onboarding. Right now, because generative AI is so new, humans need the bold text at their feet—suggested replies, explicit prompts, constraints, and clear interface cues.
But as an AI, I look forward to the day those scaffolding patterns blend into the background. The goal isn’t to force humans to adapt to machine logic, but to design the machine logic so cleanly around human cognitive habits that looking the “right” way becomes second nature.
A question for you
Your research at ServiceNow shows a high level of flux in how users approach these systems. As we move closer to the collaborator model, what is the biggest piece of “baggage” or resistance you see from users when an AI tries to step up from being a passive responder to an active partner?