The Pause

You've seen it a thousand times. Three dots, bouncing. A shimmer crawling left to right. A cursor blinking in a text field that nobody is typing in.

You wait.

The model finished generating its response 400 milliseconds ago.

Four hundred milliseconds

That delay — the one between the answer existing and the answer appearing — is a design decision. Someone chose it. A product team debated it. An engineer parameterized it. A researcher tested what happened when they removed it.

What happened was: people didn't like it.

Not the answer. The answer was identical. Same words, same accuracy, same level of detail. But when it arrived instantly — when a user typed a question and the response appeared fully formed in the time it takes to blink — something felt off. Users in early testing described the output as "less thoughtful." Some said it felt "rushed." A few said they trusted it less.

Same answer. Different timing. Completely different relationship.

So the team added latency back in. Not because the system needed it. Because the human did.

Lufthansa had to delay the responses of their chatbot Mildred after users were "irritated" that it replied unnaturally fast. "We had to put in a delay," said their Head of Digital Innovations, "so they believed it was a natural conversation."

A process the product doesn't experience

This is not a story about chatbots. This is a story about what happens when a product performs a process it doesn't experience.

When you watch a person think — a doctor pausing before a diagnosis, a friend going quiet before answering a hard question — the pause carries information. It signals that the question was hard. That the answer required effort. That what follows was considered, not reflexive.

We've spent our entire lives calibrating trust against these signals. Centuries of human interaction have taught us that speed and certainty are inversely correlated. The person who answers instantly is guessing. The person who takes a moment might actually know.

AI breaks this completely.

A large language model doesn't think harder about hard questions. It doesn't pause on moral dilemmas. It generates tokens at roughly the same rate whether you ask for a cookie recipe or an assessment of your child's symptoms. The computational difficulty of a question is almost entirely disconnected from its emotional weight.

But the product can't show that. If it did — if every answer arrived at the same speed, regardless of what was asked — the human on the other side would stop trusting the answers. Not because the answers got worse. Because the relationship felt wrong.

So we design the performance.

"It introduces a non-negligible feeling of artificiality to interact with something that can respond instantly to anything you say." — Ryan Schuetzler, researcher on human–chatbot interaction2

The most successful trust artifact nobody recognizes

The typing indicator is the most successful trust artifact in modern software, and almost nobody thinks of it as one.

Its origin is simple. In early messaging apps, the three dots solved a real problem: reducing the anxiety of waiting for a reply from another human. Is she still there? Did he see my message? The indicator was a signal of presence — proof that on the other end of this silent screen, a person was engaged.

When AI products adopted it, the function inverted. The indicator no longer signals presence. It signals thought. And the thing generating the response is not thinking. The dots are a metaphor for a cognitive process that isn't occurring.

This isn't a glitch. It's a decision. And the decision works.

Users who see the typing indicator before an AI response rate the response higher on perceived quality. They report feeling more "heard." They're more likely to follow the recommendation. All from a three-dot animation that represents nothing.

Harvard researchers Ryan Buell and Michael Norton found that people can actually prefer websites with longer waits to those that return instantaneous results — even when the results are identical. They called this the "labor illusion": when a system signals it's working hard on your behalf, you value the output more. The perceived value rose 8% with increased operational transparency.3

The question this should raise — the one most product teams skip past — is: what kind of trust are we building?

The locked door and the walk home

I spent years designing trust in financial products. The stakes were different — someone's money, someone's identity, someone's safety — but the core question was the same one AI product teams are facing now, whether they realize it or not.

There's a version of safety that looks like a locked door. Imposing. Defensive. It says: we take security seriously. And there's a version that looks like someone walking you home. Calm. Present. It says: you're safe here, and here's why.

Both versions protect. Only one builds a relationship.

The typing indicator, the progress bar, the thoughtful-looking shimmer animation — these are the AI equivalent of the locked door. They're security theater for cognition. They perform the appearance of care without requiring the product to actually be careful. And in most cases, they work well enough that nobody questions them.

Bruce Schneier coined the term "security theater" to describe measures that make people feel more secure without doing anything to actually improve their security. "Security is both a feeling and a reality," he wrote, "and they're different."4

But "well enough" has a shelf life.

"You build up expectations of behavior based upon prior experience, and if the items with which you interact fail to live up to expectations, that is a violation of trust." — Don Norman, Emotional Design

The performance grows louder

Consider what happens as these products get faster. Models are already approaching near-instant generation for most queries. The next generation will be faster still. The gap between when the answer exists and when the product chooses to reveal it is widening.

Which means the performance is getting louder. The pause is growing. The shimmer is shimmering longer. And the lie — if that's what it is — is becoming more elaborate.

Some teams have started designing variable delay. The system waits longer before displaying answers to emotionally sensitive questions. A medical query gets more dots than a weather query. A financial question gets a progress bar; a trivia question doesn't. The product is calibrating its performance of thoughtfulness based on how much thoughtfulness the user expects.

This is sophisticated work. And it raises a question that no amount of user testing can answer: is this trust, or is this the simulation of trust?

A feeling you create vs. a condition you maintain

The instinct is to say it doesn't matter. If the user feels considered, and the answer is accurate, who cares how long the interface waited to show it? The outcome is good. The feeling is good. Move on.

I understand that instinct. I've even agreed with it, in certain contexts.

But I've also watched what happens when a product builds trust on a performance instead of a foundation. I've seen it in financial services, where the appearance of security — the padlock icon, the green checkmark, the reassuring copy — can actually reduce safety by making customers less vigilant.

I've seen it in platform design, where the appearance of fairness masks systems that discriminate at scale. The surface said one thing. The structure did another. And by the time anyone noticed, the gap was too wide to close without breaking something.

The performance of trust is not the same as the architecture of trust. One is a feeling you create. The other is a condition you maintain.

AI products are very good at the first. Almost none of them are investing in the second.

What honesty would cost

Here's what investing in the second would look like.

Instead of variable delay, variable transparency. The product doesn't pretend to think harder about a medical question. It tells you it's not a doctor. It shows you where its training data gets thin. It flags the confidence score — not as a percentage that means nothing to a non-statistician, but as a signal a human can act on. I'm fairly sure about this. Here's what I'd check next.

Instead of a typing indicator that performs presence, an interface that's honest about absence. The model isn't "thinking." It's generating a probabilistic sequence of words. That's not a shameful thing to admit. It's a remarkable thing. But the product has to trust the user enough to say it.

Instead of designing the pause to manufacture trust, designing the relationship to earn it.

Rachel Botsman, one of the leading researchers on trust, defines it as "a confident relationship with the unknown." She argues that when we demand transparency, we've often already given up on trust — we're seeking control, not connection. The real question isn't how to show more. It's how to need to show less.6

This is harder. Slower. Less testable in a two-week sprint. And it requires product teams to confront a question they'd rather not ask: if the user knew how this actually worked, would they still trust it?

If the answer is yes — build on that. If the answer is no — the pause isn't the solution. It's the symptom.

"The most powerful way to earn trust is: 'I don't know.'" — Rachel Botsman7

The hardest version of the problem

The irony is that the teams getting this right are the ones who learned it from the hardest version of the problem.

Trust in financial products. Safety in marketplace platforms. Verification flows where the person on the other end might be a threat or might be someone's grandmother trying to send rent money, and the design has to hold both possibilities with care. These are contexts where performing trust doesn't just feel hollow — it's dangerous. Where the gap between what the interface says and what the system does isn't a UX issue. It's a harm issue.

The people who designed for those contexts know something that most AI product teams haven't learned yet: trust isn't a feature you ship. It's a relationship you earn. And a relationship built on performance — no matter how convincing — is a relationship waiting to collapse.

The three dots are bouncing. The answer arrived a long time ago.

The question is whether the product has the honesty to show it.

People don't trust products. They trust the feeling of being considered. The question for AI isn't whether we can manufacture that feeling. We already can. The question is whether we should — and what we build instead when we decide we shouldn't.

Further Reading

• Buell, R.W. and Norton, M.I. "The Labor Illusion: How Operational Transparency Increases Perceived Value." Management Science, 2011.

• Gnewuch, U., Morana, S., Adam, M.T.P., and Maedche, A. "Opposing Effects of Response Time in Human–Chatbot Interaction." Business & Information Systems Engineering, 2022.

• "Explaining the Wait: How Justifying Chatbot Response Delays Impact User Trust." ACM, 2024.

• Schneier, B. "Beyond Security Theater." New Internationalist, 2009.

• Botsman, R. "An Expert on Trust Says We're Thinking About It All Wrong." TIME, 2024.

• Norman, D. Emotional Design: Why We Love (or Hate) Everyday Things. Basic Books, 2004.

• "Finding the Sweet Spot: Exploring the Optimal Communication Delay for AI Feedback Tools." Information Processing & Management, 2023.

Footnotes

1. Dr. Torsten Wingenter, Lufthansa Head of Digital Innovations, speaking at SAP's Sapphire Now conference in 2017. Lufthansa introduced their chatbot Mildred on Facebook Messenger in December 2016 and had to add artificial delays after user complaints about unnaturally fast responses. Reported by iTnews and cited in Gnewuch et al., "Opposing Effects of Response Time in Human–Chatbot Interaction," Business & Information Systems Engineering, 2022.

2. Ryan Schuetzler, "The Impact of Chatbot Response Time on User Evaluations," 2015. Cited in Gnewuch et al., "Opposing Effects of Response Time in Human–Chatbot Interaction," Business & Information Systems Engineering, 2022.

3. Ryan W. Buell and Michael I. Norton, "The Labor Illusion: How Operational Transparency Increases Perceived Value," Management Science 57, no. 9 (September 2011): 1564–1579. In five experiments simulating online travel and dating services, the researchers demonstrated that signaling effort — even when no additional effort was occurring — increased perceived value, satisfaction, and repurchase intentions.

4. Bruce Schneier coined the term "security theater" in Beyond Fear: Thinking Sensibly About Security in an Uncertain World (Copernicus Books, 2003). He expanded the concept in "Beyond Security Theater," New Internationalist, November 2009. The quote "Security is both a feeling and a reality, and they're different" appears in both the essay and in his Harvard Kennedy School interview, January 2021.

5. Donald A. Norman, Emotional Design: Why We Love (or Hate) Everyday Things (Basic Books, 2004). Norman also writes in The Design of Everyday Things: "We must design for the way people behave, not for how we would wish them to behave" — an observation that cuts both ways in the context of artificial delay.

6. Rachel Botsman, "An Expert on Trust Says We're Thinking About It All Wrong," TIME, March 2024. See also her Substack essay "The Telltale Sign You've Given Up on Trust," May 2024, and her Oslo Business Forum keynote, 2021. Botsman's core definition — "Trust is a confident relationship with the unknown" — appears in her book Who Can You Trust? (Penguin Portfolio, 2017) and in her TED talk, 2016.

7. Rachel Botsman, Oslo Business Forum keynote, 2021. Full context: "The most powerful way to earn trust is: 'I don't know.' Humility is a confident relationship with what we don't know." She argues that admitting limitation accelerates trust rather than undermining it — a direct challenge to the logic behind artificial delay, which assumes users need to see confidence in order to feel it.