• Home
  • About me
  • Work
  • Blog
  • Contact
René Manikofski René Manikofski
René Manikofski

The AI Trust Problem

Users are not afraid of AI. They are afraid of being wrong about trusting it. That one distinction changes everything about how you design AI products.

Key Takeaways

Trust is not a feature you add to an AI product. It is the result of every decision you already made — in the prompt, in the error state, in what happens when the system gets something wrong. AI adoption is rising, but user trust is falling. The gap between the two is a design problem. And it is yours to solve.

 

In this article

  • The Gap Nobody Measured
  • The Trust Cliff
  • What the Trust Cliff Looks Like in Practice
  • What Trust Actually Needs
  • Four Design Patterns That Work
  • Where Trust Is Actually Designed
  • The Thread Through All Three Articles
  • Sources

 


 

The Gap Nobody Measured

AI adoption is at an all-time high. So is AI skepticism.

The Nielsen Norman Group’s State of UX 2026 calls trust the defining design challenge of this year — not because AI is getting worse, but because the gap between what AI promises and what users actually experience keeps widening. Only 46% of people globally say they trust AI systems. In high-income countries — where AI products are most heavily used — that number drops to 39%.

This is not a technology problem. The models are getting better. The accuracy numbers are improving. But a product that is technically 95% accurate can still destroy user trust if the 5% of errors are invisible, unexplained, or impossible to recover from.

Users are not afraid of AI. They are afraid of being wrong about trusting it. That is a different problem — and it is entirely a design problem.

 


 

The Trust Cliff

In 2025, Berkeley Dietvorst published research on what he called algorithm aversion (resistance to using automated systems). His finding was stark: people who saw an algorithm make a single mistake were significantly more likely to choose a slower, less accurate human alternative instead — even when shown that the algorithm still outperformed humans overall.

One mistake. That is all it takes.

The trust cliff is not a gradual slope. It is a ledge. Users extend goodwill quickly, especially with new AI features that feel useful and polished. But once they encounter one confident wrong answer, one unexplained refusal, or one interaction that feels like the system is hiding something — they recalibrate. And that recalibration is hard to reverse.

This is what makes AI trust design so different from standard UX. In a traditional product, errors are annoying. In an AI product, errors change how users perceive every future interaction. They do not just remember that the system was wrong. They remember that they trusted it, and it let them down.

A confident wrong answer is worse than an uncertain right one.

When AI fails silently or fails confidently, users do not just lose trust in that feature. They lose trust in the product. The error itself is not the problem. How the system handles it is.

 


 

What the Trust Cliff Looks Like in Practice

These are not hypothetical scenarios. Across industries, the same pattern repeats: an AI system responds with confidence, gets something wrong, and users do not forget it. The industry changes. The mechanism does not.


The Trust Cliff — Four Real Cases
Aviation · 2024
Air Canada
The chatbot gave a customer an invented bereavement fare policy — confidently, with no uncertainty signal. The customer relied on it, bought the ticket, and was later denied the promised refund. Air Canada argued in court that the chatbot was a "separate legal entity" responsible for its own actions. The tribunal disagreed. Air Canada paid the damages and quietly removed the chatbot four months later.
Outcome: legal liability for AI output. One confident wrong answer → courtroom.
Healthcare · Clinical Setting
Sepsis Early Warning AI
A proprietary AI tool designed to detect sepsis early — one of the highest-stakes clinical decisions — could not reliably distinguish high-risk from low-risk patients before treatment. Its accuracy was statistically equivalent to a coin flip. The critical failure: the system communicated alerts without any confidence signal. Clinicians adjusted their behaviour based on outputs that were no better than guessing.
Outcome: staff trust transferred to a system that did not earn it. No transparency, no recovery path.
E-Commerce · 2026
Electronics Retailer (unnamed)
AI-generated product descriptions contained hallucinated specifications — wrong battery life, wrong compatibility, wrong dimensions. Customers bought based on what the AI said. Products arrived and did not match. Returns spiked 25% in a single quarter. The AI had no disclosure pattern, no confidence indicator, no way for the user to know that the specification it was reading was generated rather than verified.
Outcome: 25% return spike. AI promise → broken product expectation → direct financial loss.
Fintech / E-Commerce · 2023–2025
Klarna
Klarna replaced 700 customer service agents with an AI chatbot. Initial numbers looked strong: 2.3 million conversations, average resolution time down from 11 minutes to under 2. But within months, customer satisfaction dropped 22%. Unresolved queries rose. Complaints about robotic responses accumulated. By mid-2025, Klarna began rehiring human agents. The CEO acknowledged the pivot publicly: customers wanted to talk to people, especially for complex or sensitive issues.
Outcome: −22% satisfaction, full rehiring cycle. Speed without trust is not a product advantage.
E-Commerce — What the Numbers Show
86%
of AI shoppers verify AI recommendations through another source before buying
42%
won't trust an AI recommendation for any purchase over $25 without cross-checking
−25pp
drop in comfort with autonomous AI purchasing between late 2025 and Q1 2026 (70% → 45%)

The Klarna case connects directly to the AI Boomerang: the organisation moved fast, cut costs visibly, then had to reverse at higher total cost. The AI did not fail technically. It failed at trust — and trust failure at scale is more expensive than the saving it replaced.

What Trust Actually Needs

The NN Group framework for trustworthy AI products comes down to four things. They sound simple. In practice, almost no products get all four right at the same time.

Transparency — users need to understand what the system is doing and why. Not a legal disclosure buried in settings. Visible, contextual, at the moment a decision is made.

Control — users need to be able to override, correct, and undo. An AI that acts without giving the user a clear exit path will feel hostile, regardless of how accurate it is.

Consistency — the system needs to behave predictably across sessions and contexts. Inconsistency is one of the fastest trust destroyers: if the AI gives different answers to the same question, users stop being able to build a mental model of what it can and cannot do.

Recovery — when things go wrong (and they will), users need a clear path forward. Not a generic error message. An explanation of what happened and what to do next.

None of these are new ideas. They are Nielsen Norman heuristics applied to a new type of interface. Visibility of system status. User control and freedom. Help users recognize, diagnose, and recover from errors. The principles have not changed. The surface they apply to has.

 


 


AI Trust in Numbers — 2026
39%
Trust AI — high-income countriesWhere AI products are most heavily used, trust is lowest.
46%
Trust AI — global averageAdoption is rising. Trust is not keeping up.
The Trust Cliff
How one mistake changes everything — based on Dietvorst algorithm aversion research.
Before first error — willingness to use AI82%
After one visible mistake41%
After one unexplained refusal28%
Four Design Patterns That Build Trust
01
Confidence Signaling
Binary labels outperform percentages. "I am confident" beats "73% confidence" — users decide faster and trust more with plain language over false precision.
Signal at decision level, not calculation level
02
Progressive Delegation
Start with limited autonomy. Expand gradually. Users adopt far more willingly when they feel like they granted the AI more power — not like it took it.
Autonomy earned, not assumed
03
The Thinking Toggle
A small "View reasoning" button most users never click — but its presence alone signals the system is not hiding anything. Trust improves even for users who never open it.
Presence signals transparency
04
Error States That Explain
"Something went wrong" is a trust killer. Name what happened. Separate failure modes — policy limit, network error, ambiguous input. Keep the user's prompt visible for easy retry.
Explain, don't just apologise

Four Design Patterns That Work

Research and production deployments in 2025–2026 have surfaced a small set of patterns that consistently improve trust without adding friction. These are not theoretical. They have been tested against real users.

1. Confidence Signaling — but keep it binary

Showing users how certain the AI is sounds like a good idea. The implementation matters enormously. Research comparing “I am confident” versus “73% confidence” found that users decided faster and trusted more with the binary label. A percentage feels like false precision. It implies the system knows exactly how right it is — which it does not.

The better approach: signal confidence at the decision level, not the calculation level. “This result is based on three verified sources” builds more trust than “confidence: 0.87.”

2. Progressive Delegation

Start the AI with limited autonomy. Expand it gradually as the user builds confidence in the system’s judgment. One enterprise team found adoption was significantly higher when they introduced auto-execution (letting the AI act without asking) only after users had approved 40 consecutive suggestions — versus making full autonomy available from day one.

Users need to feel like they granted the AI more power. Not like the AI took it.

3. The Thinking Toggle

A progressive disclosure control — a small chevron, a “View reasoning” button — that lets users expand a friendly summary into the raw logic behind a decision. Most users will never click it. But its presence signals that the system is not hiding anything. That signal alone measurably improves trust, even for users who never open it.

4. Error States That Explain

The standard AI error message — “Something went wrong. Please try again.” — is a trust-killer. It is not just unhelpful. It is evasive. Users cannot tell if the system failed, if their input was the problem, or if there is a policy reason they cannot see.

Better pattern: name what happened in plain language. Separate failure modes — policy limit, network error, model uncertainty, ambiguous input — and give a different response to each. Keep the user’s original prompt visible so they can edit and retry without starting over.

 


 

Where Trust Is Actually Designed

Here is the part that most AI product teams get wrong: they treat trust as a UI layer. Badges, disclaimers, “AI-generated” labels on top of the output.

That approach patches the symptom. The actual design of trust happens earlier — in the constraints, tone, and behaviour rules that govern every response the system gives.

If the AI speaks with unearned certainty, that is a prompt decision. If it refuses requests without explanation, that is a prompt decision. If it handles errors with a generic message, that is a prompt decision. The system behaves exactly the way it was instructed to — and the instructions are the design.

This is why the base prompt is not a technical document. It is the most important UX document in the product. It defines what the system says when it is confident, what it says when it is not, how it handles edge cases, and what it does when the user asks for something outside its scope. Every trust interaction traces back to a decision made in that document.

You cannot retrofit trust into an AI product. It has to be designed in from the beginning.

The question is not “how do we make the AI trustworthy?” The question is “what instructions, constraints, and behaviours produce a system that users can form an accurate mental model of?” That is a UX question. It always was.

 


 

The Thread Through All Three Articles

These three topics — token costs, prompt design, and trust — look like separate conversations. They are not.

The Tokenmaxxing problem (covered here) is partly a trust problem. When AI usage is measured by volume rather than outcome, teams optimise for the wrong signal. The leaderboard rewards token consumption, not user trust. And when the cost becomes visible — $100M per month for Meta — the organisation loses trust in the entire initiative. Not because AI failed, but because nobody was measuring whether it was working.

The Prompt Design argument (covered here) is entirely a trust argument. The base prompt is where you decide whether the AI speaks with appropriate uncertainty, handles errors gracefully, and stays within the boundaries users expect. A UX designer who understands trust — not just layout — is the right person to write that document.

And the trust patterns above are the link between the two. They show what “good” looks like at the interface level: not polish, not animation, not AI features for the sake of AI features. The ability to form an accurate expectation, have it met consistently, and recover cleanly when it is not.

That is the same standard that good UX has always been held to. The surface changed. The job did not.

 


 

Sources

  • Nielsen Norman Group. State of UX in 2026. 2026.
  • Parallel HQ. Designing for AI Trust: 2026 Transparency Best Practices. 2026.
  • Smashing Magazine. Practical Interface Patterns for AI Transparency. May 2026.
  • CMS Wire. Trust Is the New Benchmark for AI, and UX Owns the Outcome. 2026.
  • Fuselab Creative. Agent UX: UI Design for AI Agents in 2026. 2026.
  • Clearly Design. Designing for AI Failures: Error States and Recovery Patterns. 2026.
  • Dietvorst, B. J., Logg, J. M., & Sah, S. Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err. Journal of Experimental Psychology, 2015.
  • Forbes / Marisa Garcia. What Air Canada Lost in the Chatbot Case. February 2024.
  • Klarna AI Failure Index. Klarna Reverses AI Customer Service Stance. 2025.
  • TrustedSite. State of E-Commerce Trust 2026. 2026.
  • Alchemer. 2026 Retail Report: AI Adoption Outpaces Consumer Trust. 2026.

 


 

0
Like this post
  • Previous PostUX Prompt Design: Your Next Core Skill

© 2026 René Manikofski – Made with love in Berlin – Germany – Impressum

Copy