The fundamental design paradox with AI trust.
Consumers demand AI transparency, but that very transparency can erode confidence in AI use.
This paradox creates a silent erosion dynamic where brand relationships dissolve without warning signals that traditional metrics can capture.
When Transparency Erodes Confidence
The Disclosure Dilemma Nobody’s Talking About
In my last piece, I mapped how intelligent interfaces are competing for agency - the right to decide what happens next in our lives.
But I left something crucial unresolved: the paradox at the heart of this transformation.
Data reveals a counterintuitive point that most brands miss: people demand transparency, but this very transparency can erode confidence by highlighting the artificial nature of interaction.
It’s not a communications problem we can solve with better disclosure statements.
It’s a fundamental tension in how humans form trust with non-human actors.
In Deloitte’s 2025 Connected Consumer survey:
20% of consumers say tech providers are “very clear” about data collection or usage
Only 20% say it’s “very easy” to control what’s collected.
Just 48% believe the benefits of online services outweigh privacy concerns—down sharply from 58% in 2024 and the lowest level since they began tracking this sentiment in 2019.
The erosion of trust is real and accelerating.
Meanwhile, 82% of consumers say they’d have more trust if humans were involved in AI development, and over 80% believe AI-created material should be clearly labelled.
That sounds like a clear mandate for transparency.
But we need to dig deeper.
Research from multiple behavioural science labs shows that AI disclosure itself can reduce trust by lowering perceptions of legitimacy, even when accuracy is demonstrably high and outcomes are superior.
How This Impacts Consumer and Brand Dynamics
The trust paradox creates a silent erosion dynamic.
Brand relationships dissolve without warning signals that traditional metrics can capture.
While 67-73% of customers leave brands due to poor service, Qualtrics data reveals that only 29% give feedback and 30% tell no one - including friends - after bad experiences.
This means brands optimising AI for cost deflection rather than problem resolution create what research identifies as “satisfaction without attachment”: it’s when customers report being satisfied in surveys (satisfaction scores holding at 77-79%) yet exhibit zero behavioural loyalty, quietly stopping their purchases without complaint or explanation.
When nearly 1 in 5 customers who tried AI support saw no benefit at all (a failure rate four times higher than average AI tasks), the message is clear: customers can tell when AI serves the company’s efficiency goals rather than their relief, and their response isn’t voice but exit.
Why the Paradox Exists: The Neuroscience
The answer isn’t where most companies are looking, it’s actually happening in the brain when people encounter AI systems.
Recent neuroscience research shows that trust formation is encoded in emotional brain circuits, not rational evaluation. When we interact with AI systems, our brains process cues at multiple levels simultaneously. Research using fMRI and biometric tracking reveals that emotional arousal patterns differ significantly when people know they’re interacting with AI versus humans, even when interaction quality is identical. The disclosure itself changes the neurological response.
This creates what researchers call the “cogwheel effect”—trust acts as a mechanism that transforms static characteristics into behavioural intentions. But it operates through dual pathways that can contradict each other. The heuristic pathway processes gut-level signals (Does this feel human?), while the systematic pathway evaluates evidence (Can I verify these claims?). Your systematic brain might calculate 95% accuracy, but your heuristic brain detects artificiality, triggering hesitation that overrides statistical evidence. Trust doesn’t average these signals; negative heuristics dominate.
When asked what activities they’d trust AI to do, less than a quarter trust it for high-stakes tasks. The pattern is stark: the less risky the activity, the more trust. An agent that’s 95% accurate at recommending restaurants gets adopted quickly; the same accuracy for medical treatment, or buying, generates widespread resistance.
This isn’t rational—it’s neurological.
Last week, I chatted with Isabelle Zdatny, Head of Thought Leadership at Qualtrics XM Institute, about their 2026 Consumer Experience Trends Report just released.
While topline numbers looked encouraging (satisfaction at 79%, trust at 76%), she noted: “Under the hood, things are not as rosy.” The gains cluster in easy-switch categories where loyalty was already weak. For every ten poor experiences, five customers simply cut spending without warning.
The Qualtrics data shows 73% of consumers use AI for daily tasks, yet customer support is rated the worst AI use case. The top concern isn’t capability; it’s misuse of personal data, and no human to connect with when things go wrong.
Only 29% trust organisations to use AI responsibly.
Isabelle’s phrasing was precise: “Companies are deploying AI to deflect cost, not resolve problems. Customers can tell.” When people are stressed or frustrated (the likely state when seeking support) their cognitive fluency drops, and emotional processing dominates. AI systems optimised for efficiency trigger “coldness perception,” activating distrust pathways regardless of whether the AI actually solves the problem.
The Empathy Spectrum: A Governance Framework That Works
Given this neurological reality and the regulatory landscape emerging worldwide, I’ve developed a framework that addresses both the trust paradox and compliance requirements. My Empathy Spectrum provides a red-line policy for emotion-aware features, creating clear boundaries that balance capability with ethical responsibility.
The Four Tiers
Supportive (De-escalation, Accessibility)
Rule: Allowed by default
Why: Primarily reduces harm
Examples: Detecting frustration to offer human handoff, accessibility features for neurodivergent users, crisis intervention triggers when language suggests self-harm risk
Measurement: Time-to-relief, escalation prevention rate, accessibility usage patterns
Assistive (Tone and Pacing)
Rule: Allowed with disclosure
Why: Adapts interaction style; users should know it’s happening
Examples: Slowing cadence when user seems stressed, softening language when confused, adjusting information density based on comprehension signals
Measurement: User satisfaction with pace, comprehension rates, override frequency when users manually adjust
Persuasive (Conversion Tuning)
Rule: Restricted and auditable
Why: Explicitly uses emotional data to influence outcomes
Examples: Timing offers based on detected engagement peaks, adjusting messaging based on urgency signals, optimising checkout flow based on hesitation patterns
Measurement: Regret rate (reversals within 24-48 hours), override frequency, conversion lift vs. ethical cost
Manipulative (Exploiting Distress)
Rule: Prohibited entirely with enforcement mechanisms
Why: Exploits vulnerability for commercial gain
Examples: Targeting ads during detected emotional lows, increasing prices when the system detects desperation, using detected anxiety to pressure purchases, lengthening cancellation flows when the user shows frustration
Measurement: Red-line violations tracked, automatic alerts to governance team, and user harm reports
Why This Framework Matters
The Empathy Spectrum isn’t just ethical guidance. It’s a practical governance tool that aligns business incentives with responsible use.
If your incentive structure only rewards growth, systems will inevitably drift toward manipulation because that’s what gets optimised. You’ll feel it in your support queue long before you see it in your dashboards: watch for patterns of confused or frustrated users who feel pushed toward outcomes they didn’t actually want.
The framework provides four critical functions:
Clear boundaries for teams building emotion-aware features
Audit criteria that specify what to measure and when to intervene
Enforcement mechanisms that have real teeth, not just principles
Alignment with emerging regulations, particularly the EU AI Act
How the Empathy Spectrum Maps to the EU AI Act
The EU AI Act, which becomes fully applicable August 2, 2026, takes a risk-based approach that closely mirrors the Empathy Spectrum.
This isn’t a coincidence. It’s validation that thinking clearly about emotion-aware AI leads to similar conclusions about appropriate boundaries.
Supportive → Minimal Risk
EU classification: AI with negligible impact on rights/safety
Compliance: Voluntary codes of conduct encouraged
Examples: Crisis detection, accessibility enhancements
No formal registration or assessment required
Assistive → Limited Risk
EU classification: AI that interacts with users without significant implications
Compliance: Transparency obligations—inform users they’re engaging with AI
Examples: Chatbots must disclose their AI nature, and emotion-adaptive interfaces must explain pacing changes
Applied from August 2, 2025
Persuasive → High Risk
EU classification: AI in critical areas with significant impact
Compliance: Register in EU database, conduct conformity assessment, implement risk management systems, ensure human oversight, report incidents
Examples: AI screening job candidates based on video interviews, credit scoring using behavioural signals, dynamic pricing based on detected urgency
Full compliance by August 2, 2027
Penalties up to €15M or 3% global turnover
Manipulative → Unacceptable Risk (Prohibited)
EU classification: Subliminal manipulation, exploiting vulnerabilities
Compliance: Banned entirely from February 2, 2025
Examples: Voice-activated toys encouraging dangerous behaviour, systems influencing voting without knowledge, social scoring, and cognitive manipulation of vulnerable groups
Penalties up to €35M or 7% global turnover
The alignment is striking. What I initially developed as an ethical framework turns out to anticipate the world’s first comprehensive AI regulation. I’d like to think this suggests both approaches are grounded in a realistic assessment of where emotion-aware AI creates value versus harm.
The US Context: Fragmentation Creates Opportunity
While the EU leads with comprehensive regulation, the United States takes a markedly different approach.
As of 2025, there is no federal AI law. Instead, companies navigate a patchwork of state legislation, sector-specific rules, and existing consumer protection laws applied to AI.
The Trump administration reversed Biden’s Executive Order on AI safety in January 2025, shifting focus to “removing barriers to innovation” with minimal federal oversight.
Meanwhile, all 50 states introduced AI legislation in 2025, with approximately 100 measures enacted across 38 states. Colorado passed the first comprehensive state AI law in May 2024, effective February 2026. California, New York, Utah, and Tennessee have followed with varying requirements around disclosure, deepfakes, voice cloning, and high-risk systems.
In July 2025, the Senate voted 99-1 to reject a proposed federal moratorium on state AI laws, signalling that states will continue filling the regulatory void with local frameworks. The result: companies face compliance complexity across jurisdictions with no unified standard.
This fragmentation makes the Empathy Spectrum even more valuable.
Positioned as “EU-compliant + beyond,” it provides companies, particularly US firms operating globally, with a practical framework that works across jurisdictions and anticipates where regulation is heading. Rather than scrambling to meet different state requirements or waiting for federal clarity that may never come, organisations can implement a single governance structure that satisfies the EU’s comprehensive requirements while addressing emerging US state laws and existing federal enforcement through agencies like the FTC, EEOC, and CFPB.
The framework’s risk-based approach aligns with Colorado’s high-risk AI definitions, California’s disclosure requirements, and the broader direction of state legislation focusing on transparency and consumer protection. For companies operating across borders, this means building once and deploying everywhere, rather than maintaining separate compliance strategies for each market.
Practical Implementation
For emotion-aware AI in customer service specifically, apply this lens:
Detect frustration to route to human agent → Supportive → Deploy with monitoring
Adjust explanation complexity based on comprehension signals → Assistive → Deploy with disclosure (”I’’m adjusting my explanations based on your questions”)
Optimise upsell timing based on engagement signals → Persuasive → Requires assessment, human oversight, regret tracking
Extend retention flows when detecting cancellation hesitation → Manipulative → Prohibited
The key metrics that keep you honest:
regret rate (how often users reverse decisions),
override frequency (how often they manually intervene),
novelty acceptance (whether they engage with suggestions outside normal patterns),
time-to-calm (how long recovery takes after system errors).
These matter as much as conversion because they reveal whether you’re building trust or burning it for short-term gains.
What to Build Next: Aligned Action
Based on this framework and the regulatory reality, here’s what I think works:
Human + AI, one resolution path. Simple issues get AI end-to-end with clear summary. Messy situations get AI draft plus human finish with automatic handoff and no repetition. Measure combined first-issue resolution, not bot deflection rates. This hybrid approach satisfies efficiency and trust needs because customers experience continuity.
Design for modes, not profiles. Detect signals of hurried, careful, exploring, stressed, or committing states. Tune steps and guardrails to each. Use minimal, situational data rather than accumulating comprehensive profiles. State the benefit plainly: “Because you did X, we did Y, which saved you Z.” This matches how brains form trust—through clear cause-and-effect relationships in the moment.
Make privacy human. One-page plain-language explainer. Real control centre where users see, edit, pause, or delete data. Publish “you said, we changed, here’s the result” updates where customers actually are. Trust is an outcome of behaviour, not a promise at signup.
Build early warnings. Fuse calls, chats, behavioural trails, and operational data into systems that prevent problems, not just document them. Tie every alert to an owner and a fix-by date. This creates the transparent logging and accountability that research shows is necessary for trust with non-human systems.
Ship a trust brief with every feature. Document where humans stay in the loop, what data minimisation looks like, retention periods, and provide a one-sentence explanation a customer would accept. If you can’t explain it simply, don’t ship it. This addresses the disclosure dilemma through contextual, moment-relevant transparency.
Report CX like P&L. Track satisfaction, trust, recommend-intent, and buy-more-intent monthly. Link directly to revenue and cost-to-serve. Treat them as business levers predicting future performance, not vanity metrics.
Isabelle’s team ran a multi-year analysis showing emotion leads business outcomes, and the gap between emotion leaders and laggards is widening. Research using fMRI shows brand decisions activate the same neural regions as personal relationships. Systems delivering perfect efficiency but no emotional resonance create “satisfaction without attachment”—customers report satisfaction but show zero behavioural loyalty, switching for minor improvements because there’s no emotional cost to leaving.
The Line Forward
Consumers are open again. Ready to try new things. But they won’t tolerate cheap-feeling AI, murky data habits, or dead-end support revealing you deployed automation to cut costs rather than solve problems.
The Empathy Spectrum provides the governance structure to navigate this. It’s not just ethical, it’s strategic.
Design for modes, not profiles. Keep fast paths to humans for moments when emotional state demands it. Treat trust as a system you build through consistent behaviour aligned with your stated framework, not a slogan you print on marketing materials.
The research proves it. The regulation requires it. The question is whether you’ll implement it before your competitors figure out that the trust gap is the new competitive moat.
Coming Next: Identity Fluidity and the Interfaces That Adapt
The intelligent interfaces I’ve described still assume relatively stable preference profiles, but identity itself is becoming radically fluid. I’m not the same person when I’m parenting versus strategising versus grieving versus celebrating. The contexts shift, the situationships demand different versions of me, and the modes I occupy change faster than any preference file can track.
If agents optimise for consistency when humans are actually fluid, they’ll systematically misread signals and make recommendations that feel tone-deaf or intrusive. In my following pieces, I’ll explore how interfaces can detect modal transitions through contextual signals, how brands can build offers that adapt to situational identity rather than fixed personas, and why the future of personalisation depends on recognising that I contain multitudes—and those multitudes deserve to be seen rather than averaged into demographic fiction.
Thanks for reading! Michael.
This essay brings together both UNCX and continues the Intelligent Interfaces series. I’m shaping a guide that bundles my frameworks—Axis of Consumer Agency, Empathy Spectrum, Shared Agency Loop, and Identity Fluidity—with research citations and practical checklists. You’ll see pieces land here first, then in the book.