Shuichiro Ogawa
日本語

Notes · updated 2026-06-28

The AI Expectation-Capability Gap and UX — Industry Trends (2026)

An integrated summary of 33 collected sources: vendor primary sources (T1v: 8 items), public and consulting surveys (T2: 10 items), and practitioner insights (T3: 13 items). Position talk (each company’s promotional claims of superiority) has been separated. The internal working ledger with provenance tracking, confidence ratings, and position assessments is source/review/ai-expectation-gap-ux/industry.md (repository-internal, not published). For the academic side, see ai-expectation-gap-ux-literature. Collection stream: industry source collection via WebSearch. Protocol: .claude/collection-protocol.md (zero fabrication, provenance tracking, position-talk removal).

TL;DR

As of 2026, the AI expectation gap has been made quantitatively visible. According to Pew Research, 44% use ChatGPT while only 29% trust its output — a 15pp usage-trust gap. Stanford HAI reported that the perception gap between experts and the public on AI’s employment impact reaches 50pp. Gen Z’s excitement about AI fell 14pp in one year, while anger rose 9pp.

Vendors have begun building design guidelines to close this gap through UX. The major frameworks are Microsoft HAX (18 guidelines, 4 temporal phases), Google PAIR (Mental Models chapter), IBM’s 6 principles (peer-reviewed at CHI 2024), and Apple HIG (with a newly added Generative AI section). In the practitioner community, new patterns responding to the agentic AI era (progressive delegation, proportional awareness, trust calibration spectrum) are rapidly taking shape.

The shared design principle is not “raising trust” but “forming trust commensurate with AI’s capabilities (calibrated trust),” and even on the same model, UX design can produce diametrically opposite trust outcomes.

1. Vendor Guidelines (T1v): Each Company’s Design Guidance

All seven major AI companies have published design guidelines, positioning expectation management as a central concern.

Comparing the Structures

VendorFrameworkOrganizing AxisCore of Expectation Management
MicrosoftHAX Toolkit (18 guidelines)4 temporal phases: Initially / During / When Wrong / Over TimeBehavior on failure and long-term adaptation
GooglePAIR Guidebook6 chapters: User Needs / Data / Mental Models / Explainability+Trust / Feedback+Control / ErrorsThe Mental Models chapter prescribes gradual onboarding
IBM6 principles (CHI 2024)Principle-based: Mental Models / Trust & Reliance / Generative Variability / Co-Creation / Imperfection / ResponsibleAddressing Generative Variability (different output every time)
AppleHIG for ML/AIInput/Output: Feedback / Calibration / Corrections → Mistakes / Options / Confidence / Attribution / LimitationsExplicit indication of where AI is used + disclosure of limitations
OpenAIApps SDK UX PrinciplesTask-flow centeredShow concrete value within 60 seconds
AnthropicResponsible Scaling Policy + System CardsSafety Levels + defense-in-depthAcknowledging uncertainty in capability thresholds
MetaResponsible AI PracticesTransparency-basedExplicit, plain communication

Three prescriptions are common across the companies. (1) Explicitly communicate AI’s capabilities and limitations. (2) Design graceful recovery that does not destroy trust on failure. (3) Give users means of control and feedback.

Generative Variability, identified by IBM’s Weisz et al. (CHI 2024), is a challenge specific to generative AI. Traditional UX presupposed the consistency of the same operation returning the same result, but generative AI breaks that. No existing guideline has sufficient patterns for designing expectations around this non-determinism.

2. Quantitative Evidence of the Expectation Gap (T2)

2.1 The Usage-Trust Gap

The 15pp usage-trust gap shown by Pew Research (June 2026) (44% usage vs 29% trust) is the most direct quantitative indicator of the expectation gap. 60% read AI-generated search summaries, but only 24% say AI has a positive effect on education, and just 23% for employment.

2.2 The Expert-Public Gap

According to the Stanford HAI AI Index 2026, 73% of experts are positive about AI’s employment impact versus 23% of the public (a 50pp gap). For economic impact the figures are 69% vs 21%; for healthcare, 84% vs 44%. Pew Research (April 2025) likewise confirmed a gap of nearly the same magnitude: 56% of experts positive vs 17% of the public.

2.3 The Intra-Organizational Gap

McKinsey (2025) reported that executives underestimate employees’ AI usage rates by a factor of three (executive estimate 4% vs employee self-report 13%). BCG (2025) showed that with leader support, employees’ positive sentiment rises from 15% to 55%, while only 29% of companies provide formal AI training. Persona-based learning reportedly achieved an adoption rate 20 times that of one-size-fits-all training.

2.4 The Acceleration of Disillusionment

Gen Z’s excitement about AI fell from 36% to 22% in one year, while anger rose from 22% to 31% (Stanford HAI 2026). 50% of Americans say their concern about the growth of AI outweighs their excitement (up from 37% in 2021), and two-thirds think AI is advancing too fast (Pew 2026).

These figures suggest that the expectation gap is not a problem of ignorance (“anxious because they don’t know AI”) but a problem of experience design: trust fails to form as a result of use.

3. Practitioner Patterns (T3): From the Design Front Lines

3.1 Three-Phase Patterns for the Agentic AI Era

Smashing Magazine (February 2026) organized agentic AI UX patterns into three phases. The wider industry discourse on Agentic Experience (AX) is organized in agentic-experience-industry.

  • Pre-Action (control): the Intent Preview pattern — “Here is what I am about to do. Is that OK?”
  • In-Action (context): Explainable Rationale + Confidence Signal — displaying “why” and “how certain”
  • Post-Action (safety): Action Audit & Undo + Escalation Pathway — post-execution auditing, undo, and escalation

3.2 Progressive Delegation

A pattern proposed independently by agentic-design.ai and UXmatters (December 2025). Agent autonomy starts low and expands gradually in line with the user’s approval history. This prevents a single failure from driving the user to abandon the product entirely.

3.3 Proportional Awareness

Proposed by Smashing Magazine (April 2026) and designative.info. More transparency is not better. Excessive alerts produce alert fatigue and trust collapse, while too little transparency produces overreliance. What is required is design that provides the necessary transparency at the necessary moment in the necessary amount.

3.4 Trust Calibration Spectrum

Grand Studio’s (2026) formulation is lucid: “Most AI feature failures in 2026 are not model failures. They are design failures.” Two products on the same model — one builds trust while the other destroys it, and the only difference is UX design. Excess trust leads to passive dependence; insufficient trust means the purpose of delegation goes unfulfilled.

3.5 The Shift in Error Handling

UXmatters (November 2025) stated that trust is earned not by reducing errors to zero but by how errors are handled. Confidence score display (“I am 85% confident”), human-understandable explanations of reasons, and graceful acknowledgment of mistakes (“I may have misunderstood — could you tell me more?”) are recommended. At the same time, a warning is issued against trustwashing (the mere appearance of transparency).

3.6 Onboarding Principles

OpenAI and Userpilot independently point to the principle of showing value within 60 seconds. In a flow where results appear only after answering five questions, most users drop off. Intent-based onboarding (structuring the flow around the user’s goal) outperforms feature-highlighting. For AI chatbots, making the tasks they can handle visible through suggested prompts and example queries quickly forms a mental model of “what it can do.”

4. Correspondence with Academic Findings

Industry trends align closely with academic research.

Industry PatternCorresponding Academic Finding
Microsoft HAX 18 guidelinesAmershi et al. (CHI 2019) is the academic foundation
IBM Design for ImperfectionKocielnik et al. (CHI 2019): advance disclosure of imperfection is effective for expectation calibration
progressive delegationBansal et al. (HCOMP 2019): mental models of error boundaries form gradually
proportional awarenessBuçinca et al. (CSCW 2021): adding explanations alone does not reduce overreliance
Emphasis on error handlingYin et al. (CHI 2019): observed accuracy affects trust
trust calibrationLee & See (2004): the three-axis definition of appropriate reliance
Caution toward anthropomorphismCrolic et al. (J. Marketing 2022): anthropomorphism amplifies expectation violations

However, whereas most academic research consists of one-off experiments, industry deals with design premised on long-term product operation. This difference in time horizons makes the two bodies of knowledge complementary.

5. Open Problems

  1. The guideline→implementation gap: Companies have published guidelines, but translation into implementable patterns remains insufficient (as UX Collective points out).
  2. Expectation design for non-determinism: No established pattern exists for calibrating user expectations around generative AI’s “different output every time.”
  3. Trust design for agentic AI: Expectation management for agent-type AI that takes actions differs qualitatively from conversational AI whose output is merely viewed. Gartner forecasts that over 40% of agentic AI projects are at risk of cancellation.
  4. Rebuilding trust after disillusionment: As the Gen Z data shows, no patterns are in place for addressing the cycle of excessive expectations → disillusionment → abandonment.
  5. Unverified methodology behind the NNGroup statistics: The figures of 63% (effect of confidence display) and 72% (influence of language) are widely cited, but the original study’s methodology and sample remain unverified [requires primary verification].

References

← All Notes · Home