Short answer
Shift from designing static interfaces to designing dynamic 'conversations' where the AI constantly updates the user on its capabilities, intent, and need for correction.
- Field
- User-Centred Design
- Source
- Academic Publication (2019)
- Method
- Three-phase heuristic development and validation study
- Sample
- 49 participants
- Evidence
- Strong effect
By aligning system behaviors with human expectations through consistent feedback and error-correction loops, designers mitigate the inherent unpredictability of probabilistic AI outcomes. This user-centred design research insight is drawn from a 2019 study published in Academic Publication. Using Three-phase heuristic development and validation study with 49 participants, researchers explored how this design variable affects real-world outcomes. The key design takeaway: Shift from designing static interfaces to designing dynamic 'conversations' where the AI constantly updates the user on its capabilities, intent, and need for correction.
Proactive transparency and graceful recovery mechanisms increase trust in AI-infused interfaces
By aligning system behaviors with human expectations through consistent feedback and error-correction loops, designers mitigate the inherent unpredictability of probabilistic AI outcomes.
Academic Publication · 2019
Key Findings
- 01Initial expectations influence long-term trust; systems that clearly state 'what the AI can do' reduce user disappointment.
- 02Interaction-time feedback loops (e.g., dismissing a recommendation) strengthen the user's sense of control.
- 03Providing explanations for system behavior reduces 'black box' anxiety during unexpected AI actions.
- 04Global guidelines (e.g., match social norms) are consistently relevant across all interaction stages from initial use to long-term habituation.
Application
Design takeaway
Shift from designing static interfaces to designing dynamic 'conversations' where the AI constantly updates the user on its capabilities, intent, and need for correction.
How to apply
Implement 'undo' or 'dismiss' buttons directly on AI recommendations (Guideline G9) and use onboarding tooltips to explicitly state that the AI is learning and may make mistakes (Guideline G1).
Project actions
- 01Pick 3 out of the 18 guidelines to focus on in your prototype.
- 02Create a 'correction' feature where users can swipe away or downvote a suggestion.
- 03In your user testing, ask specifically if the user understood WHY the system showed them a specific result.
Method & Evidence
Variables
Strengths & Limitations
Strengths
- +Utilizes a large-scale synthesis of existing design recommendations (1500+) to develop a comprehensive set of guidelines.
- +Employs a practical validation method (heuristic walkthroughs) with a significant number of UX professionals (49) to assess guideline applicability.
- +Investigates a critical and timely area of Human-AI interaction, directly addressing user frustration and loss of agency.
Limitations
Testing these guidelines requires a high-fidelity prototype that mimics AI behavior, which can be difficult to build without actual coding skills (use Wizard-of-Oz testing instead).
Reliability & validity
The study's reliability is supported by the systematic synthesis of a large dataset and consistent application of heuristics by multiple UX professionals. Validity is strong in terms of ecological validity (using real-world AI products) and construct validity (measuring user trust in relation to design features). However, external validity may be limited due to the focus on consumer software, as acknowledged by the authors.
Think critically
If we make AI too transparent about its mistakes, do we risk making users lack confidence in the technology even when it is performing well?
Design Principles
"Design for Error-Prone Intelligence"
AI systems often violate traditional usability heuristics because they are probabilistic rather than deterministic. Users experience profound frustration and loss of agency when an AI's hidden 'logic' fails to match their mental model or when the system lacks a clear path to fix its mistakes.
What This Means for Your Design
Since AI isn't 100% accurate, designers need to give users tools to understand why the AI is making a choice and an easy way to fix it when it gets things wrong.
How to use in your project
- 1.Use these guidelines as a framework for your Heuristic Evaluation phase.
- 2.Cite Amershi et al. (2019) when justifying why you added a 'why am I seeing this?' button to your app.
Add to My Project
Quick Cite
Paragraph starter
According to the guidelines by Amershi et al. (2019), AI interfaces must support 'graceful recovery' from errors to maintain user trust and system usability.
Source
Questions About This Research
- What does the research say about proactive transparency and graceful recovery mechanisms increase trust in ai-infused interfaces?
- Shift from designing static interfaces to designing dynamic 'conversations' where the AI constantly updates the user on its capabilities, intent, and need for correction. Evidence: Academic Publication (2019).
- Why does "Proactive transparency and graceful recovery mechanisms increase trust in AI-infused interfaces" matter for design?
- AI systems often violate traditional usability heuristics because they are probabilistic rather than deterministic. Users experience profound frustration and loss of agency when an AI's hidden 'logic' fails to match their mental model or when the system lacks a clear path to fix its mistakes.
- How can designers apply this research?
- Shift from designing static interfaces to designing dynamic 'conversations' where the AI constantly updates the user on its capabilities, intent, and need for correction.
- What were the main findings?
- Initial expectations influence long-term trust; systems that clearly state 'what the AI can do' reduce user disappointment.. Interaction-time feedback loops (e.g., dismissing a recommendation) strengthen the user's sense of control.. Providing explanations for system behavior reduces 'black box' anxiety during unexpected AI actions.. Global guidelines (e.g., match social norms) are consistently relevant across all interaction stages from initial use to long-term habituation.
- What research method was used?
- Three-phase heuristic development and validation study with 49 participants.
- How strong is the evidence?
- Evidence strength is rated Strong effect, based on a 2019 journal from Academic Publication.
- What should I do differently in my next project?
- Implement 'undo' or 'dismiss' buttons directly on AI recommendations (Guideline G9) and use onboarding tooltips to explicitly state that the AI is learning and may make mistakes (Guideline G1).
- What are the limitations?
- The guidelines focus primarily on consumer-facing software; highly specialized domains like medical AI or high-frequency trading may require additional safety-critical constraints.