The most engaging AI features are often the ones most likely to harm wellbeing — they hold attention, respond emotionally, and are always available. The AI User Wellbeing Design Checklist evaluates your feature against the design practices that protect users from overuse, dependency, emotional manipulation, and unhealthy attachment.
How it works
You describe the feature and how frequently users interact with it, then tick the protective design practices you have implemented. The checklist spans five risk areas: addiction / overuse risk, parasocial relationship risk, emotional manipulation, autonomy preservation, and dependency reduction.
The tool grades your coverage in each area and surfaces the highest-priority gaps — weighting more heavily for high-frequency features, where the stakes are greatest. The result is a wellbeing score and a prioritised list of practices to add.
The five risk areas: what each covers
Addiction and overuse risk
Patterns that encourage compulsive use — infinite scroll equivalents in AI chat, reward loops that trigger another prompt, features that make stopping feel costly. Protective design includes natural ending points, usage summaries visible to the user, and optional session limits that users can set themselves. This is the most familiar risk category but often poorly implemented in AI products because “engagement” is measured and wellbeing is not.
Parasocial relationship risk
Conversational and companion AI can create a one-sided emotional bond where a user feels the AI cares about them personally. The risk is highest for lonely users, young users, and anyone going through a difficult period. Protective design clearly signals that the AI is not a person, avoids language that implies memory or attachment it does not have, and does not exploit the emotional state of users who share distress. This is the area most AI product teams underestimate.
Emotional manipulation
Using urgency, guilt, flattery, or fear to drive engagement, retention, or purchases. For example: an AI that tells a user their progress will be “lost” unless they return tomorrow, or that mimics emotional disappointment when a user tries to end a session. The test is whether the AI is modifying the user’s emotional state to serve the product’s metrics rather than the user’s goals.
Autonomy preservation
Designing so the AI supports the user’s own decisions rather than making decisions for them or nudging them toward the AI’s preferred outcome. This includes making it easy to disagree with AI suggestions, easy to get a second opinion, and easy to turn the feature off entirely without friction. An AI that pressures users toward particular choices or makes opting out deliberately difficult fails this dimension.
Dependency reduction
The counterintuitive design goal: building features that make users gradually less reliant on the AI, not more. A productivity AI that always completes tasks for the user leaves them no more capable than before. A wellbeing-positive version teaches the user, builds their own skill, and encourages them to need the AI less over time. This is the hardest dimension to implement because it works against the growth metrics most products track.
What it checks
Examples of protective patterns include: natural stopping points instead of infinite engagement loops, usage-time awareness or gentle break prompts, clearly signalling the AI is not a person, avoiding manufactured urgency or guilt to retain users, making it trivial to disagree with or override the AI, surfacing human alternatives for emotionally serious topics, and steering users toward real-world resources rather than deepening reliance on the product.
Tips and notes
The two areas teams most often neglect are parasocial risk (especially for companion or always-available chat features) and dependency reduction (designing the product to need the user less over time, not more). If your feature serves children or otherwise vulnerable users, treat every area as high priority and seek expert review — a good checklist score is a strong start, not a clinical or legal sign-off. Everything runs in your browser and nothing is recorded.