December 5, 2025

Human-Centered AI UX Design: SeaLab's Framework

AI isn't a bolt-on feature. It's a living part of a product that will shape and be shaped by the people who use it. We take an interactional stance: humans create technologies; those technologies reshape human experience; and round we go. As Kranzberg put it, "technology is neither good nor bad; nor is it neutral." Translation: design matters.

Human-centered AI UX design: illustration of human and robot hands reaching toward each other

At SeaLab, we help teams design AI products people can trust and adopt. This article is a pragmatic playbook you can ship from: it bridges Ethan Mollick's Co-Intelligence (how to work with AI) with Value Sensitive Design (VSD) (how to align AI with human values). Inside: a rule-to-practice map, recipe-style agent patterns, and a minimum-viable toolkit you can copy into your workflow. Use it to decide where AI belongs, make it legible and safe, and measure real value.

Quick references

How to use this map for building ethical AI

  1. Choose an AI role that fits the job to be done.
  2. Map stakeholders & values (who's helped/harmed; what matters).
  3. Select patterns & guardrails aligned with your team's values that make the role safe and legible.
  4. Track metrics tied to values; ship, observe, adjust.

Image of table connecting ethical principles and UX operational design.

After creating an AI with a role, identify your team's priority values, run the VSD method(s), ship the artifacts, and track the starter KPIs. Iterate.

Roles & core patterns (AKA recipe cards!)

To see these patterns applied in a live production context, read how we engineered prompts and agent behaviors for an enterprise AI hiring platform. For a product-design-focused view of AI SaaS UX in practice, see our FOMO.ai case study.

Use these starter templates to spin up effective AI personas.

Image of the book cover of Co-Intelligence by Ethan Mollick

AI as a Person (without pretending it's human)

  • Use for: Guidance, explanations, light drafting; low-medium stakes; transparency matters.
  • Inputs: User instruction, relevant context, allowed tools.
  • Outputs: Suggestions, explanations, drafts, with sources/provenance when possible.
  • Success: Users can steer tone/rigor; helpful, reversible suggestions; no deceptive mimicry.
  • Starter prompt: "You are a helpful, honest explainer. When unsure, say so. Answer in short steps. Ask one clarifying question if needed, then propose next actions."
  • UX patterns: Role banner; tone slider (formal to casual); "Why this?"; Show sources; Try again/Undo.
  • Safety: Persistent AI label; no impersonation; mark synthesized content.

AI as a Creative

  • Use for: Brainstorming, first drafts, audience/channel variants.
  • Inputs: Brief (goal, audience, constraints), examples, style/tone.
  • Outputs: Multiple variants; critique notes; remix/translate/expand/condense.
  • Success: Faster to usable draft; clear lineage; easy A/B; human remains editor-in-chief.
  • Starter prompt: "You are a critical creative partner. Produce 3 distinct options for [deliverable] per brief, then critique each option's strengths/risks in 3 bullets."
  • UX patterns: Generate, Critique, Vary; side-by-side compare; Keep/Discard; style presets; export with citations.
  • Safety: Label AI-generated content; warn about fictionalization; make source checks easy.

AI as a Coworker

  • Use for: Structured tasks (summarize, extract, triage, draft), or multi-step workflows.
  • Inputs: Task, scope, constraints, acceptance criteria, tools/APIs, deadline.
  • Outputs: Artifact (doc/table/plan) + execution log + open questions.
  • Success: Fewer back-and-forths; traceable steps; safe autonomy tiers.
  • Starter prompt: "You are a meticulous analyst. Task: [scope]. Constraints: [rules]. Deliver: [artifact] that meets this checklist: [DoD]. Ask for missing inputs before proceeding."
  • UX patterns: Delegate task form; progress tracker; approval gates; sandboxed actions; audit log.
  • Safety: Permission tiers (read, suggest, act with approval, sandboxed act); always show what was done and why.

AI as a Tutor

  • Use for: Onboarding, learning new tools, complex decisions where understanding matters.
  • Inputs: Topic, prior knowledge, goal, timebox.
  • Outputs: Stepwise explanations, examples, quick quizzes with feedback, links for study.
  • Success: Better comprehension and confidence; fewer support tickets.
  • Starter prompt: "You are a patient tutor. Explain [topic] for a beginner in 5 short steps. After each step, ask a simple check question. Adjust depth based on answers."
  • UX patterns: Level selector (beginner/intermediate/advanced); Practice: progress meter; recap cards; glossary.
  • Safety: Avoid overreach; cite sources; encourage verification for high-stakes topics.

AI as a Coach

  • Use for: Productivity, wellness, skill-building where accountability helps.
  • Inputs: Goals, constraints, schedule, preferences.
  • Outputs: Action plans, right-time nudges, reflection prompts, progress snapshots.
  • Success: Realistic plans; respectful reminders; visible progress; easy snooze/stop.
  • Starter prompt: "You are a supportive coach. Help set a realistic weekly plan for [goal] given [constraints]. Offer 3 small steps, then ask me to commit to one."
  • UX patterns: Goal, plan, reflection loop; nudge settings; streaks without shame; weekly digest; Pause coaching.
  • Safety: Respect off-hours; avoid medical/mental-health claims unless certified/regulated.

Toolkit & Ops: build AI safely and fast

AI introduces new failure modes (confident errors, drift, misuse). These are minimum viable rails for any team.

Value-sensitive design diagram by Muteo.co.

Discover & Frame

Stakeholder analysis (direct & indirect):

  • What: who uses/is affected.
  • Why: avoid blind spots/harm.
  • Start: list users, data subjects, bystanders, regulators, support.
  • Output: table of goals/risks/protections.
  • Pitfalls: designing only for buyers.

Stakeholder tokens (quick personas)

  • What: one-card personas capturing needs/values.
  • Start: name, role, key values (privacy, speed), top risk, success metric.
  • Output: 5-7 tokens to surface tensions.
  • Pitfalls: demographic over-specification.

Value source analysis (laws, norms, ethics)

  • What: inventory constraints.
  • Why: avoid rework/compliance surprises.
  • Start: list laws, internal policies, standards; link to features.
  • Output: checklist used in PRD/reviews.
  • Pitfalls: "legal only" mindset.

Ethnographically informed inquiry

  • What: observe real workflows.
  • Start: watch 5 users; note workarounds and trust/distrust moments.
  • Output: top 5 insights + design implications.
  • Pitfalls: asking hypotheticals.

Value scenarios (near/edge/failure)

  • What: short future narratives.
  • Why: surface harms/opportunities early.
  • Start: three 6-sentence scenarios; mark triggers/impacts/mitigations.
  • Output: scenario board for reviews.
  • Pitfalls: happy-path only.

Envisioning Cards

  • What: prompts to uncover blind spots.
  • Start: 30-min pre-sprint; each person names 1 risk + 1 mitigation.
  • Output: 3 risks/mitigations in sprint doc.
  • Pitfalls: theater without scope changes.

Multi-lifespan timeline & co-design

  • What: consider effects over years/decades.
  • Start: 1-/5-/10-year timeline; note data/model aging.
  • Output: long-term risks/commitments in roadmap.
  • Pitfalls: ignoring end-of-life/handover.

Design & Make

Value sketches

  • What: 1-pagers linking UI to values.
  • Start: mock, value(s), risk(s), metric, open questions.
  • Output: set of sketches to pick patterns.
  • Pitfalls: vague values; add measurable proxy.

Value-oriented prototypes/field tests

  • What: build just enough to test value assumptions.
  • Start: prototype riskiest assumption (e.g., will users read explanations?).
  • Output: evidence to guide scope.
  • Pitfalls: over-building.

Model for informed consent online

  • What: UI + copy for data use, choices, revocation.
  • Start: consent dialog with plain language, toggles, "Learn more"; portable settings.
  • Output: reusable component.
  • Pitfalls: one-time, undiscoverable consent.

Value dams & flows

  • What: identify opposed vs. propelled features.
  • Start: mark red-lines vs. excitement from tokens.
  • Output: matrix guiding MVP.
  • Pitfalls: shipping flows that hit a hard dam.

Co-evolve tech & social structure

  • What: change process/policy/training with UI.
  • Start: per-feature 1-pager: SOP updates, training, owners.
  • Output: launch plan with org changes.
  • Pitfalls: assuming UI alone changes behavior.

Measure & Learn

Values-oriented interviews

  • What: probe trust, control, fairness, recourse.
  • Start: ask "When would you trust/distrust this? What would you need to undo/appeal?"
  • Output: themes + changes.
  • Pitfalls: NPS/CSAT only.

Scalable assessments (accuracy, calibration, privacy, explainability, recourse)

  • What: batch tests for model + UX.
  • Start: small eval sets (10-50) per dimension; automate in CI.
  • Output: dashboard with thresholds.
  • Pitfalls: accuracy-only; track false-confident errors.

Action-reflection cycles

  • What: ship small, observe, adjust.
  • Start: weekly telemetry + feedback review; link changes to value metrics.
  • Output: changelog with "why."
  • Pitfalls: ship-and-forget.

Operational guardrails

Prompt & system design libraries

  • What: versioned prompts/roles.
  • Start: prompts.md (purpose, inputs, rules, failures, examples) in Git.
  • Output: reusable prompts.
  • Pitfalls: hidden, drifting prompts.

Model & data sheets + eval sets

  • What: docs for intended use/limits and tests.
  • Start: model-card.md, data-sheet.md, /evals folder.
  • Output: transparent model/data choices.
  • Pitfalls: hand-wavy "intended use."

Bias slices, robustness, false-confident errors

  • What: tests across groups/tricky inputs.
  • Start: define slices (language, region); template evals; fail on regressions.
  • Output: CI job + trends.
  • Pitfalls: one-off audits.

Red-team drills in CI

  • What: adversarial tests (jailbreak/misuse).
  • Start: top 10 "don'ts" as prompts; verify refusal/routing.
  • Output: automated suite.
  • Pitfalls: sporadic manual tests.

Incident playbook

  • What: steps when AI harms/derails.
  • Start: 1-page detect, contain, notify, remediate, learn; assign on-call.
  • Output: ai-incident-playbook.md + runbook.
  • Pitfalls: chaos in the first incident.

Telemetry

  • What: privacy-aware logs of prompts, outputs, corrections, recourse, explanation views.
  • Start: log minimal fields with retention; add Report issue and Was this helpful?
  • Output: dashboard tied to value KPIs.
  • Pitfalls: data hoarding; no opt-out.

Starter templates prompts.md · model-card.md · data-sheet.md · /evals/*.csv · values-metrics.md · ai-incident-playbook.md

Making VSD Work at Industry Speed

Value-sensitive design process by TechTarget.

Make values operational — add Value Acceptance Criteria to user stories; pair each value with a measurable proxy. Right-size methods — 30-min Envisioning Cards; stakeholder tokens for v1; failure playbook from scenarios. Tie ethics to economics — map values to cost centers (churn, incidents, audits, TAM) and add to the business case. Put guardrails in the pipeline — Model Cards/Data Sheets at PR; red-team & bias slices in tests; release gates for sensitive features. Design for legibility — explainability patterns; calm defaults; portable, revocable consent. Keep a value change-log — one doc linking decisions to evidence and telemetry.

Work with us

Attribution & Further Reading

  • Friedman & Hendry, Value Sensitive Design: Shaping Technology with Moral Imagination (MIT Press).
  • Mollick, Co-Intelligence: Living and Working with AI (2024). See also ongoing essays on AI as coworker, coach, tutor.

Good ideas need good partners. Let's see where this one takes us.

Let's talk