SafeSight
Making a safety-sensitive AI experience easier to understand—and harder to overtrust
Role: UX contributor focused on onboarding and guidance
- Direct artifacts
- Safe Sight Get Started Guide and SafeSight visual artifact
- Context
- AI-assisted historical safety analysis, trends, visualization, and predictive direction
Choose a safety topic to explore
Dashboard
Explore trends from historical safety reports.
- Tarmac Damage TrendSurface and equipment-contact reports
- Ground equipment eventsRamp vehicles and tooling
- Weather-related reportsSeasonal operating conditions
Recent chats
- Tarmac Damage Trend
- Ramp equipment events
The 30-second read
- Who was doing the work
- Safety analysts asking questions of historical safety reports.
- What I owned
- Onboarding and guidance: authored the Get Started Guide and created the SafeSight visual artifact.
- What changed my approach
- Added structure—scope, limits, and review cues—at the cost of a frictionless first answer.
- Artifact evidence
- The Get Started Guide and the SafeSight visual artifact.
- Where it landed
- Not a shipped-product or safety-performance claim. Predictive capability was a direction, and a May 2025 record shows the work being put on hold.
Guiding analysts through an AI safety tool
Guiding analysts through an AI safety tool
SafeSight was an AI-driven safety-analysis experience intended to help analysts ask questions about historical safety reports, identify patterns, generate visualizations, and explore proactive or predictive signals.
My directly documented contribution was onboarding and guidance. I authored the Get Started Guide and created a SafeSight visual artifact. That work required translating a complex AI capability into practical questions, expectations, and feedback behavior without suggesting that an AI-generated result should replace human safety judgment.
A confident answer can invite overtrust
A confident answer can invite overtrust
The product’s promise created the central UX risk. Faster access to safety information could be useful, but a confident answer or visualization could also be mistaken for a complete or verified conclusion.
Users needed to understand what to ask, what data the system was drawing from, how to interpret trends, when to ask a follow-up question, and where human review remained necessary. Differences in role and data fluency made a one-size-fits-all explanation unlikely to work.
A fluent answer still needs boundaries
Recreated for portfolioTarmac Damage Trend
Summary
Tarmac damage reports increase across the period, with most activity during ground handling.
Important things to know
- Reports rose in the second quarter, concentrated around pushback.
- Most entries describe minor surface contact.
- Reporting volume changed during the period.
Treating the prompt as part of the interface
Treating the prompt as part of the interface
The onboarding work treated the prompt as part of the interface. Guidance needed to help analysts form useful questions, understand the distinction between historical retrieval and predictive direction, and provide feedback when an answer was incomplete or misleading.
The experience principles were:
- Make question framing concrete and role relevant.
- Expose the scope and limits of the available data.
- Present trends and visuals as evidence to interpret, not decisions to accept.
- Keep human review visible before operational action.
- Give users a clear feedback and support path.
- Describe predictive capability as a direction, not a proven safety outcome.
Onboarding and guidance states
Representative concept1 · Ask a concrete question
Alaska Safe SightWhat would you like to explore?
- Trends in a report category over time
- Compare locations for a period
Show tarmac damage trends for the last two quartersSubmit2 · See the data scope
Alaska Safe SightWhat this answer draws on
- Historical safety reports
- Selected date range only
- Predictive views: a direction, not a forecast
3 · Interpret with uncertainty
Alaska Safe SightSummary
Here’s a pattern worth reviewing: reports rise around pushback.
Some reports lack location data. Ask a follow-up or check the source reports.
4 · Review sources before action
Alaska Safe SightAnalyst review before any operational action.
Open source reportsCopy and feedback controlsFlag as incomplete or misleading → feedback to the team.
Guidance that keeps people in charge
Guidance that keeps people in charge
The direct evidence supports my authorship of the onboarding guide and visual artifact and my contribution to making the experience understandable. Product materials describe beta, training, reactive and predictive questions, and later research planning. A May 2025 record states that work was being put on hold.
My scope was the guidance layer—how people learn to ask, interpret, and question the system—rather than the model, the roadmap, or safety-performance results.
Evidence to interpret—not a decision to accept
Recreated for portfolioTarmac Damage Trend
Reports rise in every phase; gate activity shows the largest change. This is a pattern to investigate, not an explanation.
Human decision
Review source reports before operational action.
Ask SafeSight about safety trends…
SubmitResponsible AI is the structure around the answer
Responsible AI is the structure around the answer
Responsible AI UX is not only a warning message. It is the structure around the answer: better questions, visible limits, interpretable evidence, human accountability, and a feedback loop that makes uncertainty actionable.
Want to see how the guidance handled uncertain AI answers?
The Get Started Guide and visual artifact are available to share privately, under NDA where needed. Happy to talk through the guidance.
Let’s Connect