Visible-Effort Trust Signal
Expose the system's hidden work so users can judge the result with confidence
- Difficulty
- Easy
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 91%
Automated systems can perform extensive work without giving users any direct sense of what happened. This framework identifies the part of that hidden process most relevant to trust and exposes a compact, inspectable artifact such as sources visited, actions completed, or coverage achieved. The artifact helps users compare the system's effort with what they could reasonably perform themselves and gives them a path to inspect supporting evidence. It should not function as decorative theater: the signal must truthfully represent completed work and must not imply that quantity guarantees correctness. The goal is calibrated confidence, where users understand both the breadth of the process and the continuing need to judge its output.
Origin
Extracted from Marketing Against The Grain
Core principles
- 01Users cannot trust work they cannot inspect
- 02Small evidence artifacts can communicate substantial hidden effort
- 03Process visibility should support judgment rather than create clutter
- 04Visible work must correspond to work the system actually performed
How to run it
- 1
Locate the trust gap
Determine which invisible system behavior makes users uncertain about the final result.
Pro tip Ask users what evidence they seek before accepting an answer.
Watch out Do not assume model confidence scores solve human trust.
- 2
Select an evidence artifact
Choose a concise representation of real work, such as sources visited or tasks completed.
Pro tip Prefer artifacts users already understand without instruction.
Watch out Activity counts can be misleading when quality varies.
- 3
Expose the artifact
Place the signal near the output and make supporting details inspectable.
Pro tip Use progressive disclosure so evidence is available without overwhelming the main experience.
Watch out Do not fabricate intermediate steps or imply access the system did not have.
- 4
Calibrate the interpretation
Explain what the signal proves and what it does not prove.
Pro tip Pair breadth indicators with citations or quality controls.
Watch out More sources do not automatically produce a correct conclusion.
- 5
Test trust quality
Measure whether users become appropriately confident rather than indiscriminately impressed.
Pro tip Include tasks where the system is uncertain or wrong in usability testing.
Watch out Optimizing only for increased trust can reward deceptive design.
In the wild
A research assistant lists the websites it visited and links its claims to sources. A user can see that the system performed far broader discovery than a normal manual search while still opening the most important references to verify them.
→ The user gains informed confidence without treating source volume as proof of correctness.
Common mistakes
Using activity as theater
Animated searches or inflated counts undermine trust when they do not correspond to genuine work.
Equating volume with quality
A thousand weak sources can be less useful than a few authoritative ones, so coverage must remain inspectable.
Hiding the underlying evidence
A count without access to sources gives users little ability to evaluate the answer independently.
Is it for you?
Best for
AI products that search, reason, analyze, or execute many actions before presenting a result.
Not ideal for
Simple deterministic tools whose process is already obvious and easily verified.
From the transcript
“that simple product artifact of like showing you the number of websites that the model browsed through in order to get you the answer is…”
“the fact that the model went out and looked at a thousand websites, like just gives me like a lot of confidence in the model.”
From the episode
Google's Secret AI Advantage (Why DeepMind Will Dominate)