AI Product Psychology

Labor Illusion: Make AI Show Its Work

Lesson 2 said waiting needs feedback. This lesson’s takeaway is more counterintuitive: some waiting shouldn’t be eliminated at all. Products that show their work make users wait a few extra seconds—and satisfaction and trust go up. The effect is called the labor illusion. Run the experiments first, then the research, then draw the boundaries.

Experiment first · Two bots—which do you trust?

Send the same question to two support bots at once. The answers are word-for-word identical; the only difference is how you get them. A replies instantly; B spends 3 seconds lighting up a work log line by line before answering. Then vote by gut.

Instant reply vs. visible process Run, then vote
Hit “Ask both,” watch both sides, and notice which way your gut leans.

Bot A · Instant reply

Answers in 0.3 seconds

“What risks are in this contract’s non-compete clause?”
Hit “Ask both” below to start

Bot B · Show the process

3.5 seconds, work log lights up line by line

“What risks are in this contract’s non-compete clause?”
Hit “Ask both” below to start
Both answers are identical
The research · The site that made users wait 60 seconds won

The effect comes from a 2011 Harvard Business School experiment by Buell and Norton. Subjects searched flights on a simulated site: one group got results instantly; the other waited 30–60 seconds while the screen scrolled “Checking Delta… Checking United…”. The group that saw the labor scored higher on satisfaction; under some conditions subjects preferred the site that made them wait 60 seconds but showed the work. Below, run the experiment yourself.

Recreate Buell & Norton’s flight experiment 0 / 2 versions tried
Search once on each version. In the real study the labor version waited 30–60 seconds; here we compress it to a few seconds—the idea is what matters.
Shanghai PVG → Seattle SEA · Oct 14
Pick a version, then hit “Search flights”
Satisfaction for this version
--
Mechanism in one line: people use visible effort to infer invisible quality. Users can’t see whether a search was thorough or sloppy, so they treat “how much work appears on screen” as a proxy for quality. A consultant who hands you a plan on the spot looks cheap; saying “let me analyze this for three days” looks professional—same idea.
Hands-on assembly · Put AI’s effort on stage

Engineering already gives you three levers: a reasoning model’s thinking process, RAG retrieval sources, and an Agent’s tool-calling log. They started as debug info; put them on stage and they become experience assets. Flip the three switches one by one—watch what appears in the chat on the right, and how much felt trust climbs.

Visible-process assembler Trust 42
Visible thinking

Open the chain of thought: which cases it split, where it changed course—let users see the skeleton of the reasoning.

Light up retrieval sources one by one

“Retrieved 3 sources” plus clickable entries—hang the answer on checkable citations.

Tool-calling log

What it searched, read, and compared—one step per line on screen.

Felt trust
42
“It answered fast, but would I forward this to HR as-is… maybe I should ask someone else.”
Look this up for me: an employee is let go during probation—how is compensation calculated?
Thought for 4 seconds ⌄
The user is asking about compensation when probation is terminated… split into two cases: the company can prove “failed to meet hiring conditions,” and when it can’t… look up the Labor Contract Law text first; don’t invent numbers from memory.
Retrieved 3 sourcesLabor Contract Law Art. 21Labor Contract Law Art. 39SPC Interpretation (I) on Labor Disputes
✓ search_law("probation termination compensation") · 3 results✓ read_doc(Labor Contract Law §21, §39)✓ cross_check(Interpretation I §44)
Two cases: if the company can prove you “failed to meet hiring conditions,” it may terminate with no compensation; without proof it’s unlawful termination and pays the 2N standard. The burden of proof sits with the company—you only need to keep the offer letter and evaluation records.
With everything on, tally what you got—it’s more than one illusion: ① Visible effort creates the labor illusion—this lesson; ② Sources and logs give users a verifiable handle—the star of Lesson 5 on trust calibration; ③ Users incidentally learn how the system works, and a mental model grows on its own—Lesson 8’s theme. Three birds, one stone: top products put debug info on stage precisely for these three layers.
Sharp eyes · Catch the fake labor

Process display can be real or fake. The three UIs below all “show effort.” Judge each one: real labor, or fake performance? The cost of getting caught is revealed after all three.

Fake-progress spot-the-difference 0 / 3
Samples ② and ③ are live—watch a few seconds before you decide.
Sample ① · Retrieval log
✓ Search “resale deposit refund” · 4 docs
✓ Read Civil Code Art. 587
✓ Compare 2 similar cases
A deposit is refundable if the other party breaches or both agree to unwind [Civil Code §587]. Your case is seller breach—you can claim double return.
Sample ② · Deep thinking
Deep thinking…
Thinking for 0 seconds
Sample ③ · Progress bar
4%
Why fake labor is a trust bomb: Get caught once, and users generalize the trick to every corner of your product—real retrieval logs and real thinking get scrutinized as performance first. The labor-illusion dividend rests on “the labor actually happened.” Fake it once, and the real stuff devalues with it.
Three boundary questions · Clear three gates before you perform

The labor illusion works—but it has boundaries. Three questions, each a practical red line. Clear them and you can use this lesson’s levers with confidence.

Q1 · When should you hold the performance back Pick one
In which of these three scenarios should you hide the process and go straight to the result?
AA first-time user generating an annual business analysis
BA translate button users hit dozens of times a day
CLegal-clause review where users can’t easily verify right vs. wrong
Q2 · How long is too long to perform Pick one
Where’s the red line for how long process display can run?
APerform as long as the user can stand it
BPerform at most as long as the real labor took
CA bit longer than competitors, to look more thorough
Q3 · Can you deliberately slow things down Pick one
The backend finished in 0.4 seconds. A PM proposes hard-coding a 3-second delay with an “effort animation.” Take it?
ATake it—the study proves slower makes users happier
BReject it—the study assumes real labor; a forced 3 seconds is fakery
CCompromise: add 1.5 seconds—lower chance users notice
Sources and further reading: The main labor-illusion study is Buell & Norton (2011), The Labor Illusion: How Operational Transparency Increases Perceived Value, Management Science. Later work framed it as operational transparency; Buell, Kim & Tsay (2017) showed in restaurant settings that two-way transparency can raise both customer satisfaction and employee performance. A premise often skipped in the original paper: the labor illusion only works when the result is acceptable. If the answer bombs, the effort users watched becomes a new complaint: you checked 12 docs and still got this?

What this lesson wants to share