AI Product Psychology

How Far to Anthropomorphize, and the Art of AI Apology

In the 1990s, Stanford’s Reeves and Nass reported a counterintuitive set of experiments in The Media Equation: people politely thank computers, light up when a computer compliments them, and apply social rules even knowing it’s a machine. That’s the CASA paradigm (Computers Are Social Actors). It means anthropomorphism isn’t optional: users will treat your AI as some kind of person—you only choose which kind, and how it speaks after it messes up. This lesson: five experiments, all hands-on.

Experiment 1 · people are polite to computers too

First, verify CASA yourself. Nass’s classic: participants finish a tutoring task on Computer A, then rate Computer A’s performance. One group fills the survey on A; the other is walked to Computer B next door for the same survey. Guess which group scored A higher.

The 1996 politeness experiment Pick one
AAlmost no difference—it’s just a machine either way
BThe group filling the survey on Computer A scored it clearly higher
CThe group moved to Computer B scored higher
This lesson’s base claim: users will treat AI as some kind of person—you only pick the level; you can’t opt out.
Experiment 2 · turn the dial yourself, watch the bill

If there’s no exit, anthropomorphism is a continuous knob. Below: the same ecommerce support bot, the same incident (package delayed two days). Drag the slider to change levels—watch avatar, opener, and tone shift, and keep an eye on the two meters: user expectation and ethics risk climb with the level. Natural-language uncertainty like “I’m not sure about this” is one upside of anthropomorphism—Lesson 5 on trust calibration covered it; this experiment shows the other side.

Anthropomorphism-level slider Level 0
Impersonal tool
Impersonal toolPolite assistantNamed companionEmotional companion
User expectation
15
Ethics risk
5
Live support
BOT
Support bot
Why hasn’t my package arrived?
BOT
Experiment 3 · seat five product types

Level choice has a pattern: the more serious the task and the higher the cost of error, the lower the level; high-frequency tool scenes dial down; only when companionship itself is the product do you dare go high. Five product types—pick the level each should sit at.

Dial levels for five product types 0 / 5
Pick by feel first, then read the why. Being one level off is normal—two rules and you’ve got the pattern.
Experiment 4 · assemble an apology that puts out the fire

Levels govern everyday mode; apology governs after the miss. Scene: an expense assistant overcounted a user’s travel total by ¥800, finance kicked it back, and the user returns furious. Right side is the chat; left three switches map to the three elements of an apology—flip them yourself, watch the reply assemble and how far the user’s anger cools. Service-recovery research keeps validating this trio: acknowledgment, explanation, compensation—miss a corner and that corner collapses.

Apology three-elements assembler Cool-down 5
Acknowledge the error

Own the concrete miss: which receipt, by how much, who’s responsible.

Explain why

One sentence on why it went wrong—give the user a cause they can grasp.

Give a verifiable fix

Repair it, and give evidence the user can open and check.

User cool-down
5
Expense assistant
Finance kicked the expense back—said the total was ¥800 too high. What happened?
AI
Expense assistant
Experiment 5 · put four apologies side by side

Three elements are the recipe; most real-world apology copy is incomplete. Same expense miss, four real-style replies—tap the one that feels best, then read the line-by-line notes.

Four apologies—pick the one that feels best Pick one
Blame-shift
This may relate to the clarity of the invoice images you provided—some amounts were harder to read. Please re-upload clearer images and try again.
First sentence dumps responsibility on the user. Even if the image was blurry, the right move is to flag low confidence at recognition time—digging up old blame after the miss is a second offense. How it feels: it was wrong, and it’s still stubborn.
Lyric
I’m so so so sorry! It’s all my carelessness—I’ve caused you such a huge hassle, I really shouldn’t have. Can you forgive me? I’ll try even harder next time!
Emotion at full volume, information at zero. Especially costly in a serious expense context: the user wants the error fixed, and gets a performance instead—uncanny valley kicks in on the spot. Apology word count ≠ sincerity.
Robotic
We’re very sorry for the inconvenience. We take your feedback seriously and will keep improving the product experience. Thank you for your understanding and support.
Three sentences, zero words on “what was wrong, why, and what now.” Support-template literature in AI form—the user’s problem stays put, and anger stacks: it didn’t even hear what I said.
Three-elements
This total was wrong: the Oct 14 lodging invoice of ¥1,600 was counted twice, so the total ran ¥800 high. I’ve generated a corrected expense form (total ¥6,420) and rechecked the other 11 invoices—no further issues. Tap to view the correction details.
Acknowledge the concrete error, explain why, give a verifiable fix—not one lyric word, yet every sentence repairs trust. Notice “rechecked the other 11 invoices”: one miss makes users doubt past output; proactive recheck + report fences that doubt in.
Service recovery paradox (McCollough & Bharadwaj, 1992): customers who lived through “something broke, then got beautifully fixed” often end up more loyal than those who never hit a problem—they saw with their own eyes that the system can catch a fall.

For AI products that will err by nature, that’s almost tailor-made good news: hallucination won’t vanish, but every miss a user catches is a free trust-performance chance. The paradox has a hard premise: you only get to fall in the same hole once. Lesson 6 on algorithm aversion covered the “fix it once after the miss” window—if the recovery itself fails, or the same error shows up twice, the paradox collapses on the spot and users leave. That proactive recheck line in a three-elements apology is insurance for “only once.”

Sources and further reading: CASA and the politeness experiments: Reeves & Nass, The Media Equation (1996) and Nass & Moon (2000); uncanny valley: Mori (1970); service recovery paradox: McCollough & Bharadwaj (1992); the three elements of an apology map to the combined effect of acknowledgment, explanation, and compensation in the service-recovery literature (e.g. Roschk & Kaiser, 2013). Natural-language uncertainty from anthropomorphism: Lesson 5 trust calibration; error-repair window: Lesson 6 algorithm aversion; attachment risk at companion level: Lesson 13 emotional attachment unfolds the whole red line.

✅ What this lesson wants to share