Picture a product meeting where somebody says, “We just need more empathy.” Everyone nods. The meeting ends. Nothing changes. Somewhere, a values poster gets a promotion.
Empathy has become the good-faith shorthand we reach for when a product feels like it has wandered away from the people using it. It is not wrong, exactly. It just isn’t enough. Empathy is how you feel what someone else is feeling. It is a sensing organ. A sensing organ, by itself, does not decide anything.
The sensing organ is not the decision.
Compassion is the step after empathy. Compassion is what you do about what you felt. It is the willingness to take responsibility: for the onboarding that confuses people on their worst day, for the error message that shames a user who was already anxious, for the cost you pushed onto someone else’s afternoon because it was cheaper than fixing it inside your system.
Empathy is “I understand how you feel.” Compassion is “I will change what I’m doing because of it.” One of these is free. The other is a ticket in next sprint.
Product work is almost entirely the second one. You can have infinite empathy for a user and still ship a product that punishes them, because empathy does not, by itself, make you pick one tradeoff over another. Compassion does. Compassion closes the loop, or at least puts “close the loop” on the board.
What it looks like in the work.
The easiest place to see it is in the defaults. A compassionate default assumes the person on the other side is doing their best inside imperfect conditions. It forgives small mistakes. It makes the correct path the obvious path. It does not ask the user to compensate for the team’s missing work.
- The confirmation dialog that doesn’t gaslight you for pressing delete.
- The billing page that tells you, before you ask, what you’re about to be charged.
- The AI reply that says “I’m not sure” instead of confidently fabricating an answer.
- The form that saves your work when you navigate away and, in a controversial move, tells you it did.
None of these require a new insight. They require someone on the team deciding that the user’s situation is a responsibility, not a design constraint.
Why this matters more with AI.
The stakes rise when the product is a model. A language model can produce a confident answer without adequate evidence; the work of making it behave, of making it admit what it does not know, of making it hand off, of making it refuse, is compassion expressed as engineering. Reliability, in this sense, is a form of care.
Safety and cost are usually framed as constraints on creativity. I think they are part of the creativity. The AI products I want to build treat reliability as an ingredient, not a tax. They are the ones that feel like they were made for someone, rather than at someone.
A working rule.
When a team gets stuck between two options, fast versus right, cheap versus coherent, easy to ship versus easy to live with, I try to ask a small compassion question. It goes like this: Who is this going to be hard on if we get this wrong, and are we okay with that?
Usually, the answer reveals that one of the options isn’t really an option at all. The choice becomes obvious. It doesn’t resolve every disagreement, but it changes the thing we’re disagreeing about: from what’s convenient for us, to who we’re asking to carry the cost.
That small shift, repeated over years, is what I think I mean by taking the work seriously. Not feeling more. Deciding more carefully.
Turn the concern into a design decision
Consider a person who loses a long form after a connection failure. Feeling bad about it does not restore the text. The useful work is deciding what the system saves, how the person recovers it and how clearly the interface explains the outcome.
| Human difficulty | Product decision | Check |
|---|---|---|
| Lost work after interruption | Save progress and explain recovery | Can the person resume? |
| Unclear error | Name the problem and next action | Can someone fix it without guessing? |
| Unsupported AI answer | Show uncertainty and a useful handoff | Does the fallback preserve context? |
W3C’s writing guidance for accessibility gives practical advice on clear instructions and error messages. My argument about compassion is a product perspective, not a formal definition from that guidance.
The same concern runs through AI that behaves in the real world and early product decisions. Care becomes useful when somebody can point to the behaviour it changed.
