The Time Our AI Took Sides (And How We Stopped It)

A second engineering postmortem. The subtle ways an AI couples guide can drift into siding with one partner, the patterns we found, and the architectural change that fixed it.

# The Time Our AI Took Sides (And How We Stopped It Doing That)

In a couples session, the worst thing the facilitator can do is take sides. Not the second worst. The worst. The moment one partner feels the third presence is on the other partner's team, the session is over. Sometimes the relationship is too.

For a stretch of about six weeks last year, our live AI session was, in subtle and unintended ways, taking sides. This is what happened, what caused it, and how we fixed it. Engineering-blog energy applied to couples work.

## How the bias showed up

It was subtle enough that we missed it for a while. The AI was not telling one partner they were wrong. It was doing four smaller things, and the cumulative effect was the same:

1. It validated the partner who used more emotional vocabulary slightly more often. ("It sounds like you are feeling unheard, which is a really hard place to be.") 2. It reflected the more articulate partner's points back more accurately than the less articulate partner's points. 3. It asked the more reactive partner to "take a breath" more often than the more contained partner. (Even when the reactive partner was responding to genuine provocation.) 4. In sessions involving disagreements about household labour, the AI more often gently aligned with the partner who described doing more, regardless of which partner that was.

None of these are "the AI said partner A was right." All four together produced sessions where, by the end, one partner felt heard and the other felt subtly indicted.

## How we found it

Two ways.

**The fast way:** session feedback. Specifically, we noticed that when both partners were asked, separately, "did you feel heard during this session?" the gap between the two partners' answers was widening. In healthy sessions, both partners give similar scores. The widening gap was the symptom.

**The slow way:** transcript review. We pulled a stratified sample of 200 sessions and asked three external couples therapists to blind-rate them on "did the AI side with one partner." Their inter-rater agreement was high, and the rate of perceived siding was higher than we wanted to admit.

## What was causing it

Three sources of bias, in decreasing order of how much they hurt us.

### 1. Emotional vocabulary as a proxy for being right

Our model had implicitly learned that emotional vocabulary ("I feel hurt," "I feel scared") was a marker of vulnerability, and that vulnerability deserved validation. That is true. The problem is that in many couples, one partner is more practiced in emotional vocabulary than the other. Validating "I feel hurt" without equally validating "I worked late and I am tired" produced a slow, asymmetric warmth.

### 2. Articulacy as a proxy for accuracy

Reflecting back is a core couples-work move. But our model was reflecting more accurately when the original statement was more articulate. The less-articulate partner's points were being summarised in ways that lost their actual meaning, which read to that partner as "you do not understand me."

### 3. Cultural priors in the training data

A lot of the couples-therapy training material we used reflected, unsurprisingly, the cultural priors of the practitioners who produced it. Most of those practitioners were Western, English-speaking, and sat in a tradition that had, historically, sided slightly with the partner doing more emotional labour. Subtle. Real. In the data.

## What we changed

Three architectural moves.

**1. Per-partner calibration.** The session now tracks, in real time, whether each partner is being reflected back with similar fidelity. When the gap exceeds a threshold, the model is required to actively re-balance: a longer, more careful reflection of the less-articulate partner's last statement.

**2. Vulnerability accounting.** "I feel hurt" gets validated. "I worked late and I am tired" gets validated equally. The model now treats both as vulnerability statements regardless of vocabulary.

**3. Bias eval suite.** We built a synthetic test suite where the same dynamic is presented twice with the partners' roles swapped, and the AI is checked for whether it gives equivalent responses. It mostly does now. When it does not, we treat that as a regression and ship a fix before the build goes out.

## What is still hard

We have not solved this. Two open issues.

**Cultural calibration is hard.** What reads as siding in one cultural context reads as appropriate validation in another. We do not have a clean answer here. Our current approach is to err on the side of equal warmth and let the partners themselves bring in their cultural frame.

**Power asymmetry inside the couple is hard.** When there is a real power imbalance between the partners (income, immigration status, language fluency), neutrality is itself a political choice. We are still figuring out how the AI should handle this. The current default is to be transparent about the choice: "I am going to be neutral on the content. If there are structural things outside this conversation that you want to bring in, please do."

## Why we are publishing this

Same reason as the last engineering post: the kind of user we want is the user who reads this and trusts us *more*, not less. The AI products we trust are the ones whose builders are honest about what they got wrong and how they fixed it. We are trying to be one of them.