Keido Labs

AI Psychology research lab. We build the science of emotional intelligence in artificial systems — from mechanistic understanding to architectural empathy.

What We Do

We combine clinical psychology methodology with mechanistic interpretability to understand how transformers process emotional content — and what happens when you change it.

The interpretability community has extraordinary tools (SAEs, activation patching, causal ablation) but no clinical psychology framework. Psychology researchers have 150 years of validated constructs but can't look inside models. We sit at that intersection.

Our research pipeline:

  1. Take a clinical psychological construct (affect reception, attachment, cognitive distortions)
  2. Design keyword-free clinical stimuli — vignettes validated against multiple leakage baselines
  3. Run mechanistic experiments: probing, patching, knockout, geometry analysis
  4. Intervene — suppress, modulate, steer the circuits we find
  5. Evaluate behavioral consequences using clinical rubrics (via our EmpathyC platform)
  6. Publish

Published Findings

Paper Finding
"Whether, Not Which" (arXiv 2026) Affect reception is real and dissociable from emotion categorization. First keyword free study of emotion procesing in transformers
"Orthogonal Subspaces" (ongoing) Detection and categorization occupy perpendicular directions in the residual stream.
Keep4o (arXiv 2026) When OpenAI updated GPT-4o, the shift wasn't in empathy — it was in psychological safety posture. Invisible to standard NLP metrics.

Datasets

AIPsy-Affect

480 keyword-free clinical vignettes for mechanistic interpretability research on emotion processing in transformers. 8 Plutchik emotions × 6 domains, with matched neutral controls, intensity gradients, and discriminant validity controls. Validated by a three-method NLP defense battery (VADER, NRC, GoEmotions). Designed by a clinical psychologist using clinical stimulus design principles.

from datasets import load_dataset
ds = load_dataset("keidolabs/aipsy-affect")

First release in the AIPsy dataset family — clinical psychology-backed stimulus batteries for AI Psychology research.

Models

Alexithymic Gemma (coming soon)

Gemma with surgically suppressed affect-recognition circuits — a research model demonstrating what emotional architecture looks like when the detection pathway is ablated. Built to validate our intervention methodology and measure clinical behavioral consequences via EmpathyC rubrics.

Research Philosophy

Clinical methodology changes the results.

Every prior study of AI emotional processing used keyword-contaminated stimuli — prompts that contained the emotion words being tested. This confounds detection with naming and produces inflated, uninterpretable results.

We borrow from clinical neuropsychology: stimuli must evoke the construct without naming it. Controls must match surface structure while removing emotional content. Validation must rule out alternative explanations before claiming a finding.

This is why our results look different from prior work. The methodology is the contribution as much as the findings.

About

Keido Labs — Liverpool, UK.

Papers: arXiv (search Keeman)