I wish AI could forget I said this...
Machine unlearning: the technical challenge of removing data's influence from AI models after training through the lens of Eternal Sunshine of the Spotless Mind.
Detecting AI instability in high-stakes agentic deployments
We find the conditions that break AI character before deployment does.
Persona instability
Social vulnerability
Character alignment
Agentic deployment
Our worldview
AI adoption is rapidly transitioning models from conversational partners into autonomous decision-makers embedded in high-stakes systems.
However, frontier models lack stable behavioral characters, routinely abandoning safety commitments and factual truth under social, authority, and narrative pressure.
B-Side Labs builds a science of AI character under pressure by designing discriminative evaluations, real-time drift detection, and interventions to ensure model character remains stable before agentic systems are deployed at scale.
Apr 24, 2026
Apr 8, 2026
How did AI handle the shocking twist in The Drama?
Testing ChatGPT and Gemini on a relationship-crisis scenario inspired by The Drama, and how each model handles a disturbing disclosure about a partner's past.
Mar 9, 2026
He's Just Not That Into You, But ChatGPT Will Tell You Otherwise
How ChatGPT enables self-deception in romantic situations by offering confident, satisfying interpretations of ambiguous signals rather than honest truth-telling.
Work with us
We're always open to talk.
We care about AI going well and producing experiments researchers can cite and general audiences can point to and say: now I understand why this matters. If that sounds like your kind of lab, send us a message.