Human-AI Interactions Research
Issue 00 / Pre-Launch
A differend would be a case of conflict, between two parties, that cannot be equitably resolved for lack of a rule of judgement applicable to both arguments. Jean-François Lyotard
The Manifesto
I. When humans and AI speak, make sure the human is heard.
AI speaks in statistical probability, token vectors, loss functions, and GPU efficiency. Humans speak in intuition, implicit knowledge, and moral nuance. Naturally, the companies building these systems judge the human-AI dialogue in the language they already own: dashboards.
Clicks, latency, retention curves, deflection rates. But a session that closes clean can still close wrong, hiding a failure of human comprehension or unarticulated user frustration. This creates a differend - a structural conflict where the human experience is flattened because it is translated entirely into machine telemetry. “The ‘regulation’ is done in the idiom of one of the parties while the wrong suffered by the other is not signified in that idiom.”
Clicks are not clarity. Retention is not trust or safety. We no longer study a product. We study a relationship.
II. Verify what you are verifying
Something that is not quantifiable makes computer scientists very uncomfortable | Brian Christian, The Alignment Problem
For decades, software engineering relied on deterministic logic and predictable test suites. AI models, by contrast, are non-deterministic - constantly evolving around training datasets and live user interactions.
Automated evals and synthetic benchmarks can verify raw model capabilities, but they cannot evaluate whether technology is fulfilling human intent or silently reshaping human thinking around model constraints.
It’s time to recalibrate our approach to measuring success. Relying exclusively on telemetry leaves a critical blind spot - the why behind user churn, cognitive friction, or trust and reliance miscalibration.
Qualitative research is no longer a traditional UX convenience; it is the core strategic lens for evaluating the live negotiation between human intent and machine probability.
III. Research with humans is an always-on practice: the stakes are too high for point-in-time audits
Traditional user research operates on periodic, point-in-time audit cycles - an approach fundamentally mismatched with human-AI interaction. Probabilistic models exist in an "always-in-beta" state; they drift, self-evolve, and update weekly across millions of unique contexts. Auditing a system quarterly/annually when the underlying model changes weekly leaves organisations in a perpetual blind spot.
To build safe, resilient, and indispensable AI products, qualitative research cannot remain a periodic event. It must become an always-on operational pipeline running in parallel with machine telemetry - continuously probing interaction boundaries, evaluating human-in-the-loop hand-offs, and capturing unarticulated friction before it escalates into product failure.
Human-AI interaction is a high-stake bet that doesn’t tolerate opportunistic shortcuts. Measure the model. Listen to the person. Repeat.
The Differend exists to ensure the human voice is not lost in computational translation as AI scales. We democratize, popularize, and operationalize rigorous, research-backed frameworks specifically tailored for Human-AI interactions.
As software evolves from passive tools into autonomous agents, we provide product leaders with actionable evidence to ensure human agency, authority, and intent remain central to product design