Why We Built Receptiviti Labs

The human-AI interaction is psychological in nature, and language is the vehicle that conveys the psychological information. Measure the language, and the psychology of the interaction becomes visible.

Receptiviti has spent a decade measuring exactly this. The confidence, doubt, urgency, cognitive load, or distress a person brings to a conversation is carried in the words themselves, in the small, structural features of language that reflect a person's cognitive and emotional state. The science behind measuring it, LIWC, was invented by Receptiviti co-founder James Pennebaker and has been validated across three decades of peer-reviewed research. That research now sits behind more than 34,000 independent citations and all of Receptiviti's dimensions. Receptiviti was founded in 2015 to make that science available to organizations that need to understand the people who matter to them.

Receptiviti Labs, established in 2026, is focused on deeply understanding the human-AI interaction: what these systems infer about the people they interact with, and the dynamics of the relationship that forms between them, so these systems can respond more effectively, support people better, and operate more safely.

The field now has a name for the problem this creates. Kirk and colleagues call it socioaffective alignment: aligning a system while accounting for the reciprocal influence between the model and the user's psychological state. OpenAI has studied affective use and emotional dependence at scale, Anthropic has published on sycophancy and personal guidance across hundreds of thousands of conversations, and the 2026 International AI Safety Report devotes a section to emotional dependence on chatbots. The problem is now a recognized frontier.

The unsolved part is measurement. The frontier labs have started building classifiers for dependency, distress, and problematic use, which is exactly the right instinct. Almost all of it works by having a model classify the user's state, one model judging another's inputs. That judgment comes from the same kind of system being judged, so nothing validates whether the judgment is accurate. Model-produced classification also varies across versions and phrasings; the same conversation can yield a different label on a different day. It influences the response without ever becoming a variable that can be inspected or verified.

That's the gap Labs exists to close. We're building the measurement layer for the human side of AI: validated measurement of a user's cognitive and emotional state, derived from language, independent of the model being evaluated, and structured so the systems responsible for a model's behavior can see and respond to it. The same input yields the same dimensional score, and every dimension traces back to published science. That independence is what makes the measurement useful: a model cannot generate independent evidence about itself, and independent evidence is what governance, auditing, and correction require.

Here is what we think is true, and why it points at measurement.

1. Human agency is the variable these systems are least equipped to protect. When people hand their reasoning, judgment, and decisions to an AI, those capacities can weaken from disuse. Peer-reviewed work already documents over-reliance, reduced critical engagement, and eroding epistemic independence as measurable effects of sustained AI use. A system that genuinely preserves a user's agency will sometimes need to introduce friction, prompt reflection, and slow the user's deference, even when a smoother, more compliant response would feel more helpful in the moment. But a system can only know when to do that if it can see how its interactions are affecting the user over time. Without that signal, a system defaults to being fast, confident, and frictionless, which over time is what erodes agency.

2. The signals that are critical to alignment, user safety, and the viability of these systems are the ones being measured least rigorously. Every conversation carries a layer of linguistic information about the user's cognitive load, emotional state, distress, and psychological risk. It's some of the most consequential information in the interaction. The models already respond to it, they infer something about the user's state and adjust accordingly, but that inference is inconsistent, sometimes inaccurate, and exists only inside the model. There is no mechanism for inspecting what it concluded, checking whether it was right, or detecting when it went wrong. Making user state an explicit, validated variable turns those opaque judgments into something you can inspect, audit, and correct.

3. Understanding how AI understands users may be the most underexplored entry point in interpretability. A model's inference about the user happens between input and output, and it influences every response. Measuring it outside the model has a further advantage. Models are increasingly able to tell when they're being tested, and some evidence suggests they can behave differently under evaluation than in ordinary use, a problem the labs themselves have started flagging. A measurement taken from the user's language doesn't inherit these problems, because it doesn't come from the model. It isn't subject to the model's biases, its tendency to hallucinate, or its awareness of being evaluated. It's computed from the words the user actually wrote, the same way every time, and grounded in validated psycholinguistic research. That gives it a stability and empirical grounding the model's own inference doesn't have. Making the interaction's psychological state visible, explainable, and consistent is one of the more practical paths to understanding what these systems are doing: how they perceive the people using them, and why they respond the way they do.

4. Real alignment will require systems to act on measurement of what's happening to each user, across time. A system tuned for the average user is misaligned with almost every actual one. We think alignment has to become user-specific, grounded in what is happening to a particular person, in the current session and across their whole relationship with the system. That needs a kind of measurement that barely exists today: measurement that can see cognitive load rising across a session, confidence eroding conversation by conversation, dependency deepening over weeks. None of that shows up in a single interaction. It only appears across time.

None of these positions require new model architectures or retraining; they require a measuring signals that already exists in the language, made explicit and put where the systems governing the model can use it. That signal is derived from the interaction itself, computed the same way every time, and traceable to published science, which is what separates measurement from the model's own changing inference about itself.

That's our focus. Field Notes is where we'll think out loud about it: engaging with the research we find compelling, the arguments we're still working through, the people doing serious work in this field, and what we're learning as we build. Some of it will be technical, some of it argument, some of it us working through problems in public.

If any of this is close to what you're working on, we'd like to hear from you.