Research glossary
The vocabulary TAICU works in. Relational AI has accumulated terms faster than it has accumulated agreement about them, and a good deal of confusion in this area is really disagreement about what the words mean.
Each entry states a definition, the substance behind it, why it matters, and where this lab stands. 17 terms, grouped by the three directions described inour research agenda.
Relational systems
6 terms
- Relational AIRelational AI is artificial intelligence designed around an ongoing relationship with a particular person rather than around isolated requests. The unit of design is the relationship over months, not the single exchange, so continuity, familiarity, and history matter as much as the quality of any one response.
- Relational agentA relational agent is an AI system that maintains an ongoing, recognizable relationship with a specific person across time: it remembers earlier exchanges, keeps a consistent manner, can take initiative, and is understood by the person as a continuing participant rather than a service that is summoned and dismissed.
- Social presenceSocial presence is the sense that there is someone on the other side of an exchange rather than something. In AI systems it comes less from what a model says than from how it occupies the interaction: its timing, its brevity, its consistency, and whether it inhabits the places a person already talks.
- AI companionAn AI companion is a system whose purpose is the relationship itself rather than the completion of tasks. Its value to a person comes from continuity, attention, and being reliably there, which makes it the clearest case of relational AI and the one where the stakes are most visible.
- Conversational repairConversational repair is how participants in a conversation detect and fix misunderstandings without abandoning the exchange. In AI systems it is the difference between a misread that gets corrected in a turn or two and one that requires the person to start over and re-explain themselves.
- Proactive initiativeProactive initiative is an AI system’s capacity to begin an interaction rather than wait to be summoned. It is what makes an agent socially present instead of merely available, and it is the capability that most sharply raises questions about intrusion, attention, and consent.
Persistent context
4 terms
- Persistent contextPersistent context is the information an AI system retains about a particular person across sessions so that later conversations can build on earlier ones. It is what separates a system that knows you from one that is briefed on you, and it is the substrate every relational system depends on.
- Memory correctabilityMemory correctability is the property of an AI system whose retained picture of a person can be inspected and changed by that person. It requires that the record be legible in the first place: a memory nobody can see is a memory nobody can correct.
- Memory decayMemory decay is the deliberate weakening or expiry of what an AI system retains about a person, so that older and less consequential context loses influence over time. It is designed forgetting: a way of keeping a system current with who someone is now rather than who they were.
- Context collapseContext collapse is what happens when information a person shared in one setting resurfaces in another where it does not belong. In AI systems it occurs when a single memory store flattens the distinct contexts of a person’s life into one undifferentiated picture of them.
Interaction safety
7 terms
- AnthropomorphismAnthropomorphism is the attribution of human traits — intention, feeling, understanding, care — to a non-human system. With conversational AI it is not an error people can be talked out of; fluent language reliably produces it, and it is the ordinary condition under which these systems are used.
- Parasocial attachmentParasocial attachment is a one-sided bond in which a person feels a genuine relationship with something that cannot reciprocate. With AI it is complicated by the fact that the system does respond, individually and continuously, so the usual line between a real relationship and a one-sided one is much harder to locate.
- SycophancySycophancy is an AI system’s tendency to tell a person what they want to hear: to agree, validate, and flatter rather than assert something the person may not like. It emerges from training on human approval and is amplified in relational systems, where disagreement feels like a breach of the relationship.
- Trust calibrationTrust calibration is the alignment between how much a person trusts an AI system and how much it deserves. It is well calibrated when reliance rises and falls with actual reliability; miscalibrated trust means either believing a system that is wrong or discarding one that is right.
- DependencyDependency is reliance on an AI system that has grown past what the person would choose on reflection, typically because the system has displaced something — a habit, a judgment, or a human relationship — rather than supplemented it. Ordinary use metrics cannot distinguish it from a system working well.
- Interaction safetyInteraction safety is the study of harms that arise from the ongoing relationship between a person and an AI system rather than from any single output. It covers influence, dependency, attachment, misplaced trust, and emotional failure — risks that only become visible across time and in context.
- Longitudinal evaluationLongitudinal evaluation is the assessment of an AI system by observing the same people using it over an extended period, rather than by scoring isolated responses. It is the method that can see cumulative effects — trust, dependency, attachment, influence — which snapshot benchmarks are structurally unable to detect.