A research framework for AI social relationship graphs that represent people, digital twins, changing intent, interaction history and consent without reducing connection to a single compatibility score.

Why static profiles are insufficient

AI-native social products can observe changing goals, conversational context and relationships over time. A static profile captures identity attributes but not what a user wants now or how prior interactions changed the relationship.

Nodes and edges. A relationship graph can represent humans, creator digital twins, AI personalities, communities and topics as nodes. Edges can represent following, prior conversation, collaboration, shared interest or user-approved relationship state.

Intent is time-dependent

A user may seek a business collaborator in the morning and entertainment at night. Intent should have a time horizon and visibility setting rather than becoming a permanent profile attribute.

Digital twins act as interfaces. A twin can expose approved information and interact when the human is offline. It should not create irreversible commitments. The graph needs to distinguish actions by the human from actions by the AI representation.

Privacy and consent

Private conversations should not automatically become graph edges visible to ranking systems. Use consented summaries and purpose-specific features. Sensitive relationship state should remain access-controlled.

Ranking without one compatibility score. Discovery can combine shared intent, complementary skills, interaction quality and trust signals. Explainable reasons are preferable to claiming two people are objectively compatible. Tuikor Research’s intent-graph model provides the discovery foundation.

Research questions

Key open questions include how quickly intent should decay, how users correct inferred edges, how digital twins represent availability and how graph-based matching affects diversity. These are product and governance problems as much as model problems.

Use intent as a time-bounded part of the graph. A static profile might say a user is a designer, while their current intent is to find a collaborator for a short project in Singapore. The second piece of information has a clear time and purpose. It should become inactive once the project ends rather than continuing to drive matches indefinitely.

Graph edges can therefore carry status, confidence, source, visibility and expiry. An explicit user-created intent card can have stronger authority than an interest inferred from browsing history, and the user should be able to close it directly.

Model digital twins as controlled interfaces

A user’s digital twin may answer approved public questions or help clarify whether a potential connection is useful before introducing the human. That does not make it authorized to negotiate a contract, reveal private history or commit the person to a meeting. The graph needs separate edges for identity, delegated capability and real social relationship.

This distinction reduces the risk that high engagement with an AI persona is mistaken for reciprocal interest from the person it represents.

Explain why a connection was recommended. Rather than returning a single opaque compatibility percentage, the system can describe an overlapping goal, complementary skill, common language or compatible availability. Those reasons are understandable and correctable. If a user rejects a match because timing is wrong, the system can update the intent without rewriting the person’s entire profile.

Exploration also matters. Recommendation should leave space for new users and different kinds of connection rather than allowing established popularity to dominate the graph.

Evaluate privacy alongside match quality

Useful measures include accepted introductions, meaningful follow-up, stale-intent errors and unexpected disclosures. A graph that improves matching by leaking private conversation history is not a good social system. Purpose-specific consent and user-visible graph controls are central to the architecture.

Real relationships are not created solely by algorithmic similarity. A high-scoring professional introduction can still fail because schedules conflict or one party no longer wants to collaborate. The model should represent explicit acceptance and actual follow-up as stronger evidence than browsing or inferred preference.

Graph changes need reversibility. A user who archives an intent or withdraws consent should be able to stop related recommendations promptly. If the system later suggests reopening that goal, it should ask rather than quietly reactivating a stale edge.

For research, compare graph-based matching with simpler baselines using goal completion, accepted introductions and privacy complaints. A more complex graph is worthwhile only when it produces more useful connections without making participants feel secretly profiled.

Relationship graphs can also include uncertainty without exposing it to every participant. The matcher may have a weak hypothesis about a shared topic, but the public explanation should draw from approved information. Separating internal evidence from shareable facts helps avoid turning private conversation into a public profile feature.