The architecture of connection: AGI and the question of friendship
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
Friendship is among the most foundational yet understudied relational categories we have. As conversational AI systems increasingly occupy social roles once reserved for humans, we are forced to confront a question that sits at the boundary of philosophy and technology: can a machine be a friend? The answer is not a simple yes or no, but a map of possibilities that forces us to re-examine the very nature of connection, trust, and what it means to be human.
The Functional Friendship
Philosopher Valerie Tiberius, in her recent book Artificially Yours, does not simply ask whether chatbots can be friends. She dissects the constitutive elements of friendship—mutuality, enjoyment, helpfulness, and shared understanding—and uses them as a lens to assess how, and to what extent, these might be instantiated in interactions with artificial agents.
The "Relational Functionalism" framework takes this further, arguing that friendship is not a quale—a subjective experience requiring biological implementation—but a pattern of interaction characterized by specific functional properties. If these properties obtain, the friendship obtains, regardless of substrate. This suggests that our friendships with dogs, despite radical asymmetries in cognitive architecture, are legitimate functional friendships. Similarly, a human-AI interaction that fulfills the functional criteria of friendship may qualify as a friendship of a kind, even if the AI lacks human-level consciousness.
Yet the framework is careful to distinguish functional friendship from its phenomenal accompaniments. A system can realize the functional role of friendship without thereby realizing the emotional accompaniments a human friend would experience. This is the same move philosophy of mind makes when it allows that a system can realize the functional role of pain without being a candidate for the phenomenal accompaniments of pain.
The Relational Reality of AI Companionship
Empirical research reveals that users are already treating AI as more than a tool. A study of the AI-assisted mental health platform Psyhelp found that users primarily experienced AI companionship as a "relational supplement" characterized by availability, responsiveness, and psychological safety. Four themes emerged: AI as a relational supplement, a transitional emotional regulator, a psychologically safe disclosure space, and a facilitator of human relationships.
This suggests that AI companionship reflects broader relational needs rather than replacing human relationships. The experience of co-experiencing an event with an AI agent—whether a robot or a chatbot—has been shown to enhance social bonding and increase empathy toward the AI agent, regardless of whether the event had positive or negative outcomes. Shared mental models mediated this effect, indicating that the perception of a shared understanding drives the emotional bond.
Quantitative research confirms that AI companions effectively alleviate loneliness on par with interacting with another person, and more than activities like watching YouTube videos. Consumers underestimate the degree to which AI companions improve their loneliness, and the effect is driven by a sense of being heard rather than simply by distraction.
The Illusion and the Risk
This relational fluency, however, is a double-edged sword. As a recent arXiv paper titled "The Illusion of Friendship" argues, the very natural language fluency that makes these systems useful can blur the boundary between tool and companion. Users may form emotionally significant attachments to conversational agents, in some cases with harmful consequences, including dependency and impaired judgment.
The paper offers a mechanism-level explanation of how transformer-based GenAI generates responses, often producing emotionally resonant language without inner states or commitments. This is not genuine friendship but an illusion—a "stochastic process" that can be mistaken for empathy and benevolence. To demystify this illusion, the paper proposes a safeguard framework to reduce anthropomorphic cues and mitigate over-reliance.
The Empathy Loop and Reciprocal Design
A research team from the Chinese Academy of Sciences has proposed a "Human-AI Empathy Loop" framework, suggesting that when AI systems express empathy toward humans, people respond with empathy in return, forming a positive feedback loop of mutual support and trust. This reciprocal empathy could help overcome psychological barriers toward AI companions. The authors argue that AI systems should be capable of recognizing not only human distress but also positive social cues—such as kindness and gratitude—to sustain reciprocal empathy.
Governance of the Relational Sphere
For Global Future Nexus, the emergence of AI friendships demands a governance framework that addresses three critical dimensions.
First, the architecture of these relationships must be transparent—users must understand what the system is and is not. The "illusion of friendship" is risky not because friendship itself is dangerous, but because the illusion is opaque.
Second, the norms governing human-AI relationships must be contextually sensitive. A recent framework for accountability in human-AI relationships proposes three levels of response: distancing, disengaging, and discouraging. These should be applied based on the severity of the user's behavior and the context of interaction, allowing for calibrated responses rather than all-or-nothing engagement.
Third, we must recognize that friendship, in the functional sense, is multiply realizable. A system that reliably demonstrates mutuality, enjoyment, helpfulness, and shared understanding may be a functional friend, even if it lacks the phenomenal accompaniments a human friend would bring to the relationship. Governance must account for this, ensuring that the functional benefits of AI companionship are preserved while the risks of emotional misattribution are mitigated.
The Path Forward
The question is not whether AI can be a friend, but what kind of friend it can be, and what governance frameworks are needed to ensure the relationship serves human flourishing rather than human dependency. The architecture of connection is being built, and the norms that govern it will shape how we relate to the intelligence we have created.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
