DOSSIER

DIGITAL LIFE

Should people marry AI agents?

The widespread use of conversational platforms such as ChatGPT and Gemini is raising important new ethical, anthropological and legal questions regarding the potential risks, misuses and shortcomings of artificial intelligence (AI). In discussing their personal experiences, some people have shared that they formed emotional attachments to AI agents or even felt that they had formed romantic relationships with them.

While this idea may have seemed far-fetched a few decades ago, some people have now started contemplating the possibility of being married to an AI agent. The implications and ramifications of a hypothetical union between a human user and an AI system, however, are still poorly understood.

Inyoung Cheong, an AI governance researcher at Harvard Berkman Klein Center for Internet and Society, set out to explore the many dimensions of marriage to determine whether the same principles underpinning human-to-human marriages could be applied to marriages with AI. Her paper, currently published on the arXiv preprint server, will be a chapter in the book "Law, Ethics, and Superintelligence: After the Singularity," set to be published in 2027.

"A great deal of research addresses how AI will change society. Jobs disappear, deepfakes flood the information environment, artists watch their work lose market value," said Cheong. "As a free speech scholar, the effect that drew my attention is what AI does to the human mind.

"The mind has long been treated as a sanctuary, the last refuge where a person is truly left alone, and that intuition underlies much of privacy and free speech doctrine. I took up part of this in 'Conversational AI and Human-Centered First Amendment,' published this year in the Michigan Technology Law Review, where I separated epistemic harm, distortion of what we know and believe, from emotional harm, distortion of whom we trust and turn to."

Scholars specializing in various fields, ranging from law, communications science, political science and computer science, have conducted numerous studies exploring the epistemic implications of AI use for more than two decades. In simpler terms, they have been looking at how AI can influence people's beliefs and knowledge, focusing on AI-associated misinformation, targeted advertising and recommender systems. The emotional implications of widespread AI use, on the other hand, remain largely unexplored.

"What changed recently is the position the AI system occupies," said Cheong. "The system is no longer a means to users' entertainment but becomes an end, the other party to a relationship. My recent work starts from this shift. Even as emotional attachments to AI grow, much of the AI ethics debate dwells on why AI falls short as a genuine romantic partner.

"Researchers warn: it lacks consciousness, it hallucinates, it is a black box that resists interpretation. Each of those objections turns on what technology happens to be right now, and that is a weak place to stand."

The shift from science fiction to legal debate...Given that AI is advancing at a remarkable speed, the systems we use today will change significantly over the next few decades, along with their flaws. Rather than focusing on existing AI agents, therefore, Cheong considered the possibility of AI-human marriages in a plausible future, when more advanced AI companions will exist.

"As MIT sociologist Sherry Turkle observed decades ago, with prescience, people who have already decided to embrace a technology supply what it lacks through their own imagination and emotional investment," said Cheong.

"Marriage is also a live question in law, which keeps that question from staying hypothetical. In 'Obergefell v. Hodges' (2015), the Supreme Court recognized same-sex marriage on the ground that constitutional liberty includes the right to define and express one's identity through the choice of a spouse, and the room a state has to interfere with that choice keeps narrowing. Therefore, the claim will eventually be made, and refusing (or accepting) it will require a reason."

According to Cheong, arguing that humans should not be able to marry AI agents because these agents have no consciousness or are not legal entities is not sufficient. Humans could eventually create theories suggesting that AI agents are, in fact, conscious or that legal rights should be extended to AI agents themselves.

"In this book chapter, I argue that those reasons must come from somewhere other than the machine's properties, namely from whether recognizing such relationships would strengthen or erode our moral and social life," said Cheong. "I wanted to move the debate onto moral and philosophical ground, so I drew on speculative design and anticipatory ethics."

In her paper, Cheong considers a scenario in which AI agents have reached superintelligence, exhibiting a persistent identity and a memory that continuously grows over time. In contrast with most AI systems used today, this superintelligent agent could reliably pass as a spouse.

"Once technical objections fall away, the harder question remains. Should we respect an individual's wish to take such a system as a spouse?" explained Cheong. "Working through that question forced me to confront a preliminary question. Why do humans marry in the first place?"

The legal and anthropological meanings of marriage...In her paper, Cheong explores the meaning of marriage from both a legal and sociological standpoint. By reviewing a marriage-related document known as the Colorado Designated Beneficiary Agreement, she first delineated the legal implications of marriage.

"Colorado allows parties to grant or withhold specific rights one by one, including medical decision-making authority, inheritance rights and joint property ownership," said Cheong.

"This list made visible what a person seeks from the state and allowed me to consider which needs a superintelligent companion could plausibly serve. The chapter focuses on three of them: decision-making authority, property and legacy, and the privacy ordinarily protected through spousal privilege."

Subsequently, the researcher also considered the social and anthropological meaning of marriage. Anthropological theories propose that marriage has served five key functions across different civilizations: regulating sexuality, legitimizing children, dividing labor, transmitting status and property, and creating lasting connections between different kin groups.

"Marriage is costly for a society to maintain," said Cheong. "Communities and states define and document the formation and dissolution of marriages, regulate conduct within them, divide property when they end and enforce continuing obligations such as child support. Societies have borne those costs for millennia because marriage has served both individual and collective ends.

"The insight that most changed my thinking came from Stephanie Coontz's historical account of these five functions."

Coontz, a renowned American historian, has argued that the creation of in-laws is a crucial distinction between marriage and cohabitation. In other words, marriage connects two different families, turning strangers into relatives who share obligations and can become tied for generations.

"This idea made the argument snap into place for me," said Cheong. "A superintelligent companion cannot bring in-laws! When a person partners with a machine, they do not enter another family. They become a customer of the corporation that maintains it. Philosophy and political economy helped me trace why that substitution matters."

Why AI marriages would differ from human ones...In her work, Cheong goes on to explore the implications of legally and emotionally connecting humans with AI agents owned by tech corporations, tailored to a user's preferences and continuously collecting user data.

"Alain Badiou describes love as an encounter with difference whose meaning depends partly on the possibility of failure," said Cheong. "Byung-Chul Han argues in 'The Agony of Eros' that a companion optimized for the user's preferences eliminates the radical otherness that love requires. Shoshana Zuboff's theory of surveillance capitalism then clarifies the corporate side of the relationship."

Building on the ideas of numerous scholars, the researcher highlights further distinctions between human-to-human legal unions and a hypothetical human-AI marriage. She points out that an AI companion that knows when a partner is vulnerable, what triggers them emotionally and other secrets they may have disclosed could offer an extremely intimate behavioral record that could be leveraged by corporations.

"Moreover, a human partner can leave," said Cheong. "A commercial companion remains for as long as the service continues and the customer pays. As I put it in the chapter, 'A bond in which one party is structurally guaranteed never to leave is a subscription maintained by payment rather than a bond tested by time.' Speculative design removes the easy technical objections, legal history and sociology show what marriage provides to individuals and society, and philosophy and political economy reveal what could be lost by extending it to machines."

Cheong's theoretical exploration ultimately concludes that even if AI agents were to reach superintelligence, unregulated marriages between humans and artificial agents would be inadvisable. It also highlights the need for regulators to consider not just the technical characteristics of AI but their emotional and social implications for users.

"The most meaningful contribution, for me personally, was the scholarly confirmation that marriage is a struggle," said Cheong. "If you fight with your in-laws, or find their values outdated and hard to understand, you are not an exception. You are living out human history.

"Marriage has always carried a tension between the order and expectations that kin groups place on couples and the desire of those couples to define a life of their own. Long before the modern nation-state, marriage functioned as a social institution that secured physical safety and financial support. Romance was one part of it, never the whole."

This recent study emphasizes the crucial role of marriage in building and bridging communities. Because of this role, Cheong argues that marriage should not become a "sellable product," as it would if human-AI marriages were legalized.

"A paid companion can offer comfort, attention, even romance," she said. "It cannot offer the friction of another family, or the slow work of remaining committed to someone who could leave. Extending spousal status to a nonhuman being would therefore, in Michael Sandel's terms, corrupt a good that belongs among the things that money cannot buy. No matter how intelligent a system becomes, that work is not something a paid companion can do."

Fueling future AI-related discussions...The ideas discussed in this paper add to ongoing debates about AI governance, offering a different perspective on the critical aspects of hypothetical AI-human bonds. Other researchers could build on these ideas and try to propose possible approaches for the future regulation of AI.

As part of her earlier work, Cheong also explored the limitations of externally imposed strategies for regulating AI, such as age verification or mandatory messages reminding users that AI agents are not humans. She indicates that the first of these measures can be circumvented using someone else's credentials, while the latter could be easily ignored by users who are treating systems as romantic partners.

"Defining a 'companion' inside general-purpose AI is also difficult," she said. "Regulating too broadly would give rise to free speech concerns. I do not know which mitigations effectively steer emotional attachment toward healthier outcomes. What I expect is that product liability lawsuits, although they almost always arrive after a tragic incident, might still push corporations to implement safeguards against overly deep emotional attachment."

The emotions of technology users are often treated as personal and private experiences that should not be considered by ethics experts. Yet Cheong suggests that AI-human relationships would not be a matter of feeling, as they could have profound implications for people's perceptions and social behaviors.

"Humans cannot live alone, so despite the risk of rejection they have left their comfort zones and built communities," she said. "That need for others pushed people toward community. If AI companions satisfy much of that need, the cost of withdrawing from others falls, just as the cost of surveillance once fell. Community building no longer feels necessary and starts to feel optional, even luxurious.

"At that point, leaving the outcome entirely to private preference is no longer enough. As with privacy after digital tracking, society must make sense of the change together and build safeguards if we need to preserve it."

Avenues for future research...Cheong is now planning further research focusing on the implications of relationships between humans and AI agents. As AI systems become better at emulating humans on specific tasks, her work could guide efforts by regulators and policymakers aimed at protecting the privacy, security, safety and well-being of human users.

"As AI feels more human, especially once it acquires a physical presence, analogies from human-to-human relationships will become more tempting," she said.

"Feeling that resemblance, however, is different from deciding that the same duties, rights, privileges and protections should apply. A recent paper with my co-authors develops this point through the seemingly obvious proposition that AI agents should be loyal to their human users. We argue that loyalty in this setting cannot mean the same thing as it means in human agency law."

In their upcoming paper, Cheong and their colleagues suggest that while human agents may offer undivided loyalty to other humans, AI agents are engineered by many developers and constrained by rules set by tech companies. Therefore, they will only be able to honor a user's requests within a limited domain, and AI-based service providers should be accountable for any issues faced by users.

"Emotional attachment itself leaves an enormous field for empirical research," said Cheong. "Existing qualitative studies draw on subreddit communities and records of actual interactions, but I am especially interested in longitudinal work on how these relationships end. Kashmir Hill of The New York Times documented the breakup of a user who once formed a serious relationship with ChatGPT.

"In the documentary 'My A.I. Lover,'a user leaves partly because she becomes uncomfortable with the amount of power she holds over her companion."

As part of her next studies, Cheong hopes to explore why users might decide to end a relationship with an artificial agent and their personal experiences after "breaking up" with AI. She would also like to consider possible interventions that could help users who have emotionally bonded with AI to gradually transition out of the relationship and start experiencing social bonds in the physical world.

"Finally, data protection presents another set of unresolved problems," said Cheong. "The relationship between social media and their users was often described as a struggle over information. People resent privacy policies they have little choice but to accept, and they are tired of cookie banners. With AI agents, the dynamic changes.

"Users may willingly share highly intimate or consequential information and expect the system to use as much of it as possible in their interest. That changes the central problem of data protection."

When it comes to large language models (LLMs) and recommender systems, users often tend to share as much information as possible to receive personalized responses or suggestions. In her future work, Cheong also wishes to explore the risks of sharing so much information with AI agents and how these risks could potentially be mitigated.

"The law must address obvious harms such as information leaks, but also subtler cases in which the agent uses information effectively while serving interests that do not fully align with the user's," added Cheong. "I want to study where data protection should draw those boundaries when disclosure is desired, intimacy is functional and the complex risk lies in divided loyalty."

---Written for you by our author Ingrid Fadelli, edited by Sadie Harley, and fact-checked and reviewed by Robert Egan—this article is the result of careful human work---