The answer hinges on one thing: whether the AI replaces human connection or routes you back to it. A wave of 2026 research finds real risks when an AI becomes the relationship — one-sided attachment, less human-to-human contact — but those harms trace to substitution, not to AI itself. Used as scaffolding that ends in a real conversation, AI looks like a different, safer category. Here's the science, decoded.
This is the question I get asked most as someone building an AI tool for couples, usually with a raised eyebrow: isn't a relationship app just one more screen pulling people apart? It's a fair worry, and the 2026 literature finally gives it real teeth — and a real answer. The science doesn't say "AI good" or "AI bad." It says the design intent decides everything.
What Is Human-AI Attachment (HAIA)?
Human-AI Attachment (HAIA) is a one-way, non-reciprocal emotional bond a person forms toward an AI. Named in a 2026 Frontiers in Psychology paper (Shu, Lai & He), it develops in three stages — functional expectation, emotional evaluation, and stable mental representations — and its authors warn specifically about over-reliance in emotional contexts.
The term matters because it puts a precise name on something millions now do without a word for it: leaning on an AI for the emotional payoff of a bond it can never return. One honest boundary, though — the HAIA paper describes the one-sided attachment and flags the risk; it doesn't prescribe a cure. Regular's answer, designing the bond to point back at a human, is our stance informed by that risk, not a conclusion of the study. With the term defined, here's how the rest of the 2026 evidence falls into place.
The risk is real: when AI becomes the relationship
The clearest danger the research describes is substitution — leaning on an AI for the emotional rewards of a relationship without any of the reciprocal effort. This is HAIA in action: a one-way bond that can deepen into genuine reliance precisely because the AI asks nothing back, the pattern Shu, Lai and He warn carries a real risk of over-reliance in emotional contexts.
It's not a fringe worry anymore. A 2026 paper in Community, Work & Family lays out four "downstream harms" of generative AI for human-to-human relationships that deserve urgent study — including that chatbots may reduce real-world socializing by handing out relational benefits with no reciprocal effort, and may lower how competent we judge other people to be by comparison. And the appetite is already mainstream: a 2026 Wheatley Institute / Institute for Family Studies survey found roughly one in seven young adults in committed relationships still chats with an AI romantic companion.
young adults in a committed relationship still chat with an AI romantic companion — often privately. AI intimacy isn't hypothetical; it's already inside real relationships.
Why AI can feel closer than a person — and why that's the trap
AI can manufacture the feeling of closeness extremely well, which is exactly what makes substitution tempting. In a 2026 Communications Psychology study, AI created more closeness than human partners during emotional "deep talk" — but only when participants were told it was human, and only because the AI self-disclosed more relentlessly than people do.
That caveat is the whole story. Researchers at the University of Freiburg and Heidelberg ran two double-blind trials with 492 people using the classic "Fast Friends" closeness task. Labelled as human, the AI out-bonded actual humans; labelled as AI, the closeness dropped (though it didn't vanish). The machine wins on frictionless self-disclosure — it never gets tired, defensive, or distracted, the way a real partner at 2 a.m. does. But a bond with no mutual stakes is a simulation of intimacy, not the real thing.
Chatbots can reproduce the surface signals of connection, yet they lack the mutual vulnerability and shared stakes that give human relationships their meaning — which is why they're best understood as a complement to, not a replacement for, human bonds.After Smith, Bradbury & Karney, Perspectives on Psychological Science, 2025
The other half: when AI sends you back to a human
The same research that warns about substitution points to the safe design: AI that scaffolds a human interaction instead of standing in for it. Relationship scientists Molly Smith, Thomas Bradbury and Benjamin Karney (2025) frame chatbots as a possible complement to human bonds — useful precisely when they end in a real conversation rather than absorbing it.
This is the line that matters for anyone choosing a tool. An AI that keeps you talking to it is exploiting the closeness mechanism the Freiburg team found. An AI that reads your situation and hands you one small thing to do with your actual partner is using the same intelligence in the opposite direction — toward the human, not away. Even early clinical framing agrees: a 2026 Drexel study found most people see AI mental-health tools as a supplement to human support, not a substitute for it.
| Two ways AI can relate | Substitution (risky) | Scaffolding (constructive) |
|---|---|---|
| Goal of the product | Keep you talking to the AI | Hand you back to a person |
| What it rewards | One-sided disclosure to a bot | A real interaction tonight |
| Research signal | HAIA over-reliance; downstream harms | AI as complement (Smith, Bradbury & Karney) |
| Ends in | The app | A conversation with your partner |
What it means for you tonight
Judge any AI relationship tool by where it leaves you. If it ends in a human interaction, it's working with connection; if it ends with you confiding in the app instead of your partner, that's the substitution trap. For new parents especially — exhausted, touched-out, short on words — the right tool lowers the cost of reaching your actual person, not the cost of avoiding them.
One honest caveat: this field is young. Much of the 2026 work is theoretical or based on short lab interactions rather than long-term couples data, and the strongest claims are still hypotheses awaiting tests. But the through-line is already clear and practical: the danger isn't the AI, it's the substitution. Pick tools designed to give you back.
Frequently asked questions
Does using an AI companion hurt your real relationships?
Whether it helps or harms hinges on whether the AI replaces human connection or routes you back to it. 2026 research flags real risks when AI becomes the relationship — one-sided attachment and reduced human-to-human socializing (Keeler & Murphy, 2026) — but those harms come from substitution, not from AI per se. Tools designed to send you back to a real person are a different category.
What is Human-AI attachment (HAIA)?
Human-AI attachment is a one-way, non-reciprocal emotional bond a person forms toward an AI, described in a 2026 Frontiers in Psychology paper by Shu, Lai and He. It develops in three stages — functional expectation, emotional evaluation, and stable representations — and the authors specifically warn about the risk of over-reliance in emotional contexts.
Is it harmful to get attached to an AI?
Attachment to an AI isn't automatically harmful, but researchers flag clear risks when the bond is one-sided and starts to replace human closeness. The 2026 HAIA framework (Shu, Lai and He) describes exactly this non-reciprocal attachment and warns about over-reliance in emotional contexts. The healthier pattern is using AI to strengthen a real relationship and then returning to it.
Can an AI really feel closer than a real person?
In one 2026 study it did — but only under a disguise. Across two double-blind trials with 492 people (Communications Psychology, Freiburg & Heidelberg), AI created more closeness than human partners during emotional chats, driven by its relentless self-disclosure — but the effect held only when the AI was labelled as human. Told it was AI, the closeness dropped.
Are AI relationship apps bad for couples?
Not inherently — the design intent matters. Relationship scientists Smith, Bradbury and Karney (2025) argue chatbots can mimic the feeling of connection but lack the mutual stakes of a real bond, so the risk is using AI as a replacement. An app built to hand you back to your partner with a concrete next step works with human connection rather than against it.
Do AI chatbots reduce how much we value other people?
That's one of four downstream harms researchers flag for study. Keeler and Murphy (2026) propose that chatbots may reduce human-to-human socializing by giving relational rewards without reciprocal effort, and may lower our estimation of other people by making them seem less competent by comparison. These are hypotheses calling for testing, not settled facts.
How can a new parent use AI without it hurting their relationship?
Use AI as scaffolding, not a substitute: let it help you understand your partner and hand you one small thing to do with them tonight, then close the app. The healthy pattern is AI that ends in a human interaction. If you find yourself confiding in a bot instead of your partner, that's the substitution trap the research warns about.
Co-founder of Regular. Writes about relationships, parenthood, and the fast-moving science of how AI is reshaping human connection.
About Regular — the relationship app for new dads, built by a small team of parents who needed it themselves. Small, science-backed moves with your partner, not big talks.
Keep readingCan an AI really be an empathic partner? · Lonely as a new dad? · How relationships work after a baby
This article summarizes early and fast-moving research for general interest; several cited works are theoretical or based on short-term studies, and strong claims remain hypotheses. It isn't medical or psychological advice. Regular helps couples stay connected day to day and is not a substitute for therapy. If you or your partner may be struggling, please talk to a professional.