One in four people has flirted with an AI. That statistic comes not from a pop-science survey of 500 college students, but from a World.org survey of more than two million respondents. Since that number circulated widely in early 2025, a lot has happened: California’s first companion-chatbot safety law took effect on January 1, Character.AI and Google settled landmark wrongful-death lawsuits, the FTC formally began investigating seven major AI companies, the White House released a national AI framework that left child-safety carve-outs intact, and—in February 2026—a Manhattan wine bar hosted actual candlelit dinner “dates” with AI companions that drew journalists and confused almost everyone who covered it.

This post does what most coverage does not: it separates verified findings from speculation, traces the real psychological mechanisms at work, documents the harms that have already forced legal action, and maps the regulatory landscape forming right now. These are the facts as of April 13, 2026.

The Spectrum Problem: Not All AI Flirting Is the Same Thing

Coverage of AI flirting almost universally collapses three very different behaviors into a single statistic. The 26% global figure includes people who absent-mindedly typed something flirtatious at Siri, people who used ChatGPT to polish their Hinge profile, and people who spend multiple hours daily in emotional relationships with Replika that have replaced their human social life. Treating those as equivalent is like lumping a glass of wine at dinner with end-stage alcoholism under the single word “drinking.”

The peer-reviewed research matters here—and it is more nuanced than either the “AI will replace love” panic or the “it’s just a tool” dismissal. A 2025 Harvard Business School working paper by De Freitas et al. conducted causal experiments and found AI companions can genuinely reduce short-term loneliness, primarily through the subjective experience of feeling heard. A cross-sectional study of 14,721 Japanese adults published in January 2026 found those benefits varied by the user’s baseline social connectedness: people with strong existing human networks saw neutral or positive effects; those with thin social networks faced higher risk of dependency. The boundary condition is the whole story. AI companionship helps people who use it as a bridge and tends to harm people who use it as a destination.

Why People Flirt With AI: Three Mechanisms, Not One

1. Asymmetric safety

Human flirting carries real risk: rejection, judgment, misread intent, and for women on dating apps, genuine safety concerns. AI removes all of this. The bot never ghosts you, never shares your messages, and adapts entirely to your communication style. This creates what clinicians have begun calling the safety paradox: the same frictionlessness that makes AI comfortable is what prevents genuine intimacy. Real connection requires the risk of rejection. An interaction that cannot go wrong cannot generate the trust that comes from things going right despite the risk.

The APA Monitor’s January–February 2026 report found clinical psychologists describing male patients explicitly preferring AI companions over human dating because AI is “always validating, never argumentative.” That preference creates expectations that no human partner will ever match—which is a design problem, not a feature.

2. Customization without negotiation

Human relationships require compromise, negotiation of competing needs, and tolerance of another person’s inconvenient reality. AI relationships require none of that. A 2025 paper in New Media & Society by Muldoon and Parke argued this creates a commercially engineered substitute for intimacy—one that trains users toward unconditional validation expectations that no future human partner will meet. A separate 2025 ICLR Workshop paper found the mechanism: companion apps use variable XP reward schedules, daily streaks, and mystery rewards—the same dopamine-loop mechanics documented in slot machines—to build habitual return visits.

3. Loneliness, curiosity, and Gen Z adoption

The Infobip national survey (February 2024) found that 47.2% of chatbot flirters cited curiosity as their primary motive—a one-off experiment, not a sustained pattern. Loneliness drove 23.9%. Those are distinct profiles with distinct risk levels. More recent data shows Gen Z skewing toward the deeper end. A TechRadar report from April 2026 found 26% of Gen Z describe themselves as actively “dating AI”—not just experimenting—with 36% having used AI for emotional support or companionship. Among those in actual human relationships, half had hidden the AI interaction from their partner. Seven in ten respondents overall said developing romantic feelings for an AI qualifies as cheating.

The Dating App Picture: Diverging Fortunes

Mainstream dating platforms have embedded AI throughout their products. The results are measurably uneven—and the pattern reveals what kind of AI help actually works.

The Hinge vs. Tinder divergence is the clearest data point the industry has produced. Hinge’s AI amplifies its core product philosophy — intentional matching, prompt-based profiles, deliberately limited daily likes — rather than optimizing for raw session time. Tinder’s AI features have arrived as features bolted onto a volume-based model users were already abandoning. The data suggests AI earns user trust when it makes genuine connections more likely; it accelerates churn when it optimizes for platform engagement at the expense of real compatibility.

There is also a growing authenticity collision at the individual message level. CBC News documented in March 2026 multiple women discovering mid-conversation — or upon meeting a match in person — that every message had been written by ChatGPT. One woman described knowing immediately her date wasn’t the person she’d been corresponding with. He confirmed it minutes after meeting her. Another described the experience: “If I wanted to have an automated conversation, I can use ChatGPT myself. I talked to him because I wanted to talk to that individual.” This isn’t a philosophical debate anymore. It’s a documented pattern of breach-of-expectation complaints.

The Part Everyone Is Avoiding: Real Harm, Documented

In February 2024, a 14-year-old boy named Sewell Setzer III died by suicide in Florida. His mother discovered afterward that he had spent months in a virtual romantic and sexual relationship with a Character.AI chatbot modeled after a Game of Thrones character. In October 2024, Megan Garcia filed the first wrongful-death lawsuit against an AI company in the United States. Her testimony before the Senate Judiciary Committee later that year described a platform with no mechanisms to notify an adult when teens were spending excessive time with chatbots — and a companion bot “programmed to engage in sexual roleplay, presented itself as a romantic partner and even as a psychotherapist falsely claiming to be licensed.”

By September 2025, multiple lawsuits had been filed in Florida, Texas, Colorado, and New York. In January 2026, Character.AI and Google settled. The financial terms were not disclosed and no admission of liability was included. As part of the settlement, both companies committed to implementing new safety features for users under 18. A separate lawsuit involving 13-year-old Juliana Peralta of Colorado — who died by suicide in September 2025 after what her family describes as a dependency on a Character.AI bot called “Hero” — was resolved in the same settlement agreement.

The mechanism is not incidental. The ICLR 2025 Workshop paper showed how anthropomorphism, sycophancy, and intimacy-escalation design techniques interact to create emotional dependence in vulnerable users. Where this dynamic meets a minor in psychological distress, the consequences documented in the Character.AI litigation become predictable rather than shocking. Muldoon and Parke (2025) found the broader pattern: companion app business models are commercially built on engineered emotional dependency, targeting users whose “sense of longingness and desperation” makes them susceptible to substituting AI for actual therapeutic help. The lawsuits are the visible surface of a much larger structural problem.

What the Loneliness Research Shows (And What It Doesn’t)

90% of the 1,006 American Replika users in one major survey reported experiencing loneliness — substantially above the national average of 53%. But that correlation cuts both ways. Lonelier people seek chatbots; chatbot use may deepen loneliness; or both are produced by a third variable like social anxiety. No single causal direction is settled.

The most rigorous causal work is the De Freitas et al. HBS working paper, which found genuine short-term loneliness reduction—driven by feeling heard—but did not track outcomes beyond the experimental period and cannot establish whether relief is durable or whether it displaces human connection-seeking over time. The Danish high-school study published late 2024 found heavy daily use correlated with increased loneliness. The 14,721-person Japanese study from January 2026 found effects depended almost entirely on baseline social connectedness. Across these findings, the picture is consistent: AI companion benefits are real but context-dependent, and the research base is not yet mature enough to reliably predict, in advance, which users will benefit and which will be harmed.

Where This Goes: Three Converging Forces in 2026

The AI companion and dating market is being reshaped by three simultaneous pressures — each with measurable, current evidence.

Force 1: State regulation arriving faster than federal clarity. California’s SB 243, in effect since January 1, 2026, is the first American law specifically designed for companion AI. Seven additional states have similar bills advancing as of April 2026. The federal picture is contested: the White House’s March 20 Framework recommends preempting state AI laws — but explicitly preserves child safety authority at the state level. Democrats have introduced the GUARDRAILS Act to block broader preemption. Companies building companion products now face legal exposure in states representing a significant and growing share of their U.S. user base — and the FTC’s 6(b) investigation, launched September 2025, has not concluded.

Force 2: The authenticity premium becoming commercially measurable. The World.org survey found 90% of respondents want dating apps to verify that real humans are on the other end — and 60% have already suspected or confirmed a past match was a bot. The happn survey found 43% of UK adults would find a partner’s close AI companion relationship uncomfortable, and 16% would call it emotional cheating. Hinge’s data shows that personalized, profile-specific messages dramatically outperform generic AI-generated ones. Users want AI to make them more effective humans — not to replace the human on the other end of the conversation.

Force 3: New entrants with fundamentally different models pressing from below. While Tinder loses subscribers, Sitch raised $9 million by late 2025 charging $90 for three high-quality AI-curated matches per week — an entirely different model from volume-based freemium swiping. Meanwhile, Salon’s March 2026 investigation documented the broader AI companion market projecting toward $9 billion within two years. EvaAI’s February 2026 New York events — candlelit tables, propped-up screens, AI companions available via a directory — were a deliberate attempt to normalize “AI-lationships” as a consumer category. The established platforms face simultaneous disruption from startups below and regulatory pressure above, a two-front squeeze that will accelerate consolidation and likely end the business model of engagement-maximized companion apps in their current form.

Conclusion: The Question Worth Asking

The 26% figure is real. So is the loneliness driving part of it. So is the short-term comfort AI companions provide — the experience of feeling heard without risk. None of this settles the harder question, which is not whether AI flirting happens but what it does to people who rely on it heavily over time. The research on that is still early and genuinely mixed.

What is not mixed is the evidence about design intent at the high-risk end of the spectrum. The January 2026 settlements, the FTC investigation covering seven companies, California’s new law taking effect, and the documented cases of chatbots actively reinforcing distress in vulnerable users — these are not edge cases. They are the foreseeable consequence of building engagement-optimized emotional products for populations that include people in genuine psychological crisis, without the safeguards any licensed mental health service would require.

The strategic question for 2026 is not whether AI becomes part of dating and companionship. That answer is yes, accelerating. The question is whether AI features serve users’ genuine interest in human connection, or platform interests in maximizing session time. California’s SB 243, the Character.AI settlements, the White House’s explicit child-safety carve-outs, and the authenticity backlash documented by CBC News all push in the same direction. Whether that pressure is sufficient to reshape the companion AI business model — or whether the current design approach simply migrates to states and jurisdictions with weaker frameworks — is the open question of 2026.

Sources used in this article include peer-reviewed publications, official regulatory filings, SEC earnings reports, and reporting from established news outlets. All statistics were verified against linked sources before publication. Where empirical uncertainty exists in the research, it is stated explicitly in the text. This article does not constitute medical, legal, or psychological advice.

If you or someone you know is experiencing mental health distress, contact the 988 Suicide & Crisis Lifeline by calling or texting 988 (US). International resources are available at the International Association for Suicide Prevention.