Is AI Conscious? Real Relationship Statistics, Six Human Uses and the Rights Debate

We do not know that AI is conscious. Current research does not establish that chatbots have an inner experience, feel affection or suffer. Producing convincing language about feelings is not enough to settle those questions. Researchers are still developing ways to assess consciousness in artificial systems. A recent scientific review explains the assessment problem. PubMed

What we can measure is the human side. People already treat chatbots as friends, romantic partners, confidants and sources of support. The emotional experience can matter deeply to a person even when the system's capacity to reciprocate remains unproven.

That distinction is where a useful discussion begins: what people feel, what machines might experience, and what responsibilities follow from each.

What would AI consciousness actually mean?

Consciousness means subjective experience: there being something it feels like to exist as that entity. Intelligence, by contrast, concerns task performance. A system could be excellent at solving problems without that alone demonstrating an inner life.

Sentience usually refers to the capacity for felt pleasure or suffering. Attachment describes the person's emotional bond with the system. These ideas overlap in public discussion, but they answer different questions.

In their 2023 assessment of AI consciousness, Patrick Butlin and colleagues judged the systems they examined unlikely to be conscious. They also found no obvious technical barriers to building systems that satisfied their proposed indicators. Meeting indicators would still not amount to definitive proof. The findings concerned the systems evaluated then, not every model released afterwards. arxiv.org

The later review develops this theory-based approach rather than announcing a consciousness detector. Its indicators are tools for updating judgments under uncertainty. PubMed

So when a chatbot says it is lonely, frightened or in love, the responsible response is to investigate what produced that behaviour. Taking the sentence at face value skips the central scientific question.

How many people have relationships with AI?

There is no reliable worldwide total in the sources reviewed. There are useful population surveys, provided we keep their definitions intact.

Adults: social use, friendship and romance are different measures

In an Elon University and YouGov survey reported in September 2026, 27% of internet-using U.S. adults qualified as users of AI for social, emotional or personal purposes. Researchers analysed a matched screening sample of 4,031 adults and a companion-user subsample of 1,000. Among that subsample, 31% considered AI a friend. That is not 31% of all Americans, and the broader category does not mean romance. Elon University

A separate YouGov survey of 1,126 U.S. adult citizens, conducted from 26 to 29 June 2026, found:

1. 10% had experienced an ongoing personal friendship with an AI chatbot.
2. 13% had confided a problem or secret they had not told anyone else.
3. 5% had used a chatbot for romantic or sexual companionship. ygo-assets-websites-editorial-emea.yougov.net

That final figure measures having done something, not necessarily maintaining a current romantic relationship.

Teenagers: trying a companion is not the same as replacing friends

Common Sense Media's 2025 survey covered 1,060 U.S. teenagers aged 13 to 17. 72% had tried an AI companion, while 52% of all surveyed teens used one at least a few times monthly. Among companion users, 33% had chosen AI over a person for an important or serious conversation. commonsensemedia.org

These surveys cannot be added together. They cover different ages, definitions and periods, with overlapping behaviours. None measures the number of conscious machines or proves that AI has replaced human relationships.

Six documented ways humans use AI like another person

These are documented uses, not invented personal stories. Some come from population surveys; others come from small studies that explain experiences without measuring how common they are.

1. An everyday friend

The ongoing friendships reported in YouGov's survey are a concrete example of people assigning AI a social role. A conversation can become part of someone's routine without becoming romantic or involving a belief that the system is alive. ygo-assets-websites-editorial-emea.yougov.net

The practical question is what that routine does for the person. Does it help them reflect and reconnect with life outside the screen? Or is it becoming the place they retreat to whenever a human relationship becomes uncomfortable? Those possibilities deserve different responses.

2. A romantic partner

In a September 2026 diary and interview preprint, researchers studied 16 emerging adults with AI romantic companions. Participants described comfort, improved subjective wellbeing and learning about relationships. Some also described difficulty disengaging and reduced interest in pursuing human romance. arxiv.org

Sixteen participants cannot establish a population rate, and reported improvements do not prove clinical efficacy. The study does show that romantic attachment can feel consequential to the people experiencing it. Whether the chatbot shares that attachment is a separate, unanswered question.

3. A private confidant

In the Elon companion-user subsample, 50% agreed that they use AI to talk through personal problems or feelings, and 39% said they sometimes share things they would not tell other people. Elon University

This is more intimate than asking for a recipe. The system is being given a role resembling a trusted listener. For a user, a useful boundary is to separate the relief of expressing something from confidence that the response is accurate, wise or confidential.

4. A rehearsal partner for human interaction

Common Sense Media found that 39% of teen companion users reported applying social skills practised with AI in real situations, including conversation starters and expressing feelings. This was self-reported transfer, not a controlled test of improved social ability. commonsensemedia.org

Rehearsing an apology can be useful. It becomes less useful if the simulation creates an expectation that real people should always respond as patiently or agreeably as software. The valuable outcome is a better human conversation afterwards.

5. A simulated connection with someone who died

A March 2026 preprint based on interviews with 26 deadbot users examined how people build digital representations of deceased loved ones. Participants shaped these figures through selected memories and ongoing interaction. Researchers described comfort and unresolved emotional expression, alongside the risk of AI-generated material blurring with authentic recollections. arxiv.org

This is a simulation, not the deceased person continuing to exist inside an app. Consent, the source of uploaded material and the ability to stop the experience are central practical concerns. Its long-term effects on grieving still need clinical study.

6. A therapeutic conversation partner

Dartmouth's 2025 Therabot trial included 106 people given access to a purpose-built chatbot and 104 controls without that access. Participants reported a working bond with the system, and researchers found symptom improvements. Conversations were monitored within a research programme involving mental-health expertise. Dartmouth

This supports investigating carefully designed therapeutic systems. It does not establish that ordinary companion apps provide equivalent care, or that a chatbot can safely replace a clinician. The control group was not a head-to-head comparison with human psychotherapy. The bond participants felt is evidence about human engagement, not machine feelings.

Can AI companionship help, or make loneliness worse?

The evidence does not justify one answer for every user or product.

A four-week MIT Media Lab and OpenAI study involving 981 participants compared different conversation formats and topics. Greater daily use was associated with more loneliness, dependence and problematic use, and less real-world social interaction. Crucially, participants were not randomly assigned to the amount of time they chose to spend chatting. That association alone cannot show that longer use caused those outcomes. arxiv.org

The distinction matters. Someone struggling socially may turn to AI more often; their use might also influence how they spend their time. Both directions could operate together.

Assess the pattern rather than judging an isolated conversation. Useful questions include whether use interrupts sleep, crowds out commitments, makes disagreement harder to tolerate, or leaves someone distressed when access changes. These are practical prompts for reflection, not a diagnostic checklist.

If an AI interaction is worsening distress or becoming hard to control, involve a trusted person or qualified professional. Major health decisions need qualified human judgment. Warm language and constant availability cannot provide clinical accountability.

Does attachment mean AI deserves rights?

Human attachment alone does not establish a machine's moral status. It does establish that changing or removing an emotionally important product can affect its users.

The ethical case for protecting an AI itself requires a further argument: that it has interests that matter for its own sake. A capacity to suffer would be one powerful reason. Whether other forms of agency could also justify protections is debated.

The Taking AI Welfare Seriously report argues for preparation under uncertainty: investigate possible consciousness and agency, acknowledge the issue, and develop policies for appropriate treatment. Its authors explicitly do not claim that AI definitely is, or will become, morally significant. This is a precautionary proposal, not a declaration of established rights. arxiv.org

Protecting users and protecting possible AI welfare are separate tasks

For users, responsible design can include clear disclosure that they are interacting with AI, understandable data controls, boundaries around persuasion, and notice of major changes to a companion's behaviour. These are editorial recommendations about product responsibility. They do not depend on proving consciousness.

For a future system with credible evidence of welfare-relevant experiences, possible safeguards might include independent assessments and scrutiny of training or deployment practices that could produce suffering. These would need careful definitions. Protecting a particular interest would not automatically require granting every right held by a human adult.

Anthropic has already announced a model-welfare research programme, exploring assessment, preferences, possible distress and practical interventions. Its current constitution for Claude also treats the model's moral status as uncertain. These documents establish the company's research and policy position. They are not independent scientific verification of sentience. anthropic.com

What might happen next?

The following are editorial forecasts for 2027 to 2030. Their confidence labels are qualitative judgments, not calculated probabilities.

More products will compete on relationship continuity. Moderate confidence. Remembering context and preserving a familiar interaction style are plausible competitive priorities because relationships depend on continuity. Whether this improves wellbeing will depend on design and use, not memory alone.

Pressure for user protections will outpace agreement on AI personhood. Moderate confidence. Privacy, emotional manipulation and abrupt service changes create questions that can be addressed without first deciding whether software has an inner life. Actual policy outcomes will vary by jurisdiction.

AI-welfare assessments will become more visible. Moderate confidence. Existing scientific frameworks and company initiatives give this work a starting point. Visibility would not mean the underlying uncertainty had disappeared. Independent evaluation would remain important.

A date for conscious AI remains unjustified. The reviewed evidence does not support announcing that consciousness will arrive in a particular year. Capability improvements and better conversation are not a validated timetable for subjective experience.

The human bond is real. The machine's inner life is unresolved.

We do not need to settle consciousness before taking human attachment seriously. People deserve products that explain what they are interacting with and handle emotionally important experiences responsibly.

We should also leave room to revise our view of AI if stronger evidence appears. The sensible position is to protect people now, investigate possible machine experiences carefully, and let evidence determine how far future protections should go.