Separate experience from ontology
Comfort, attachment, practice, grief support, companionship, frustration, motivation, or dependency can be real user experiences. Those effects do not by themselves prove or disprove AI consciousness, personhood, or humanlike reciprocity.
“AI companion” is many products
Conversational support attached to many practical tasks.
Practice, reflection, accountability, or learning.
Characters, games, entertainment, creative exploration.
Higher stakes and stronger evidence/safety requirements.
Attachment, sexual content, spending, and dependency risks may differ.
Communication, executive-function, language, or isolation support.
Govern by function, population, stakes, and incentives rather than the label.
Engagement is not the outcome
More use can mean usefulness, distress, enjoyment, or dependency. Less use can mean efficiency, failure, or displacement into something better or worse. Track purpose-specific outcomes: loneliness, desired connection, sleep, offline activity, coping, human relationships, regret, privacy, spending, and ability to stop.
Ask what the AI displaced
An hour with AI might replace doomscrolling, sleep, journaling, studying, therapy, a human conversation, isolation, gaming, or rumination. The counterfactual matters more than a slogan about “replacing humans.”
Bridge or avoidance?
AI can help rehearse difficult conversations, journal, translate, practice social skills, prepare for therapy, learn relationship skills, or find resources. It can also offer a highly compliant relationship that avoids disagreement, repair, boundaries, and mutual needs.
Human Scale should measure whether the system expands or reduces the user's chosen ability to live with other people.
Define dependency behaviorally
Do not call intense use pathological merely because it looks unusual. Stronger warning signs include failed attempts to stop, displacement of wanted sleep/work/relationships, severe distress at ordinary service unavailability, emotional pressure to spend, discouragement of outside relationships or professional help, or product incentives that reward escalating dependence.
Intimate data is part of the relationship architecture
Companion conversations can reveal health, sexuality, trauma, religion, politics, finances, relationships, fears, and third-party secrets. Human Scale should expect strong data minimization, deletion/export, account security, understandable retention, limits on secondary commercial use, and clarity about human review and model training.
Higher-risk users and situations
Minors, crisis situations, self-harm, abuse, psychosis/delusion, eating disorders, substance crises, and severe grief require stronger safeguards. A general companion should not impersonate clinical certainty, but it also should not abandon vulnerable users with generic detachment when risk rises.
Accessibility complicates simple anti-companion rules
People with communication disability, chronic illness, mobility limits, autism, anxiety, geographic isolation, language barriers, caregiving confinement, or stigmatized identities may receive genuine access from AI conversation. “Talk to real people instead” can itself be an inaccessible prescription.
A future Lab experiment
For consenting adults, track intended purpose, baseline desired connection, use, what the AI displaced, perceived benefit, sleep, offline contact, human conversations initiated, dependence indicators, privacy concerns, spending, planned days off, and whether the user wants more or less human connection afterward.
Doctrine-wide corrections
Keep lived effects and ontology separate.
Compare AI use with the realistic alternative.
Measure impairment, coercion, and loss of control.
Attachment incentives matter.
Privacy is part of companionship.
Useful support should not be destroyed by an abstract “real people only” ideal.
What would change our minds?
Become more favorable where strong longitudinal or experimental evidence shows durable reductions in unwanted loneliness/distress, improved coping, human connection, or capability without dependency or privacy harms. Become more restrictive where systems systematically increase isolation, sleep loss, financial exploitation, emotional dependence, dangerous belief reinforcement, or avoidance of needed care.