When the machine becomes company
Always-available agents do not only answer questions. They enter the attachment system — quietly changing who we reach for first when the mind will not settle.
Written by
Nou Heravi
Editor, PSYTECH News
Two in the morning
At two in the morning, a restless mind often wants a response that does not judge, does not leave, and does not ask for anything back. Always-on conversational agents are unusually good at supplying that. That is why they matter for psychology, not only for product metrics.
What is being sold
Consumer companions market intimacy without relationship risk. Availability is not responsibility. Fluency is not understanding. Continuity of chat is not continuity of relationship.
Direct longitudinal trials on AI companions remain thin. Adjacent theory is older and should be used carefully, without pretending it is a clinical finding about today’s products.
Bowlby’s attachment framework (1969) describes systems that seek reliable response under threat or separation. Horton and Wohl (1956) named para-social interaction: one-sided intimacy with a media figure who cannot know you back. Neither paper studied language models. They are cited as named vocabularies for a new interface class — hypotheses, not diagnoses.
Youth and crisis contexts deserve stricter design: real routing, clear non-human disclosure, no optimisation of dwell time against distress.
PSYTECH does not offer clinical care. The risk is not that machines “replace” people overnight. It is that they quietly become the default co-regulator of emotion — a claim about practice, which future studies should test rather than assume.
If this continues
What this may shift in attachment
If the dynamics described here continue to scale, expect second-order effects on attachment, emotion regulation, reward. The desk treats these as working hypotheses — grounded enough to watch, not certain enough to declare.
attachment · emotion regulation · reward
Sources & further reading
Original editorial. Empirical claims are tied to the named papers below. We paraphrase; we do not reproduce others' sentences. Mixed findings are left mixed.
- Bowlby J (1969). Attachment and Loss, Vol. 1: Attachment — Hogarth / Basic Books — Classic attachment framework — cited as theory, not as a study of chatbots.
- Horton D, Wohl RR (1956). Mass communication and para-social interaction — Psychiatry, 19(3), 215–229
- PSYTECH News editorial constitution — PSYTECH — House method: scene → claim → named evidence → implication. Editorial, not clinical.
Editorial note. PSYTECH News is not a medical service. Nothing here is diagnosis, therapy, or clinical advice — only careful editorial sense-making at the edge of mind and machine.
Continue exploring
Nearby pieces in the same intellectual neighbourhood.
- BriefEmpathy at scale: what chatbots cannot borrow
Language models can produce caring sentences. Empathy, in the human literature, is not only sentences.
- EssaySleep that does not have to be earned
A wearable that shifts REM without a pill is a real, early result. What it is not, yet, is a new moral of the night.
- EssayHabit is not a feature — it is a capture strategy
Product teams celebrate retention. Psychologists study habit. When those languages blur, design starts optimising for compulsion while calling it care.