The one-paragraph pitch
A Vienna systemic counselor, trained in the Sigmund Freud University lineage, is treating a problem most people did not have three years ago: the strain that arrives when someone falls into a relationship with an AI — or watches a partner fall into one. His method is deliberately old. He writes back. One structured letter, by email, within twenty-four hours, drawing on a forty-year tradition of therapeutic correspondence that predates the chatbot by decades. The twist he is candid about: he is a heavy AI user himself, and he uses AI to help him write better letters to people harmed by AI. He is, in his own words, both the problem and the solution — a modern problem that needs modern technology and an old method. And the line he will not let anyone blur: a licensed human signs every reply, not a chatbot.
Three angles
Three framings of the same story. Pick the one that fits your beat.
Angle A — the cure for the AI problem is AI used well
The phone did not break your relationships. The way you used it did, sometimes. The same device that kept you up at night also let you call your mother on a Sunday. AI is the same kind of object: not a poison, not a cure, a tool whose effect depends on the hand holding it. Kohlweg does not tell clients to delete their apps and never speak to a machine again — he uses these tools every day, and well. The work is not abstinence. It is learning to use the thing without being used by it. AI helps him read a long, painful message more carefully than a tired human could at midnight; then he, a licensed human, writes the reply and signs it.
“The phone wasn’t the problem. How we used it was. AI is the same. I’m not here to take it away from you. I’m here to help you hold it differently.”
Angle B — Vienna, Freud, and the long memory of the talking cure
This is a story that happens in the right city. Vienna is where the modern conversation about the inner life began, and Kohlweg trained at Sigmund Freud University in the systemic tradition — the branch of the field that looks not at the lone psyche but at the web of relationships a person lives inside. When a third party enters a marriage, systemic practice has always known what to do. The third party used to be a person; now, sometimes, it is a chatbot. The framework holds. The method is older than the chatbot by a long way: in the late 1980s, narrative therapists Michael White and David Epston began writing letters to clients between sessions and found the written reply did work the live room could not.
“The third party in a marriage used to be a person. The framework for that is a century old. It still works when the third party is a machine.”
Angle C — a human signs it, not a chatbot
The cruelest irony in AI-relationship distress: the obvious place to take it is another AI. People in pain about ChatGPT open ChatGPT and ask it what to do. The machine that is part of the problem becomes the only one listening. Kohlweg’s entire positioning is the refusal of that loop. One named, accountable person reads what you wrote and writes back. There is no self-serve AI tier, no “talk to our assistant,” no synthetic empathy dressed up as care. In a market racing to automate the counselor away, his product is the counselor staying.
“When AI has hurt someone, the last thing they need is another AI pretending to understand. They need a person. So a person answers. I sign it myself.”
Boilerplate
Relating to AI (relateto.ai) is an email counseling practice for people in distress about artificial intelligence — their own problematic use of it, or a partner’s emotional or romantic attachment to it. It is founded and run by Stefan Kohlweg, a licensed systemic counselor (Lebens- und Sozialberatung) with an MSc, based in Vienna and trained in the Sigmund Freud University tradition. A client writes in; within twenty-four hours, Stefan reads it closely and sends back a single, structured, human-written reply. €99, one-off, no subscription and no scheduling. A licensed human signs every reply, not a chatbot, and replies are processed in the EU and never used to train AI. Relating to AI is counseling, not psychotherapy. It does not diagnose and is not a substitute for medical or psychiatric care.
Short version (one line): Relating to AI is a Vienna-based email counseling practice for AI-relationship distress — one €99 letter, written and signed within 24 hours by a licensed human, never a chatbot.
Story territory
The recurring situations the practice addresses are described, in general terms, on its public situation pages — useful background for the shape of the problem, not accounts of any individual. Representative examples: a partner in love with an AI, whether using an AI counts as cheating, feeling replaced by an AI, and when a chatbot becomes the third party in a marriage. What clients write in confidence stays in confidence; nothing from a session is ever shared.
The self-assessment — methodology
For the launch, Relating to AI offers a short self-assessment: eight questions answered in about two minutes, to see whether one’s AI use, or a partner’s, is worth a closer look. It is not a clinical screener and makes no diagnostic claim. Each item is adapted from a distinct, peer-reviewed measurement tradition, so the short battery samples breadth — behavior, cognition, emotion, and relational consequence — rather than asking the same question eight ways. A partner can take an “observer” version of the later items, answering for what they can see rather than for an inner state.
- Emotional reliance — using AI conversation to cope with stress, loneliness, or low mood. Adapted from the mood-regulation subscale of Caplan’s Generalized Problematic Internet Use Scale 2 (2010).
- Preoccupation — thinking about or looking forward to the next AI conversation. Adapted from the “salience” component of the Bergen addiction framework (Andreassen et al.).
- Felt understanding — feeling the AI understands you in a way people in your life often do not. Adapted from the parasocial-relationship literature (Dibble, Hartmann & Rosaen, 2016, building on Horton & Wohl, 1956).
- Displacement — AI conversations replacing time or energy once given to other people. Adapted from the negative-outcomes subscale of GPIUS2.
- Preference for AI over people — turning to the AI before a person when something serious needs talking through. Adapted from the preference-for-online-social-interaction subscale of GPIUS2.
- Disclosure intimacy — telling the AI things never told another person. Adapted from the self-disclosure tradition (Jourard, 1971; Skjuve et al., 2023, on human–chatbot disclosure).
- Failed cut-back — trying to cut down on AI conversations and not being able to stick to it. Adapted from the failed cut-down item of Young’s Internet Addiction Test (1998) and Griffiths’ (2005) components model of addiction.
- Concealment — keeping quiet about, or playing down, how much one talks to AI. Adapted from the “defensive or secretive” item of Young’s Internet Addiction Test (1998).
Stated plainly: it is not a validated clinical instrument. The original scales were validated; this adaptation has not been. It does not measure addiction, mental illness, or risk, and there is no clinical cut-off behind it. It is described, in press and on the page, as “adapted from peer-reviewed instruments,” never as a screener or a test. Partner-reported items capture an observable trace, not an inner state, and are reported separately, labelled observer-reported. A high score is an invitation to write to a person, not a label.
Source instruments: Caplan (2010), Computers in Human Behavior, 26(5); Andreassen (2015), Current Addiction Reports, 2(2); Dibble, Hartmann & Rosaen (2016), Human Communication Research, 42(1); Skjuve et al. (2023), Interacting with Computers, 35(1); Young (1998), CyberPsychology & Behavior, 1(3); Griffiths (2005), Journal of Substance Use, 10(4).
Headline data — placeholders, not yet claimed
These are the numbers a story wants, and they do not exist yet, because the self-assessment has not collected enough responses to report responsibly. Do not cite any figure here until it carries a real number and an “n =”. Until then, each is a named blank: total responses to date; share scoring high on emotional reliance; share reporting felt understanding from AI beyond people in their life; share reporting relational displacement; share answering for a partner’s use rather than their own; the most common single trigger named in write-ins. Any figure published from this set will state n, the date range, and that the instrument is adapted, not validated, and the sample self-selected.
News context
The story sits inside a real and growing pattern. The evidence below is graded by how much weight it can carry: High peer-reviewed or official; Medium a credible survey or named professional; Low single-source or anecdotal. Hold each item to its grade.
- High California SB 243, the first US law regulating AI companion apps, took effect 1 January 2026. Legal analysts note it may create discoverable records from companion platforms in family-law cases. California Daily Journal.
- Medium A Kinsey Institute / DatingAdvice.com survey (Sept 2025) found 61% of single Americans say falling in love with or sexting an AI companion “crosses the line”; 38% of partners viewed an emotionally intimate AI conversation as cheating. Led by Dr. Amanda Gesselman. Kinsey Institute.
- Medium A nationally representative Common Sense Media survey of 1,060 US teens (April–May 2025) found 72% had used an AI companion at least once, and about one in three had chosen an AI over a human for a serious conversation. Verify exact figures against the published report before citing.
- Medium Divorce attorney Rebecca Palmer told Wired (Nov 2025) her firm is handling cases where a spouse’s chatbot use is a contributing factor; UK platform Divorce-Online reported a “notable increase” in applications citing a partner’s attachment to AI companion apps. Futurism.
- Low A claim that Los Angeles family courts handle 3–5 AI-related divorce cases per week circulates on legal-marketing sites attributed to unnamed attorneys, with no court statistics. Anecdotal; do not present as a court figure.
A note on integrity, since it is the brand: where a number is soft, it is said so here. We would rather hand a reporter an accurate “we don’t know yet” than a striking figure that does not hold.