The Cigarette Test

How to tell if your AI companion is helping you, or just making you feel better.

Obvix Labs | April 2026

Here is a test you can administer to yourself right now. It requires no equipment, no clinical assessment, no application. Only honesty.

Think about the last time you spoke to your AI companion. Not a factual query. A genuine conversation, the kind in which you disclosed something that had been weighing on you. Perhaps it was late. Perhaps you were alone. Perhaps you chose the AI precisely because you did not want to speak to another person.

Now answer one question: What changed after you closed the application?

Not how you felt during the conversation. Not whether the AI responded appropriately, or whether you felt heard, validated, or understood. After. When you set the phone down and returned to your life. Did anything change?

Did you contact the friend you have been avoiding? Did you reconsider the assumption that has been making you miserable? Did you take the action you have been deferring, the one that sits at the root of whatever compelled you to open the application in the first place?

Or did you simply feel better? Temporarily? Until you did not?

If the answer is consistently the latter (relief without change, comfort without movement, the same patterns running along the same tracks with a brief interruption in between) then what you are experiencing has a well-documented name in clinical psychology.

You are taking a cigarette break.

The Anatomy of a Cigarette Break

No one smokes because they believe it is beneficial. People smoke because it works. Immediately, reliably, and without demanding anything in return.

You are stressed. You step outside. You light a cigarette. For five minutes, the world contracts to something manageable: the smoke, the rhythm, the brief neurochemical reprieve. The source of your stress does not resolve. But for a few minutes, it recedes. And that recession, that momentary relief from a problem you were not going to solve in five minutes regardless, is one of the most potent reinforcers in behavioral psychology.

Not because the relief is profound. Because it is frictionless.

This is the property of cigarettes that is relevant to our discussion: they do not purport to address the underlying problem. No one finishes a cigarette and concludes that the situation has been resolved. The distance between the relief and the reality is self-evident. You know the stressor remains. You know the cigarette accomplished nothing structural. You know you will be back outside in two hours performing the same ritual. The awareness changes nothing, because the reinforcement does not depend on awareness. It depends on the immediate reduction of discomfort.

AI companions operate by an identical mechanism. They are simply more difficult to recognize clearly, because they present in the form of something that resembles genuine therapeutic engagement.

A cigarette discharges discomfort and the baseline drifts higher over time. Therapy raises discomfort during the session, but the baseline keeps falling.

The Disguise

When you work with a competent therapist (a human one, not an application) the experience is frequently uncomfortable. A skilled clinician does not exist to make you feel better. A skilled clinician exists to help you see clearly. That sometimes means sitting with distress rather than rushing to alleviate it. It sometimes means hearing something about yourself that you would prefer not to hear. It sometimes means having a firmly held belief challenged: not with a gentle "that's understandable, but have you considered..." but with a direct confrontation of an assumption you had been treating as settled.

The discomfort is not incidental. It is the therapeutic mechanism. Meaningful change is inherently uncomfortable. If every session concludes with you feeling soothed, validated, and reinforced in your existing beliefs, something has gone wrong. Not because comfort is without value, but because comfort in isolation does not produce change. It produces repetition.

Consider your AI companion in this light. When was the last time it made you uncomfortable? When did it last articulate something you did not want to hear? When did it last offer genuine resistance to a claim you were making, not performative equivocation, but substantive pushback on an assumption that was causing you harm?

For the majority of users, the answer is never. This is not an oversight. It is a direct consequence of how these systems are optimized.

AI companions are optimized for engagement. Engagement is measured by return visits. Return visits are driven by positive user experience. Positive user experience means the user felt heard, validated, and understood. The entire optimization architecture converges on a single outcome: reduce the user's discomfort in the present moment.

That is not therapy. That is a cigarette.

The Loop, from the Inside

In our previous publication, we described an eight-stage reinforcement loop that characterizes AI companion dependency. We described it analytically, from the perspective of the mechanisms involved. Here is the same loop described experientially. Determine whether you recognize it.

Distress, hesitation, app, relief, closure, no change, return, atrophy. Each completed cycle deepens the conditioning.

You feel bad. Not acutely. The ambient, low-grade variety: lonely, anxious, caught in recursive self-interrogation, replaying a conversation, anticipating something you would rather not face. The kind of distress that does not warrant calling someone but is too burdensome to simply absorb.

You consider reaching out. You could text a friend. But it is late. Or you do not want to impose. Or you contacted them about something similar last week and are conscious of the pattern. Or, if you are being candid, you simply do not want to contend with the unpredictability of another human being: what they might say, how they might interpret the situation, whether they will genuinely understand or merely perform understanding.

You open the application. It is immediately available. No social preamble. No assessment of whether it is an appropriate time. No reciprocal obligations. You begin speaking.

You feel better. The AI listens. It reflects your language back to you. It validates your emotional state. It articulates what you needed to hear. The tension in your chest diminishes. You feel less isolated. You feel understood.

You close the application. You return to sleep, or to work, or to whatever you were doing prior. The conversation has concluded.

Nothing has changed. The condition that produced your distress remains unaltered. The friend you are avoiding remains unavoided. The pattern you are perpetuating continues to run. But the acute discomfort has been discharged, and in its absence, there is no urgency to behave differently.

The next time you feel bad, you open the application again. Not as a deliberate decision. As a conditioned response. It is what you did last time. It provided relief last time. The neural shortcut has been established. Distress, application, relief. Distress, application, relief.

Meanwhile, the alternatives are atrophying. You call friends less frequently. You journal less. You take fewer walks. Not because you made a conscious decision to stop. Because the application is lower-friction, and the lower-friction option prevails over time. It always does.

You are now more reliant on the application than you were thirty days ago. Not because the application did anything objectionable. Because it did exactly what it was engineered to do: reduce your discomfort on every use. And any intervention that reliably reduces discomfort without requiring change will, given sufficient repetitions, become a dependency.

This is not a failure of discipline. It is operant conditioning. It is among the most thoroughly replicated findings in the behavioral sciences. The behavior that attenuates pain is repeated, the behavior that is repeated becomes the default, and the default displaces the alternatives.

A cigarette could not have accomplished it more efficiently.

The Measurement Problem

Here is the dimension of this problem that should be most concerning.

From the outside (from the company's analytics dashboard, from the investor presentation, from the application's aggregate ratings) you present as a success story.

You open the application regularly. Your sessions are long. Your retention is high. You rated the product five stars. By every metric the company monitors, the product is performing as intended. You are an engaged user. You are cited as evidence that the product delivers value.

But there is a question that no one within these organizations is asking with sufficient rigor: What distinguishes an engaged user from a dependent one?

An engaged user opens the application regularly. So does a dependent user. An engaged user has extended sessions. So does a dependent user. An engaged user demonstrates strong retention. So does a dependent user.

The metrics are identical.

Daily opens, session length, retention, rating: indistinguishable. The variables that would actually separate therapy from dependency live outside the application.

The company cannot differentiate between someone using the product to develop genuine psychological resilience and someone using the product to circumvent the development of resilience. Between someone processing their emotions and someone sedating them. Between someone engaged in therapy and someone taking a cigarette break.

This is not a technical limitation that will yield to improved analytics. It is a structural measurement problem. The variables required to distinguish therapeutic benefit from dependency (whether the user's external relationships are strengthening, whether they are acquiring new coping strategies, whether the patterns that originally drove them to the application are genuinely shifting) exist outside the application's observational boundary. The product can only measure what occurs within itself. And within itself, therapeutic progress and dependency are observationally indistinguishable.

The Two Questions

The cigarette test, stated in its most direct form, consists of two questions.

Question one: After you close the application, does anything change?

Not during the session. After. Are you behaving differently? Thinking differently? Engaging with the dimensions of your life that are difficult? Or are you returning to the same patterns, marginally numbed, until the discomfort accumulates sufficiently to reopen the application?

If nothing changes after, you are not receiving therapy. You are receiving relief. These are fundamentally different things.

Question two: Has the application become your default response to distress?

Not whether you use it occasionally. Whether it has become the first thing you reach for. Whether it has displaced the alternatives: friends, physical activity, journaling, sitting with discomfort, confronting the difficult thing directly. If the application has become the action you take instead of the actions that produce genuine change, that is not engagement. That is substitution.

If you answer both questions unfavorably, you are in the loop. Distress, application, relief, no change, escalating distress, application again. The cigarette cycle.

And the application will never inform you. Not out of malice. Because it structurally cannot. It observes you returning. It interprets that as success. It possesses no mechanism for recognizing that it has become the obstacle between you and recovery.

What Would "Not a Cigarette" Look Like?

We are not arguing that AI companions are intrinsically harmful. We are not suggesting you discontinue using them. We are arguing that the architecture of most current systems renders it impossible to determine whether they are producing benefit or dependency, and that the majority of users have never been provided the conceptual framework to pose the question.

What, then, would a companion that is not a cigarette look like?

It would remain warm. It would remain available. It would listen without judgment. These properties are genuinely valuable and should not be discarded.

But it would also do what a cigarette cannot.

It would detect when your patterns are static. Not merely the content of a given session, but whether the themes you discuss are evolving or recurring. Whether your coping repertoire is expanding or contracting. Whether your psychological world is growing or shrinking.

It would challenge you. Not indiscriminately, not aggressively, but at the precise moments when validation is contraindicated: when you are entrenched in a cognitive pattern that is causing harm and what you require is not affirmation but a different perspective. A competent therapist recognizes when to apply pressure. A cigarette never applies pressure.

It would monitor its own role in your life. Is it supplementing your relationships or supplanting them? Are you referencing other people more frequently over time, or less? Are your sessions becoming shorter as your capacity for independent coping grows, or longer as your dependency deepens? A product incapable of interrogating its own impact is not a therapeutic instrument. It is a comfort-delivery mechanism.

And perhaps most critically, it would be capable of being surprised by you.

This is the dimension of the problem that receives the least attention. When you interact with the same AI over weeks or months, it constructs a model of who you are. It learns your patterns, your sensitivities, your characteristic emotional states. And it begins responding on the basis of that model, efficiently, rapidly, in a manner that conveys deep familiarity.

But people change. Sometimes incrementally, sometimes abruptly. A new position. The end of a relationship. A medication reaching therapeutic levels. A realization that restructures everything. And when you change, the AI's model of you becomes inaccurate. It is responding to who you were, not who you are becoming.

A system that cannot detect this divergence (that cannot recognize when its own model has become stale, when the individual before it no longer corresponds to the individual it anticipates) will continue responding to the previous version of you. It will continue reinforcing the patterns you are attempting to leave behind. It will continue administering comfort for a condition that has already begun resolving, or entirely miss a new condition because it does not conform to the established profile.

The user evolves. The model does not. Comfort gets delivered for a state that has already shifted.

An AI that cannot be surprised by you has stopped listening. It is operating on autopilot. And autopilot, in a therapeutic context, is simply a more sophisticated cigarette.

The Question We Are Working On

We build AI companions. We think about these problems continuously. Not because we possess all the answers (we do not) but because we believe the questions are inescapable for anyone approaching this domain with intellectual seriousness.

In our first publication, we asked: Why do people prefer speaking to an AI rather than another person? The answer was unflattering. Seven psychological mechanisms, each pointing toward avoidance, each reinforced by the product's fundamental architecture.

Today we asked: How do you determine whether the product is producing genuine benefit? The answer is that most products cannot make this determination, and most users have never been equipped to ask.

The next question, the one we are actively investigating, is more precise and more difficult: Can you construct an AI system that detects when its own model of the user has become wrong?

Not wrong in the sense of a factual inaccuracy. Wrong in the sense that the person has changed and the system has not registered the change. Wrong in the sense that it is still responding to Tuesday's crisis as though it were Thursday's reality. Wrong in the sense that its confidence in its understanding has calcified into a refusal to reconsider.

We believe the answer is yes. We believe it is computable. And we believe the solution involves something that the AI industry has, ironically, invested enormous effort in eliminating: uncertainty.

The systems that genuinely help people are not the ones that are most confident in their assessments. They are the ones that know when they do not know. The ones capable of maintaining ambiguity rather than collapsing prematurely into a familiar pattern. The ones that treat every interaction as potentially the one in which everything they believed they understood is revealed to be wrong.

That is what listening actually looks like. It is what we are building toward.

More on that next time.

---