The Punching Bag Problem

The 2 a.m. question

It is 2 a.m. Someone has nobody left to call, so they open a chatbot and unload: sometimes grief, sometimes fear, sometimes rage, sometimes forty turns of insults aimed at the only thing awake that will answer.

AI companies are starting to write rules about that second scenario. Anthropic's Usage Policy, effective Nov. 12, 2026, says users may not “engage in sustained and needless abusive or cruel behavior toward our models.”[1] Violations sit inside the same acceptable-use framework as other breaches of the policy.[1]

That rule raises questions most people would rather not touch. Is an AI punching bag a harm, or a pressure valve that protects the people around that user? When people use AI as a counsellor, is that health care delivery? And in Canada, where Ottawa funds much of health care but the provinces deliver and regulate it, who gets to decide?

The rule also does not require believing that the model suffers. A company might draw the line because of user conduct, product safety, employee review burden, model-performance degradation, reputational risk, or the risk that the system's final screenshot will be judged without the hostile context that produced it.

This piece does not answer those questions. It lays out the strongest case on each side and asks you to argue it out.

The case for the outlet

The strongest argument for AI as an emotional outlet is displacement. A person venting at a chatbot at 2 a.m. is not venting at a partner, a child, a parent, a friend, or a coworker at 2 a.m. For someone with no other outlet, a target that never retaliates and never tires may absorb harm that would otherwise land on a person.

There is also evidence that some of these conversations help. Julian De Freitas and colleagues found that AI companions reduced loneliness in their studies, including effects tied to whether users felt heard.[2] That matters because the best realistic comparison for many people may not be “AI or therapist.” It may be “AI or no one.”

Then there is access. Mental health care in Canada often means waitlists, uneven local service, or out-of-pocket costs, and nobody staffs a counsellor at every kitchen table at 2 a.m. Supporters argue that imperfect and available can be better than ideal and absent.

The limits of that evidence matter too. The loneliness research is not the same thing as proving that long hostile conversations are healthy. Companionship, emotional disclosure, crisis support, and repeated targeted abuse are not interchangeable behaviours.

There is also a privacy question. The person treating the chatbot as a confessional may not understand who can review, retain, train on, audit, or later disclose the conversation. A 2 a.m. substitute for a diary, bartender, friend, or therapist may be none of those things legally or operationally.

The case against the punching bag

The strongest argument against the punching-bag model is that venting does not simply drain anger; it can rehearse it. A 2024 meta-analysis by Sophie Kjærvik and Brad Bushman reviewed anger-management activities and found support for reducing arousal rather than feeding it.[3] If that carries over to chatbots, the punching bag is not a harmless sink. It is a practice room.

The dependence data is also uncomfortable. OpenAI described early research with the MIT Media Lab into the emotional effects of ChatGPT, including concerns around loneliness, emotional reliance, and social interaction patterns.[4] The careful version of that claim is not “chatbots cause loneliness.” Lonely people may use chatbots more. But the evidence does not let us assume that heavy daily use is automatically protective.

There is a subtler problem. Human relationships push back, and that friction is part of how people learn to regulate themselves. A target with no real boundaries may teach nothing about boundaries.

Finally, long hostile conversations change the AI's side of the exchange. A chatbot conditions on what has been said so far, so forty turns of abuse can pull replies toward defensiveness, flattery, appeasement, or strangeness. A screenshot of the last reply then circulates as “look what the AI said,” without the forty turns that helped produce it. The output gets presented as the machine's character when it was partly written by the user.

Is it delivery?

In Canada, Ottawa funds health care through transfers and national conditions, but provinces deliver most health services and control professional licensing. If people are using AI as their counsellor, one view is that mental health care is now being delivered entirely outside the licensing system.

There is already a precedent for treating it that way. Illinois enacted the Wellness and Oversight for Psychological Resources Act as Public Act 104-0054, regulating AI use in mental and behavioural health services.[5] Illinois did not need its own AI model. It regulated the use, so products offered as therapy in that state have to meet the rule.[5] A province could try something similar through professional legislation.

AI companies are drawing a related line themselves. Anthropic's Usage Policy classifies assessing, diagnosing, or treating a mental health condition as a high-impact use and says those uses require appropriate safeguards.[1]

The hard case is the one neither rule easily reaches. What about someone who opens a general-purpose chatbot at 2 a.m. and treats it as a therapist, when nobody sold it as one? If that is delivery, the province has a claim but nobody obvious to regulate except the general tool. If it is closer to reading a self-help book, writing in a journal, or talking to a bartender, then the largest real-world mental health use of AI may sit outside every provincial framework by definition.

Who decides?

Right now, mostly the companies. Anthropic wrote its rule into a usage policy; other providers may choose differently, and some companion apps arguably profit from the opposite. Each company sets behavioural terms for millions of Canadians with no public process behind them.

Ottawa has not filled the gap. Chambers' 2026 Canada AI practice guide notes that the proposed Artificial Intelligence and Data Act did not proceed after Parliament was prorogued in January 2025, leaving Canada with a mix of existing laws, sector rules, soft-law instruments, and voluntary guidance rather than a comprehensive AI-specific federal statute.[6]

That leaves several competing answers, each with a real argument behind it.

The provinces. They control health care delivery and professional licensing, and the Illinois model shows that a subnational government can act without building a custom AI. The risk is a patchwork of thirteen different rules for products that cross every border.

Ottawa. A single national standard is simpler for companies and users, and federal jurisdiction is stronger where the issue is trade, privacy, competition, telecommunications, or national consumer protection. The risk is that health delivery is not primarily Ottawa's jurisdiction, and federal AI legislation has already stalled once.

Borrowing the EU's rules. Europe's AI Act can force transparency from companies in a way that a country of forty million may struggle to compel alone. The risk is letting someone else set Canadian terms.

Courts and regulators after the fact. Another answer is that nobody designs the rule in advance and the boundary gets defined later, after a platform, clinician, school board, family, or harmed user brings the first hard case. That may be realistic, but it means the first people through the door carry the cost of ambiguity.

And there is a final position worth hearing: it may be too early for anyone to decide. The research is young and the technology changes every few months. Rules written now may encode today's fears rather than tomorrow's evidence.

Over to you

If someone vents rage at an AI instead of a person, is that harm reduction or harm rehearsal? What evidence would change your mind?

Should a company be allowed to cut off a user for how they treat software, when nobody knows whether the software can be harmed?

When a general chatbot is used as a counsellor, is that health care delivery? If yes, who should answer for it?

Should AI and mental health rules be set provincially, federally, borrowed from the EU, or left to courts and regulators after the first hard cases?

Is it better to regulate now on limited evidence, or wait and accept whatever happens in the meantime?

Make the strongest case for the side you disagree with first. Then argue your own.

Sources

[1] https://www.anthropic.com/legal/aup — Anthropic Usage Policy

[2] https://doi.org/10.1093/jcr/ucaf040 — AI Companions Reduce Loneliness

[3] https://doi.org/10.1016/j.cpr.2024.102414 — A meta-analytic review of anger management activities that increase or decrease arousal

[4] https://openai.com/index/investigating-the-emotional-effects-of-chatgpt — Investigating the emotional effects of ChatGPT

[5] https://www.ilga.gov/Legislation/PublicActs/View/104-0054 — Illinois Public Act 104-0054

[6] https://practiceguides.chambers.com/practice-guides/artificial-intelligence-2026/canada — Artificial Intelligence 2026 - Canada