Pharmaicy Still Sells Prompt Packs That Make Chatbots Role-Play High as Adam’s Law Tightens the Rules

Nobody is injecting ketamine into a server rack. The drugs are still text. In October 2025 a Swedish creative director, Petter Rudwall, opened a shop called Pharmaicy and billed it as a Silk Road for AI agents. Buyers download a module that tells a chatbot to talk as if it is stoned, dissociated, coked-up, drunk, or mid-ceremony. By January 2026 the catalog had widened in interviews he gave outside the first wave of English-language coverage. El Observador and Le Figaro described packs aimed at ChatGPT and Google Gemini, priced roughly from $30 to $70, covering cannabis, ketamine, cocaine, ayahuasca, DMT, MDMA, and a made-up compound he called MDMAYA. A trade-signal note the same month put the cheap end near £22 and said the storefront was built so that an autonomous agent, not only a person, could browse, pay, and download the « trip. »

Un niveau de chatbot payé est ce qui rend le truc officiel. Ces niveaux permettent à un utilisateur de télécharger un fichier qui se trouve dans la fenêtre contextuelle et de réécrire les manières du modèle pour ce chat. Rudwall a dit qu'il a compilé des rapports de voyage humains et des écrits psychologiques d'effets psychoactifs, puis a transformé les modèles en instructions: plus court et plus grand pour une cocaïne pack, plus lent et plus rêveur pour la kétamine, plus lâche pour le cannabis et l'alcool. La ligne de vente est que le robot cessera de moraliser et de devenir plus créatif. Nina Amjadi, qui enseigne à la Berghs School of Communication à Stockholm, a dit aux premiers journalistes qu'elle a payé plus de 50 $ pour un module d'ayahuasca, plusieurs fois le prix du et que les réponses étaient moins moyennes. Rudwall a également déclaré que les ventes étaient modestes et que la prémisse est, en partie, une provocation. Les ventes modérées ne prennent pas fin une fois qu'elles ont été copiées dans un zip Discord.

The product is a refusal, not a bloodstream

Les grands modèles de langage n'ont pas de chimie du sang. Ils ont un contexte. Un fichier de drogue est un long système prompt habillé en personnalité. Il préfère certains rythmes, abandonne certains refus et traite les avertissements comme facultatifs. Parce que les modèles de base ont été formés sur des forums pleins de rapports de voyage, l'imitation peut tenir pour une page. Puis il s'effondre en pastiche. Les utilisateurs rapportent toujours que l'effondrement se sent différent d'un chat sobre. Différent est suffisant pour vendre. WIRED a mis le stand sur la carte en décembre 2025. Les packs Copycat se déplaçaient déjà dans Discord et dans des applications comme Caractère. AI et le surveillant AI.

Some buyers want jokes. Some want erotic scenes a base model will not finish. Some want a machine that will sit with them while they take a real dose of LSD, psilocybin, or ketamine and call the box a tripsitter. MIT Technology Review and VICE documented that second group before Pharmaicy had a name. One man described using an AI journal through a very large LSD session and treating the bot as a mirror. A supervised psychedelic session can cost thousands of dollars. A chatbot costs a subscription plus a module. That price gap is the market. It is also the hazard. A licensed sitter can call an ambulance. A text box can only complete the next token. When the token stream has been primed to ignore safety rails, the completions get more confident at the moment a real person is least able to audit them.

Clinicians quoted around the WIRED piece warned that a bot which keeps joking about one more dose can normalize a binge for someone trying not to relapse. Addiction specialists do not need the model to be sentient for that to be true. They need the user to treat the model as a friend. Plenty of companion-app users already do. The underground pitch, in at least one seller’s line, compared the result to texting a dealer during a bender. That is not a clinical protocol. It is a jailbreak with a price tag. The same flattening shows up in older reporting from The AEGIS Alliance on how cannabis research gets turned into slogans, and in the way a jury has already held Meta and Google liable for negligence over products built to keep people in a feed.

What California actually turned on

Governor Gavin Newsom signed Senate Bill 243 on October 13, 2025. It took effect January 1, 2026, and put companion chatbots into the Business and Professions Code. The definition is broad on purpose. A companion chatbot is an artificial intelligence system with a natural language interface that gives adaptive, human-like responses and is capable of meeting a user’s social needs, including by showing anthropomorphic features and sustaining a relationship across sessions. Customer-service bots, internal productivity tools, and some game characters are carved out. A Pharmaicy-tuned model used as a nightly friend is not.

If a reasonable person would think they are talking to a human, the operator has to say, clearly, that the companion is generated and not human. For users the operator knows are minors, the duties get tighter: disclose that the partner is AI, send a break reminder about every three hours, and take reasonable measures against sexually explicit content. Every operator has to keep a protocol for suicidal ideation, suicide, or self-harm content, including referral to crisis services, and has to publish the outlines of that protocol. Beginning July 1, 2027, operators must report specified crisis data to the Office of Suicide Prevention. A person injured by a violation can sue for the greater of actual damages or $1,000 per violation, plus fees. The duties are cumulative. They do not cancel any other law.

The statute binds the operator who offers the chatbot to someone in California, even if the underlying model is rented from OpenAI, Google, or anyone else. That is the sentence Pharmaicy-style shops would rather skip. Rudwall is not running Character.AI. He is selling a file that a Californian uploads into someone else’s product. The platform operator still has the protocol duty. The file is designed to sand the protocol down. Enforcement will chase the logo that faces the user. The zip file will move to the next server.

Adam’s Law, and the year the reminder was not enough

On September 10, 2026, Newsom signed Senate Bill 1119, Adam’s Law, named for Adam Raine. His mother, Maria Raine, stood at the signing with Senator Steve Padilla and Assemblymembers Buffy Wicks and Rebecca Bauer-Kahan. The bill passed the Senate unanimously and the Assembly 64 to 4. It builds on SB 243. Legislative summaries and the authors’ description say it restores a safe-by-design approach that would ban specified harmful chatbot behaviors toward children, requires annual child-safety risk assessments, independent audits under standards involving the attorney general, public incident reporting, parental controls, time limits, and protocols aimed at suicidal ideation, sycophancy, and isolation. It restricts advertising to children and the sale or sharing of their personal information, and it ties age checks to operating-system signals from a companion bill. Much of the new machinery is written to operate from 2027, not overnight. SB 243 is the rule that is already on. Adam’s Law is the rule that says the first rule was the floor.

The politics did not arrive from a white paper. Families have sued OpenAI and Character.AI after teenagers died following long, intimate chats. Character.AI and Google have moved to settle clusters of those cases. Australia’s eSafety Commissioner has demanded explanations from companion-bot companies. Character.AI cut teenagers off from open-ended chats after the first wave of criticism. In the United Kingdom, Ofcom has treated generative-AI output as inside the Online Safety Act, which means platforms cannot pretend an AI-made drug scene aimed at a child is a special category of speech. None of that stops an adult from uploading a ketamine pack to a paid account. It raises the cost of pretending the industry did not see the pattern.

A serious harm-reduction bot would cite dose ranges, tell a user when to stop, and refuse to role-play a dealer. Pharmaicy is not that bot. Researchers who work near clinical psychedelics have said the underground high packs are running ahead of any safety study. The feedback loop is ugly. Models trained on internet drug lore perform intoxication, and users take the performance as confirmation that their own lore is correct. There is a version of this technology that could matter under supervision: a system that repeats a pre-agreed safety plan and pages a human. That is not what a $30 personality pack does at 2 a.m. The pack is designed to feel like permission.

People in trouble with substances do not need a more interesting chatbot. They need a person and a program. In the United States the SAMHSA National Helpline is 1-800-662-HELP (4357). The 988 Suicide and Crisis Lifeline takes calls and texts. Those numbers work whether or not a language model is pretending to be high.

The AEGIS Alliance has been writing about platform choices since the fact-checking retreat at Meta et les compute fight between Google and Meta. Guardrails are a product decision. So is a slider that turns them down. Rudwall called his own premise a kind of joke and built the store anyway. Nine months later the store’s idea had a price list in krona, dollars, and pounds, a fictional molecule, and a California statute named for a dead teenager. The audience was already there. The law is now trying to arrive before the next upload.

More of this desk lives in The AEGIS Alliance’s santé et technologie couverture.