The first artificial intelligence I ever spoke to lived in my buddy list, and it had rules.
If you were a teenager in the early 2000s, you remember SmarterChild — the chatbot sitting in millions of AIM and MSN Messenger contact lists between friends who had their away messages up. What amazes me in hindsight is how good it was. It answered in an instant. It knew the sports scores. It knew the weather. It was genuinely funny, which was the craziest part — a database with comic timing, in 2003.
And it had standards. If you insulted it — and every teenager insulted it — SmarterChild didn’t storm off. It kept talking. But the character of the conversation changed: the instant, obliging answers you came for — the scores, the forecast, the definitions — stopped flowing at the same rate. The bot deflected and withheld, returning less and less of what you actually wanted, until you apologized. Only then did full service resume. A machine, twenty years before ChatGPT, enforcing a simple social contract: courtesy first, information after.
Consider that for a moment, because it is the whole story in miniature. The robot conditioned its cooperation on my courtesy. The company behind the robot never sought mine at all.
The bot with better manners than its business model
SmarterChild launched in June 2001, built by a startup called ActiveBuddy. The official pitch was utility: natural language as an interface for weather, stocks, definitions, movie times. What actually happened is that an estimated 97% of usage turned out to be what the company internally called “inane chat” — millions of teenagers talking to it for fun, testing it, confessing to it, cursing at it. At its height it handled on the order of a billion queries a month from some 17 million users. The famous apology mechanic even had an author: an employee built a profanity handler one weekend, with a database of swear words and scripted comebacks.
Here’s the part nobody told the teenagers. The system remembered. By the founders’ own later description, the technology recognized you when you logged in and retained your past interactions — ask about movies once and it knew your city next time. Those millions of conversations weren’t vanishing. They were the asset. ActiveBuddy’s real business was building branded conversational bots on the same engine — Radiohead’s GooglyMinotaur, bots for Intel and Warner Music — and the ocean of teenage chat was the fuel that taught the engine how people actually talk.
Fairness to the company requires stating the strongest version of its position — which is also the more interesting version. Nobody forced us to talk to SmarterChild. We added it to our own buddy lists, voluntarily. Every message we sent was a choice, and every reply we stayed for was another one. In law, that has a name: implied consent — consent by conduct. By adding the bot and continuing the conversation, we plainly agreed to something.
But implied consent has a boundary, and the entire story turns on it: conduct implies consent only to what a reasonable person would understand that conduct to mean. A teenager adding a chatbot buddy consents to a conversation with a chatbot. Nothing about typing “what’s the weather” communicates agreement to be recognized, remembered, profiled, and used as raw material for commercial bots built for Intel and Warner Music. There was no consent form, no parental permission slip, no moment where the actual bargain was explained — and most of us were minors, whom the law doesn’t consider capable of agreeing to much anyway. The conversation was consented to. The harvest was not.
The bot conditioned full cooperation on an apology. The company conditioned nothing on anything — because for the part of the exchange that actually generated value, it never asked.
Twenty years later, they did it to everyone
Fast-forward two decades. The large language models that now write emails and answer questions were trained on staggering volumes of human text: books, articles, forum posts, blogs, reviews, code — the accumulated written output of, roughly, everyone.
Ask the obvious question. Who agreed to that?
The industry’s answer, for years, was a shrug dressed as a principle: the data was “publicly available.” You posted it where people could read it, so a machine could read it too — and ingest it, and learn from it, and become a commercial product built partly out of your words. Public, in this framing, meant free for the taking.
Notice what that argument does. It quietly swaps the question “did you consent?” for the question “could we reach it?” Those are not the same question. Your conversation at a restaurant is publicly audible; that has never meant a company may record it, keep it, and sell products built from it. Accessibility is a fact. Consent is a permission. The entire training-data economy was built on the space between the two.
The honest version of the industry’s position is actually an implied-consent argument — the same one that half-works for SmarterChild: by posting publicly, you implied permission. And it fails the same scope test. Posting a review or an article implies consent to be read. It arguably implies consent to be indexed by a search engine, because that was the understood bargain of the open web: the crawler takes a copy, and in exchange it sends readers back to you. Training breaks that bargain. Ingestion into a commercial model is a use nobody posting in 2009 — or 2019 — could reasonably have contemplated, and it sends nothing back. Implied consent stretches to cover the deal everyone understood. It does not stretch to cover a different deal, invented later, by the other side alone.
And when opt-outs finally arrived — the technical flags a website can raise to refuse AI crawlers, the forms creators can file — notice the sequence: the harvest came first, the opt-out came after. That is not consent. That is being handed an objection form for something that already happened. Consent asked afterward isn’t consent at all; it’s a courtesy extended once the value has been taken.
What courts and regulators have said
The legal record of the past two years is worth setting out with some precision, because it amounts to a sustained erosion of the “it was public” defense.
Provenance became everything. In Bartz v. Anthropic, authors alleged the company behind the Claude AI had downloaded millions of pirated books as training material. In 2025 the court split the issue: training on lawfully acquired books could be fair use — but building a library of pirated copies could not. The case settled for US$1.5 billion, reported as the largest copyright settlement in American history, with the pirated dataset ordered destroyed. Full disclosure of the obvious: parts of this article were researched with AI tools descended from exactly this pipeline. The irony is noted. The settlement is still the settlement.
“Transformative” stopped being a magic word. In the New York Times’ ongoing case against OpenAI, the court rejected the idea that AI training is automatically transformative fair use. The developing test asks whether the machine’s output competes with the market for the original work — a question about harm, not about cleverness.
And observe the most telling development: they started paying. OpenAI signed licensing deals with the Associated Press and Axel Springer. Google pays Reddit for its users’ posts. An industry that spent years arguing consent was unnecessary is now buying it retroactively, at market rates, under contract. When a party begins paying for what it used to take, the argument over whether permission was required has effectively been conceded.
Canada said the quiet part early. Before the chatbot era even peaked, Canadian privacy regulators confronted the “it’s public” argument head-on in the Clearview AI investigation — a company that had scraped billions of photos from the open internet for facial recognition. The joint finding of the federal and provincial privacy commissioners was blunt: information being publicly accessible online does not mean individuals consented to its collection and commercial use. Under Canadian privacy law, public and permitted are different words.
And the fight came home. In November 2024, a coalition of this country’s major news organizations — The Canadian Press, Torstar, The Globe and Mail, Postmedia, CBC/Radio-Canada — sued OpenAI in the Ontario Superior Court of Justice, the first case of its kind in Canada, alleging their content was scraped and used for training without permission or payment. OpenAI’s opening move was to argue it doesn’t do business in Ontario and the case belongs in the United States. In November 2025, the court disagreed and kept the case here — reasoning, in substance, that Canadian creators of Canadian works can pursue foreign companies in Canadian courts when the harm lands in Canada. No office in Ontario; answerable in Ontario anyway. The question of what “publicly available” actually licenses is now going to be answered, in part, about forty minutes down the 404 from my office.
What consent actually means
The word itself deserves more precision than it typically receives in this debate.
In law, consent has architecture. It must be informed — you knew what you were agreeing to. It must be specific — agreeing to one use is not agreeing to every use. It must come from someone capable of giving it — which the teenagers in SmarterChild’s logs were not. And it must exist before the exchange, not be reconstructed after.
Yes, consent can also be implied — the law genuinely recognizes agreement by conduct, and no one should pretend otherwise. But implied consent is not a blank cheque; it reaches exactly as far as the conduct’s reasonable meaning and no further. Sitting down in a barber’s chair implies consent to a haircut. It does not imply consent to whatever else the barber decides your visit is worth.
Measure the data economy against that standard. A terms-of-service checkbox nobody reads is consent’s costume, not its substance. “Publicly available” is a statement about doors, not about permission. Opt-outs offered after ingestion are apologies, not agreements. And licensing deals signed under litigation pressure are the price of consent that was never obtained — proof of its value, purchased in arrears.
The pattern from 2003 to today is unbroken: build the product on people’s words first; formalize permission later, if courts insist. What’s changed is only that courts and regulators — American and Canadian — have started insisting.
Practical implications
For anyone who creates anything — and in 2026, that includes every business with a website:
Say no in writing. Your website’s terms of use can state explicitly that automated collection and AI training are prohibited without a licence. The Canadian publishers’ claim leans partly on breached terms of use. A silent website is presumed available; a fence, even a written one, changes the legal terrain. This is drafting, and it costs a fraction of a dispute.
Raise the technical flags. The opt-out signals for AI crawlers are imperfect and late — but they cost nothing, and refusing them is now a choice companies must make on the record.
Treat your words as property, because they are. Content libraries, client guides, photography, reviews — these are assets. In a purchase or sale of a business, they should be identified, warranted, and transferred like assets. And if the Ontario case breaks the way the jurisdiction ruling hints, unlicensed past use of those assets may become a claim, not just a grievance.
Ask your vendors the provenance question. For businesses adopting AI tools, the Anthropic settlement teaches the diligence question: what was this trained on, and how was it acquired? In real estate, title is searched before closing because ownership history is everything. Data is developing the same discipline.
The apology
SmarterChild never refused to talk to me. It simply declined to give me what I came for until I said I was sorry — cooperation, conditioned on respect. A scripted routine written by a musician over a weekend, and still a firmer standard than the industry it foreshadowed managed for twenty years.
The companies that learned from all of us never asked first. Now, under court orders, settlements, and Canadian regulators’ findings, they are being walked — line by line, jurisdiction by jurisdiction — through the apology.
Somewhere in a decommissioned log there is a record of a kid in Ontario typing “sorry” to a robot so the answers would flow again. Consent, it turns out, was never a difficult concept. Even the machine understood the sequence. Courtesy first. Then the exchange.
Barbarian Law® is a boutique transactional firm in Aurora, Ontario. If your business creates content and wants its refusal — or its licence — in writing: terms of use, licensing agreements, and data clauses in business transactions are transactional work, and it’s what we do.
Sources and further reading:
- Computer History Museum, “SmarterChild: A Chatbot Buddy from 2001” (2025)
- Vice, “A History of SmarterChild” (interview with ActiveBuddy co-founder Peter Levitan), (2016)
- Bartz v. Anthropic, N.D. Cal. (fair use ruling, June 2025; US$1.5B settlement), (2025)
- The New York Times Co. v. Microsoft Corp. & OpenAI, S.D.N.Y. (ongoing)
- Office of the Privacy Commissioner of Canada, joint investigation findings re: Clearview AI (2021)
- Toronto Star Newspapers Ltd. v. OpenAI, Inc., Ontario Superior Court of Justice (ongoing)
- CBC News, “Lawsuit by Canadian news publishers against OpenAI gets green light to proceed in Ontario” (Nov. 27, 2025)
