How AI Got Into Your Tools Without Asking

The city woke to find intelligence standing at its gates — enormous, beautiful, and free. We wheeled it in, of course. This is a page about what history suggests comes next, told without villains, because the story does not need any.

The gift at the gates

Be honest about the wonder first: for a few years, anyone with a browser has held expert-grade drafting, coding, translation, and tutoring for free or nearly so. That generosity was strategic — growth before revenue, the oldest playbook in software — and it was funded by staggering sums of capital, because every "free" answer burns real electricity through real silicon. A gift this large is never only a gift. It is also a position.

What is inside the horse

Arithmetic, mostly. Inference costs do not vanish because the user pays nothing; someone must eventually be the revenue. The toolkit is standard and openly discussed in industry playbooks: subscription tiers for the few who pay, usage-priced APIs for developers, enterprise contracts — and for the large majority who never subscribe, advertising. In February 2026 the turn became explicit, and not as a leak: OpenAI announced it. From 9 February 2026 OpenAI tested ads for logged-in adults on the Free and Go tiers in the US, with the paid tiers above them exempt — and with the company stating that ads would not change how ChatGPT answers. By August 2026 ads had launched in the UK, Mexico, Brazil, Japan and South Korea, and in September OpenAI began testing “Sponsored Agents”: a business-sponsored chat a user enters by clicking an ad. Around it, a dedicated ecosystem has bloomed: ad networks built specifically for AI chat, affiliate links matched after the reply has finished streaming, and — the detail worth sitting with — "reasoning-time advertising," a closed-beta product category that weaves advertiser context into the model's thinking itself. That last one is the horse speaking through its own mouth — though, on the vendor’s own account, it cannot yet be fitted to the assistants most people use.OpenAI, Testing ads in ChatGPT, published 9 Feb 2026 with later updates, read in a browser 22 Sep 2026: “Today, we’re beginning to test ads in ChatGPT in the U.S. The test will be for logged-in adult users on the Free and Go subscription tiers. Plus, Pro, Business, Enterprise, and Education tiers will not have ads.” and “Update on August 11, 2026: ChatGPT Ads has now launched in the United Kingdom, Mexico, Brazil, Japan, and South Korea.” · OpenAI, Reimagining advertising with AI, dated 16 Sep 2026, read in a browser 22 Sep 2026: “Sponsored Agents are now being tested with select advertisers in the United States.” The pledge that ads would not change answers is from CBS News, 9 February 2026, read at source 11 Sep 2026: OpenAI “vowed that ad presence would not ‘influence or change how ChatGPT responds to user prompts’”.First-hand: until 22 Sep 2026 this paragraph cited CBS News alone, gave the Go tier as $8, and described US-only “sponsored cards below ChatGPT answers”. The primary post is now readable in a browser; it does not state the Go price, so the price was removed.Vendor detail: ChatAds, 2026 — ZeroClick is in “Closed beta with no public access or transparent timeline” and “Incompatible with major LLM platforms like ChatGPT or Claude”. Full note in the sources block below.

A story older than Troy's walls

None of this is unique to AI. The platform lifecycle has run the same three acts for two decades — writer Cory Doctorow gave its decay a famous name, but the shape predates the word: first the platform is generous to users to win them; then it tilts value toward business customers to monetize them; then it squeezes both to answer its investors. Search results grew sponsored rows; feeds grew promoted posts; marketplaces grew pay-to-rank listings. Each turn was announced as an improvement, and each was rational. The question this page asks is not whether AI platforms will walk the same road — capital structure makes some version of it likely — but how visibly they will walk it.

What reaping could look like — a watch-list, not a prophecy

• Answers with sponsors: product placements inside responses — disclosed clearly, buried in fine print, or not at all. The disclosure is the whole game.
• Sponsored reasoning: if advertiser context reaches the model's deliberation, the ad becomes indistinguishable from the advice. This category already exists in closed beta — and the one vendor selling it says it is incompatible with ChatGPT and Claude, because third-party code cannot reach their inference loop. Watch where it goes.
• The tier squeeze: capabilities you relied on migrating upward into pricier plans; the free tier becoming the showroom.
• API repricing: businesses built on a model's pricing discovering that pricing was introductory.
• Default-steering: assistants that recommend their own ecosystem's products, services, and worldview — the quietest lever of all.
• The counter-signals, honestly noted: not every firm is boarding the horse. Some vendors have publicly committed to ad-free, subscription-and-API models and stake their brand on it — and open-weight models remain an escape valve: a downloaded file cannot be taken back, though its licence still sets what you may do with it (compared on AI labs outside the US). The road forks; it is not fated.

Holding the gates

The citizens of this story are not helpless, and the defenses are unglamorous: Demand the label — normalize asking "is any of this answer sponsored?" and rewarding products that answer plainly. Keep an exit — portable prompts, exportable data, and a working knowledge of the local option convert you from captive to customer. Pay for what you value — the old line "if you are not paying, you are the product" is too cynical as stated, but its softer form is sound: revenue models shape products, so choose products whose revenue points at you. Watch the defaults — the horse's real cargo is never merchandise; it is the power to decide what you see first.

Not a siege

Here is where this museum plants its flag, carefully. Companies spending fortunes on compute are entitled to revenue; profit is not the enemy, and an internet funded by advertising built much of what we love. Troy's error was not accepting commerce — it was accepting cargo it could not inspect. So the ask is small and absolute: label the horse. Show the seams. Let the marketplace happen in daylight. The city can trade with anyone, so long as nothing enters the gates pretending to be only a gift.

Part of the Stay Human record. The economics thread: The Silver Lining · Local or Cloud. The machines themselves: The Assembly.