{"slug": "you-can-opt-out-you-just-can-t-afford-to", "title": "You Can Opt Out. You Just Can't Afford To.", "summary": "Asahi Linux is relicensing all of its repositories from MIT to GPLv2-or-later, citing AI-assisted code laundering in which in-development code is lifted from maintainers' forks and rewritten by models to hide its origin. Separately, a publisher that used Google's new AI opt-out control measured a 99% drop in AI impressions and an estimated 23% loss of search clicks, even though the control does not prevent content from being used for training.", "body_md": "Three different groups spent the last few weeks testing what it costs to tell an AI system no. All three got an itemized bill.\n\n## 1. Asahi Linux Changed Its License Because Rewriting Code Became a Prompt\n\nAsahi Linux — the reason Linux runs on Apple Silicon at all — is [moving every repository from MIT to GPLv2-or-later](https://itsfoss.com/news/asahi-linux-drops-mit-for-gpl/). The stated reason is code laundering: in-development code lifted out of maintainers' personal forks, republished without credit or rewritten by a model to hide its origin, by parties the project says want to monetize Linux on Apple Silicon. The sharp part is the project's own note on MIT — the license explicitly allows this and has no carveouts for bad-faith actors. Its one real condition is keeping the copyright notice.\n\nLaundering is not new — it meant renaming variables and shuffling functions until the diff stopped matching. What changed is the unit cost. Permissive licensing in a hostile world always leaned on one economic assumption: that laundering a nontrivial driver cost more than contributing upstream. A prompt is cheaper than either. But notice what copyleft actually buys, because it is not what the headline implies — the GPL cannot stop a determined rewrite any more than MIT could. What it removes is the closed, paid fork, and it forces a derivative's source into public view, which makes copying detectable. That is deterrence traded for detection: a smaller win than it sounds.\n\nThe collateral damage is the part worth studying. The BSDs depended on Asahi partly because it was MIT — OpenBSD's Apple Silicon support rides on its m1n1 bootloader — and relicensing without documented consent from every prior contributor drew immediate criticism from that side. The cited precedent is WINE going LGPL in 2002 to stop proprietary forks — but WINE was defending against companies. This project treats license choice as a threat model against a technique.\n\n**Why it matters:**\n\n- **For ICs:** If you maintain anything reverse-engineered or hardware-adjacent, your license was a cost calculation and the costs just moved. Re-run it.\n- **For leaders:** Permissive upstreams you rely on may harden. Audit which dependencies are one maintainer grievance away from copyleft.\n- **For founders:** \"We rewrote it with AI, so it's ours\" is not a legal strategy, it is a discovery exhibit. Keep provenance you would hand over gladly.\n- Every open source norm built on attribution assumed provenance was hard to erase. It isn't, so the norms are being rewritten as license text.\n\n## 2. A Publisher Pressed Google's AI Opt-Out Button and Measured the Crater\n\nIn August, Google shipped a control letting publishers keep content out of search AI features without blocking the crawler entirely — previously an all-or-nothing choice. One site [pressed the forbidden button and measured what happened](https://signal.zyppy.com/p/we-opted-out-of-google-ai-search): out on September 1, back in on September 29, two 28-day windows of Search Console data compared as an interrupted time series. AI impressions fell more than 99% within an hour, and returned about as fast.\n\nThe numbers are the interesting part. AI features were roughly 10% of total search impressions but about 21% of clicks, because click-through on AI citations ran near 0.9% against a sitewide 0.4% — more than double. Then a halo effect: appearing in AI answers lifted the site's ordinary organic clicks by around 2.3%, strongest on URLs with the most AI visibility. Estimated total cost of opting out, 23% of search clicks — one site, heavily optimized for citations, so treat it as a yardstick, not a constant.\n\nThe mechanism should bother you. The control is real, instant and works exactly as documented, which is precisely why it is unusable — a lever that costs a fifth of your traffic is not a choice, it is a price. It also governs the wrong thing: opting out of AI answers does not stop your content being used for training. Those are separate switches, and the easy one is the one that bills you.\n\n**Why it matters:**\n\n- **For ICs:** Before anyone argues about opting out, instrument it. You cannot have an opinion about a channel whose share you have never measured.\n- **For leaders:** Run the experiment on a segment, not the whole domain, and run it long enough to separate seasonality from the effect.\n- **For founders:** If AI citations are a fifth of your inbound, that is a platform dependency with no contract behind it. Price it like one.\n\n## 3. Google Promised No Human Would Read Your Files. Then It Edited the Promise.\n\nGemini's Workspace extensions launched in 2023 with an explicit commitment: personal content from Gmail, Docs and Drive is not seen by human reviewers. That commitment is gone. An October edit to the original announcement, and the current help pages, now [confirm that connected Drive and mail are routed into human review pipelines](https://github.com/thisisme6598/gemini_files/), advising users not to connect apps holding confidential information they would not want a reviewer to see.\n\nThe engineering detail matters more than the policy reversal. Retrieval fires ambiently — no explicit Drive or Gmail tag required; a passing mention of a trip or a bill can trigger a background Workspace lookup. And because tool-calling is probabilistic, the same prompt yields no retrieval at all, a metadata summary, a verbatim quote including account and confirmation numbers, or the entire unredacted file dropped into context. Test once, see nothing, and you have proven nothing — you observed a single draw. Automatic scrubbers do not save you: regex and entity recognition cannot deidentify narrative text, and a first name plus a city plus a date identifies someone even when each field looks harmless.\n\nThis is the consent model underneath every AI integration shipping right now: a boundary that is a sentence on a vendor-versioned policy page, over a data path nondeterministic enough that you cannot audit your own exposure from outside. The design conclusion is unglamorous: gate at the connector, not in the prompt. Decide what the agent can reach, assume everything reachable is in scope for human review, and stop treating vendor assurances as an architectural control.\n\n**Why it matters:**\n\n- **For ICs:** Never connect a data store to an assistant on the strength of a promise. Connect a scoped, read-limited view or nothing.\n- **For leaders:** Re-read the privacy terms of every AI tool you approved more than six months ago. At least one has changed under you.\n- **For founders:** If you make an AI product, a quiet terms edit is a trust event. Version your promises visibly or inherit this reputation.\n\n## The Verdict: Real or Hype?\n\n**Copyleft as a defense against AI code laundering → Real but early.** It cannot stop the rewrite; it kills the closed commercial fork and drags copies into daylight, which is narrower than the announcement suggests. **Platform opt-outs as meaningful consent → Hype.** The button works flawlessly, which is how you know the leverage was never in the button. **Connector-level data gating as engineering work → Real.** Probabilistic retrieval means a policy sentence is not a boundary, and the only boundary you control is the one you wire up.", "url": "https://wpnews.pro/news/you-can-opt-out-you-just-can-t-afford-to", "canonical_source": "https://fromtheterminal.substack.com/p/you-can-opt-out-you-just-cant-afford", "published_at": "2026-11-10 17:37:16+00:00", "updated_at": "2026-11-10 18:52:28.022333+00:00", "lang": "en", "topics": ["ai-policy", "generative-engine-optimization", "ai-search", "ai-crawlers", "ai-ethics"], "entities": ["Asahi Linux", "Google", "OpenBSD", "WINE", "Apple", "Linux"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/you-can-opt-out-you-just-can-t-afford-to", "markdown": "https://wpnews.pro/news/you-can-opt-out-you-just-can-t-afford-to.md", "text": "https://wpnews.pro/news/you-can-opt-out-you-just-can-t-afford-to.txt", "jsonld": "https://wpnews.pro/news/you-can-opt-out-you-just-can-t-afford-to.jsonld"}}