A PDF That Topped Hacker News
On September 16, independent researcher jorgegarciaherrero.com posted a PDF titled Prompt like a butterfly, sting like a tracker on his personal site.
The document sat on the Hacker News front page for roughly a day, finishing at 252 points / 72 comments. That kind of traction is rare on HN—only a few hundred posts per year break 200 points, and the ones that do are usually major company launches or significant open-source releases. A personal blog PDF hitting 252 means it hit a nerve.
The PDF takes about 15 minutes to read. The core finding, in one sentence: multiple AI chat products are leaking user conversations into the advertising ecosystem, and this isn't a bug—it's a designed data flow.
Original document: https://jorgegarciaherrero.com/wp-content/interactivos/20260916-Prompt-like-a-butterfly-sting-like-a-tracker-(clean).pdf
How the Tracking Works: Hidden Signals in Your Prompts
The paper's core argument isn't that "some AI company secretly sold your conversations." It's that the tracking chain runs deeper than most people imagine. The path has three layers:
Layer 1: Keyword extraction
After a user submits a prompt, the server performs content understanding. The researcher traced how some products extract keywords from conversation context and feed them into ad bidding systems. Unlike search keywords, which are isolated, conversation keywords carry full context.
Example: You ask ChatGPT, "My stomach has been hurting for weeks, but my company physical isn't scheduled until next month, and I'm worried." The keyword isn't "stomach pain"—it's "stomach pain + company physical + anxiety." That profile is more precise than any cookie.
Layer 2: User profile enrichment
Ad SDKs combine these keywords with existing cookies and device IDs to build complete user profiles. The researcher found third-party ad SDKs embedded in some AI chat products—SDKs that get blocked on the web but can read conversation context inside chat interfaces.
Layer 3: Cross-product tracking
Once conversation data enters the ad ecosystem, tracking works the same as any other channel inside major ad networks. The ads users see on TikTok, Meta, or Google get validated and weighted by these new keyword channels.
None of the three layers is technically complex. What's complex is that this happens in a setting users assume is private.
Scope: Not One Company, but an Industry Default
The paper explicitly tested multiple mainstream AI chat products. I won't repeat the specific names (readers can check the PDF), but the researcher's conclusion is blunt: this isn't a single vulnerability—it's a system-level data flow.
The issue isn't "which company did something bad." It's that:
- This architecture is the industry default—free and low-cost AI products need monetization paths.
- Privacy policies mention it, but bury it in language ordinary users can't parse.
- Current regulations have no specific provisions for conversation data.
The third point is especially critical. GDPR's "legitimate interest" clause has constraints on sensitive data, but most AI chat products classify conversation data as "non-sensitive." The paper quotes parts of these policies—you'll see phrasing like "we may use your inputs to improve our services." Literally true. But what's the purpose of "improving services"? Often, selling ads.
Why This Wasn't Exposed Before: The Unique Nature of Conversation Data
Users have already accepted cookie tracking. Install uBlock, use a private window, and most people feel they're "doing enough."
But conversation data operates at a different magnitude than cookies. Cookies can tell you:
- Which pages were visited
- Which links were clicked
- General interest areas
Conversations can tell you:
- The user's current real intent
- Emotional state (anxiety, anger, sadness)
- Personal identity information (name, address, employer, family)
- Future behavior (planning to buy something, considering a job change, health concerns)
The researcher emphasizes this point: conversation is the first "deep self-disclosure" scenario in product history. Previous cookies, search keywords, social likes—these are all shallow signals users actively expose. Conversations are different: people tend to tell AI chatbots things they wouldn't tell anyone else.
This means once conversation data enters the ad ecosystem, the harm is an order of magnitude greater than the cookie era.
What Users Can Do: Not Much, Honestly
I'll be straight: the advice I can offer is limited, and none of it is a root solution.
Read privacy policies—but most people won't, and those who do can't understand them, and those who understand have no alternative (you can't just stop using AI).
Be careful with sensitive inputs—but if you're already using AI, that advice comes too late.
Consider local models—Llama, Qwen, and Mistral all have decent on-device options. But ordinary users won't configure them; those who try can't run large models; and those with M2 Max / M3 Max hardware are a small minority.
# Minimal command to run a local Qwen model with Ollama
ollama run qwen2.5:14b
# Data stays entirely local, nothing leaves the device
# Cost: Mac mini M2 minimum, 14B model needs 16GB RAM
Two things could actually fix this:
- Vendor transparency—explicitly tell users which third parties conversation data flows to, and offer a paid tier that opts out of ad training and tracking.
- Regulatory intervention—classify conversation data as sensitive data with separate consent requirements.
On the first point, some vendors are already moving (Claude Pro and ChatGPT Plus explicitly promise their data won't be used for ad training). But what about free tiers? That's a question for regulators.
On the second point, the EU may be moving. The US FTC has held hearings. But the pace of legislation… everyone knows how that goes.
Editorial Take
This differs fundamentally from every previous "data privacy" problem: AI chat is the first product format where users outsource their deepest thinking to a third party.
In search, e-commerce, and social media, users output "signals"—clicks, likes, search keywords. Essentially, users actively publish content, knowing (or vaguely knowing) others can see it.
Conversations are different. When a user tells an AI, "I can't sleep lately, I'm worried about being laid off," the default assumption is "this is a private conversation." When that assumption is systematically broken, the damage isn't just to privacy—it's to user trust in the entire AI industry.
Three judgments:
First, regulation will come, but likely 2-3 years behind product adoption. The EU may move first on classifying conversation data, but federal-level US action is unlikely in the short term.
Second, paid tiers will become the new standard. The "no ad training" promise from Claude Pro and ChatGPT Plus will shift from a differentiator to an industry baseline. Free tiers will be required to clearly label data flows or carry explicit "ad-supported" tags.
Third, local models will move from geek toy to mass-market demand. As users internalize the vulnerability of conversation data, demand for Ollama, LM Studio, and on-device inference will grow 10x from current levels. The Apple Silicon + open-source model combo will capture the first wave.
What I'm uncertain about: this is a single independent researcher's work, and I can't independently verify the sample coverage. If a major security team reproduces the study, the conclusions' credibility will jump significantly. But even if some findings get refined, the severity of the underlying problem won't disappear—only the scope might shrink.
One last point: I didn't see the paper name specific tracking SDK sources linked to major players like Anthropic or OpenAI. But if top-tier companies have actually defaulted to integrating third-party ad SDKs in their main products, the situation is worse than the paper describes.
I'll keep following the reproduction research and vendor responses.