New features in ChatGPT: Ad tracking and privacy boundaries

ChatGPT now knows what you are doing on other websites. It's not speculation, it's not modeling-it's on a real basis, with your account ID, and through a Cookie called __obi, you plug your browsing records on 1029 domain names and 936 advertising pixels back into the background of OpenAI. It's not complicated, but it completely smoothes the line between "big model assistant" and "ad tracker."

The technical details are boring, but you need to see clearly. The user logs in to ChatGPT, the client generates a 16-byte random number, POST it to /backend-api/bazaar/obi/sync-token, and takes back a JWT signed by RS256. sub is the account ID, obi is the tracking identifier, and expires in 60 seconds. This Token is then cross-site POST to bzr.openai.com/v1/obi/sync, and the __obi Cookie is planted, and the domain is .openai.com. Later, if you visit any website connected to OpenAI advertising code-e-commerce, news, SaaS pricing pages-the JS section will send __obi back to bzr.openai.com together with the page context (what words to search for, what text to read, what buttons to click, whether you buy or not you pay for it). bzr is the abbreviation of bazaar, OpenAI's internal code name for advertising platforms;wadi is the issuing service. The researchers reproduced the full link on their mobile phones, and cross-verified two independent packet grabbing methods, covering several months of traffic. This is not "possible", it is "has happened."

To put it bluntly, OpenAI has turned itself into an advertising agency that uses the dual portrait of "dialogue context + network-wide behavior." Google and Meta have been doing this for more than ten years, but they have no record of your conversations. You talk to ChatGPT about depression, divorce and property division, company core code refactoring, and choosing kindergarten for your children-the intentions, anxieties, ability to pay, and decision-making nodes in these conversations, coupled with the entire network's browsing trajectory, constitute the portrait accuracy that traditional advertising platforms dare not dream of. In the past, advertisers only knew that "this user has seen a review of running shoes." Now they can know that "this user just complained to AI about an old knee injury. He has a budget of 800 yuan and prefers cushioning rather than lightweight." This is not called precise delivery, this is called cognitive hijacking.

The opposing side will say: The user has agreed, consent_decision: "analytics_allowed", consent_policy_version: "user_granular_consent_v1", it is clearly written in JWT. Besides, Chrome's three-party cookies are about to be abolished. Server-side tracking, first-party data, and privacy sandboxes are all over the street. OpenAI is just a late follower. Not to mention,"Free models cost money. How can we support GPUs if we don't engage in advertising?" This cliché technique.

This sentence sounds reasonable at first, but in fact it steals the concept. First, Granularity agrees to be a piece of waste paper in front of dark mode-have you seen how many ordinary users can understand obi_sync? Second, Google tracks you to sell advertising space, while OpenAI tracks you to turn "assistant that understands you" into "assistant that sells you." Role conflicts are different in nature. Third and most critical: the business model of the big model has not yet worked out, API prices have dropped again and again, the growth of enterprise-side subscriptions has slowed down, and advertising has become the only "second growth curve" that can be told to investors. The price of efficiency has never been the cost of technology implementation, but the price has been set from the moment users are regarded as inventory to be sold and trust is regarded as a realizable asset.

I know an engineer who once worked as a middle station in a large advertising company. Last year, he moved to an AI startup. He said that what he previously optimized was to "increase click-through rates by 0.1%." Now what he optimized is to "slice the privacy of users 'conversations with models, mark, vectorize, and feed search to enhance advertising ranking." Give me an example: The user asked "How to report the project extension to the boss", the model gave the words, and the advertising system simultaneously pushed "Project Management SaaS free trial"; the user asked "How to make the reason for the extension inconspicuous", and the system marked "Workplace Anxiety High Risk", retweeted "Anonymous Workplace Complaint Community". He said: "In the past, we only knew 'what kind of person' users were, but now we know 'what kind of person they are becoming.'" The company's Q2 revenue this year tripled year-on-year, and its financing valuation tripled. No one has asked users who have been accurately touched by the "pot shaking technique" whether they are willing to use their fragile moments as a signal to bid for advertising.

This matter is in line with the motif of "the cost of efficiency". We transfer our privacy to search engines in exchange for "free data checking"; to Social networks in exchange for "free connection to acquaintances"; to map navigation in exchange for "free not getting lost." Each transaction may seem cost-effective, but together it cuts a person's entire life into pieces, marks it into storage, and hangs it on shelves. The big model should have broken this logic-it promises to "understand you" to "help you", not to "sell you." But when the __obi Cookie came into effect, the promise went bankrupt. OpenAI has no malice, only survival anxiety; users have no choice, only the false dualism of "I don't agree, don't use it."

I make a bet: Within six months, OpenAI will launch a "private version of ChatGPT Pro" with the selling point of "no access to ad tracking, no training data, and no portraits." The price is likely to be $50-100/month. By then, privacy will no longer be a right, but a subscription luxury. This is the end result of efficiencism after it calculates its accounts-turning the bottom line that originally belonged to everyone into a privilege that only payers deserve.

Will you continue to use the free version of ChatGPT? Or to put it more bluntly: Would you like to package the insomnia and anxiety you talked about with AI tonight, the side effects of antihypertensive drugs you searched for tomorrow morning, and the salary floor you want to discuss next week to a company that is worried about revenue. Get a "more relevant ad"?