AI Chatbots Are Leaking Your Conversations to Ad Trackers — Even When You Opt Out

When users open ChatGPT, Claude, or a competitor and start typing, they assume the conversation stays between them and the model. A study published on 7 May 2026 suggests that assumption is wrong more often than the industry lets on.

Researchers examined four widely used AI platforms — ChatGPT, Claude, Grok, and Perplexity — and found that all four transmitted portions of user conversations to third-party advertising infrastructure. In some cases, data flowed to trackers even when users explicitly declined cookies.

The finding matters because the commercial model underpinning most consumer AI services depends on advertising revenue or advertising-adjacent data practices. Privacy policies are written to cover this; the actual network traffic, according to the study, sometimes tells a different story.

What the Study Found

The research team tested each platform by conducting structured conversations and monitoring outbound network requests. The goal was to identify which third-party domains received data and whether standard opt-out mechanisms blocked transmission.

The results were not uniform. ChatGPT shared conversation metadata with several ad-tech firms under conditions the study described as aggressive by industry standards. Claude, developed by Anthropic, transmitted fewer data points but still connected to at least two ad networks during testing. Grok, the platform associated with Elon Musk's X, sent user input to tracker domains at a rate comparable to ChatGPT. Perplexity, the AI-powered search engine that has marketed itself on privacy, ranked highest on opt-out compliance — but still failed a subset of the tests.

The critical detail is the cookies finding. When users rejected non-essential cookies on each platform, ChatGPT and Grok continued transmitting certain session identifiers to third-party domains. Perplexity's opt-out performed better but was not complete. The study did not find evidence of full conversation transcripts being forwarded, but partial data — device identifiers, session timestamps, query fragments — was sufficient, the researchers argued, to reconstruct behavioral profiles over time.

The Tracking Architecture Behind the Curtain

Modern advertising technology relies on an ecosystem of intermediaries: demand-side platforms, data management companies, audience segmentation firms. When a user visits a page or, apparently, initiates an AI query, the hosting domain can embed code from dozens of these intermediaries. Each one logs something — a cookie, an IP address, a referral string. Alone, each data point is trivial. Aggregated across sessions, the profile is detailed.

AI chatbot interfaces are web applications. They load ads, they embed tracking pixels, they run JavaScript from third parties. The study found that even platforms that position themselves as privacy-first were pulling in the same infrastructure as a typical news website. The difference is that an AI conversation is more intimate than a browsing session. Users type questions about health, finance, relationships, work — the kind of input that, combined with behavioral tracking, produces a high-resolution personal profile.

The practice is not illegal. Most privacy laws, including GDPR in Europe and the California Consumer Privacy Act, require disclosure and consent for non-essential cookies. But disclosure often comes buried in privacy policies of thousands of words, and consent mechanisms frequently default to acceptance. The study's finding — that some platforms transmit data even when users opt out — goes further than what most disclosures admit.

Why AI Companies Have a Structural Conflict

The AI industry is engaged in a race that requires enormous capital. Training large language models costs tens or hundreds of millions of dollars per generation. Running inference — the process of generating responses — is computationally expensive and grows with user volume. Advertising represents one of the few revenue streams capable of funding this scale.

Free AI tiers are the most obvious vector. Platforms offering ChatGPT, Claude, or Grok without charge need to monetize those sessions somehow. Subscription revenue from paid tiers covers some costs, but the growth calculus for AI companies depends on converting free users into paid subscribers — and on demonstrating audience scale to advertisers. Both imperatives create pressure to extract behavioral signal from free-tier conversations.

Privacy promises are marketing tools. When Anthropic launched Claude, it marketed the model in part around constitutional AI principles and a stated commitment to safety. When Perplexity launched, it marketed itself as an alternative to Google that did not track queries. These positioning moves generate user trust and, critically, help AI companies avoid regulatory scrutiny in the near term. But the commercial infrastructure underneath has not changed to match the marketing.

This is not unique to AI. The advertising technology industry built its current architecture over two decades on the logic that behavioral data improves targeting, and targeting improves ad prices. Platforms that monetized user attention through advertising had a financial incentive to resist consent opt-outs in ways that legal frameworks only slowly constrained. AI companies entered this ecosystem late but quickly adopted its conventions.

The Regulatory and User Stakes

If these findings hold under broader scrutiny, they represent a systematic problem rather than isolated company behavior. Four major platforms — representing the vast majority of consumer AI usage — all transmitted data to ad trackers. The variation was in degree, not in kind.

The stakes for users are concrete. Behavioral profiles assembled from AI conversations could be used for purposes ranging from targeted advertising to insurance underwriting to employment screening. A conversation with an AI legal assistant could reveal information about a pending lawsuit. A conversation with a health-adjacent AI tool could produce data of interest to insurers. The downstream uses of that data are not visible at the moment of input.

Regulators have begun to pay attention. The European Data Protection Board, which coordinates GDPR enforcement across EU member states, opened a consultation on AI and data protection in 2025. The U.S. Federal Trade Commission has signaled interest in AI platform data practices without issuing specific enforcement guidance. Neither body has yet produced binding rules specific enough to cleanly address the scenario the study describes — tracker domains embedded in AI chat interfaces, transmitting session data regardless of consent settings.

The practical consequence is that users who believe opting out of cookies protects their AI conversations are operating on a false assumption, at least for some platforms. Until enforcement catches up to the architecture, the burden of protecting conversational privacy falls on users — and the tools available to do so are limited. Some privacy advocates recommend browser extensions that block third-party requests, but these tools can break AI platform functionality, since many platforms depend on third-party CDNs and analytics services for core features.

The study does not claim that AI companies are deliberately exfiltrating full conversation logs. What it documents is structural: the advertising technology embedded in AI interfaces creates data flows that operate, at least partially, outside the consent mechanisms users encounter. That gap between stated privacy commitments and actual network behavior is where the problem sits — and where regulators will eventually have to look.

This publication's science desk tracks developments in AI governance, data regulation, and platform accountability.