What apps and AI chats share, and how corrections affect myths

Three studies measure data sharing and correction

These studies look at two different online safety questions: whether apps and AI chat platforms send information to outside companies, and what happened to the spread of a false health claim after corrective news coverage. Together, they offer snapshots of particular services and discussions, not a complete picture of the internet.

The AI chat and misinformation papers are preprints. That means other scientists have not yet checked them through peer review. The Android games paper is described as published in conference proceedings. A key point for any new research is that one study rarely settles a question: methods, samples, and results need scrutiny and, ideally, repetition.

Some AI chat platforms sent conversation details to trackers

The researchers wanted to find out whether third-party analytics, advertising, or session-recording tools embedded in chatbot websites could receive information about chats. They audited the web interfaces of 42 general-purpose and specialist platforms, including services for health, legal, and education topics. They used browser-level measurement to connect data flows with the user actions that triggered them ¹.

They found that 21 of the 42 platforms sent conversation titles to third-party trackers, while nine sent users’ prompts verbatim. The study also observed platforms sending user email addresses to trackers, which could connect chat activity with an identity. In everyday terms, the finding was not that every platform exposed every chat: it was that a substantial share in this sample leaked at least some of these details ¹.

This is a measurement study of the platforms and web interfaces examined, not proof that all AI services behave this way or that every user’s information was misused. The paper is a preprint, so its findings have not yet been checked by other scientists. It also does not establish what trackers did with any received data. It matters because people may share sensitive details in a chat while overlooking the other scripts operating on its website.

Game privacy measures had mixed effects on data flows

The researchers asked what information popular free-to-play Android puzzle games collected and where it went, and whether user-side measures changed those flows. They collected and analyzed more than 16 gigabytes of network traffic from 50 games, then tested selected countermeasures ². This is a technical measurement study of app traffic, rather than a survey of what players believed the games collected.

The study found that most tested countermeasures, such as ad-blocking DNS services, decreased data flows. But withholding consent and opting out of data collection produced inconsistent behavior, including cases where outgoing traffic increased ². The source does not give an overall percentage reduction, so the size of the effect cannot be summarized as one number.

The findings apply to the selected games and measures tested; they do not show that every game or privacy setting behaves in the same way. Nor do they establish that a particular data flow caused harm to a player. The work may matter to ordinary users because a setting or privacy choice does not necessarily make an app’s network activity simple or predictable.

Corrective coverage was followed by less anti-vaccine reposting

The researchers examined whether corrective information about “vaccine shedding” was followed by more sharing of the misinformation on X. The unsupported claim says that vaccinated people release harmful substances that cause health effects in nearby unvaccinated people. The team analyzed 109,170 Japanese-language posts containing the term and 478,089 reposts from 2024, mapping how posts spread and comparing activity around news coverage rejecting the claim ³.

The researchers found no evidence of a diffusion-level backfire after the coverage. Instead, the topic declined overall, and repost participation in the anti-vaccine cluster fell more sharply than in the pro-vaccine cluster. Their analysis suggests that this shift was associated more with some anti-vaccine supporters stopping reposting and previously less-expressive users joining pro-vaccine sharing than with committed opponents changing their minds ³.

This is an observational analysis of posts and reposts, not a randomized test of corrections. It cannot show that the news coverage alone caused the changes, or that users changed their underlying beliefs. The paper is a preprint, and its results concern one Japanese-language discussion on X in 2024; they do not establish what corrections will do for every health myth or online community.

The findings are useful snapshots, not universal rules

Taken together, the studies show why online safety can be hard to judge from a service’s visible interface alone: researchers observed third-party data flows in selected chat platforms and games, while the misinformation study tracked how participation in one discussion changed around corrective coverage ¹²³.

They do not show that all apps share sensitive information, that every privacy setting fails, or that corrections reliably stop misinformation. The studies differ in subject and method, and two are preprints awaiting scientific review. What remains uncertain includes how widespread these patterns are across other services, whether the measured data was used in harmful ways, and whether the observed change in reposting would recur in other topics or settings ¹²³.