Privacy Washing: Detecting Internal Contradictions in Privacy Policies
This paper introduces a multi-stage LLM-based pipeline to detect "privacy washing"—internal contradictions in privacy policies where commitments are undermined by other clauses—and applies it to 2026 and 2015 corpora to reveal recurring third-party sharing contradictions in a significant portion of companies, while noting that prevalence figures are lower bounds due to the lack of human expert validation.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every time you visit a website or download an app, you are asked to agree to a privacy policy. These are the long, dense documents that explain how a company handles your personal information. For decades, researchers and regulators have operated on a simple assumption: if a company writes a promise in that document, it keeps that promise. The idea is that these policies serve as a clear notice, allowing you to make an informed choice about your data. If a policy says "we do not sell your data," a reasonable person expects that the company will not sell it. However, this assumption relies on the document being internally consistent. What happens when a policy makes a reassuring promise in one section, only to describe a practice that directly undermines that promise in another section of the same text? This is the question a new study tackles, exploring a phenomenon the author calls "privacy washing."
The concept draws a parallel to "greenwashing," where a company might advertise a small environmental effort while continuing harmful practices. In the digital world, privacy washing occurs when a policy contains a commitment, such as a promise to limit data sharing, that is contradicted by a description of actual data handling found elsewhere in the same document. The contradiction is not always a simple logical error like saying "we collect X" and "we do not collect X." Instead, it is often more subtle: a broad, comforting statement like "we do not sell your personal information" sits alongside a detailed description of sharing user identifiers with advertising partners. A reader who stops at the reassuring promise may never reach the contradictory practice buried further down, creating a misleading impression of safety.
To investigate how common this is, a researcher named Thomas Brackin developed an automated system to scan privacy policies for these internal contradictions. The system does not read the documents like a human would, from start to finish. Instead, it breaks the text down into tiny, individual statements. It first identifies sentences that sound like promises or commitments, such as "we will not share your data." Then, it finds sentences that describe actual practices, like "we share data with partners." The system then pairs these statements together to see if they clash. It uses a series of filters to remove pairs that are clearly unrelated, such as a promise about security paired with a practice about data collection. It then employs artificial intelligence models to judge whether the remaining pairs represent a genuine contradiction. To ensure accuracy, the system uses a panel of three different AI models; a pair is only flagged as a confirmed contradiction if at least two of the three models agree.
The researcher applied this system to two large collections of privacy policies. One collection, gathered in 2026, included 123 companies from various industries. The other, gathered in 2015, included 115 companies. This allowed the researcher to see if the problem was a new issue or a long-standing feature of how these documents are written. The results showed that privacy washing is a recurring pattern. In the 2026 collection, at least one confirmed contradiction appeared in about 12 percent of the companies. In the 2015 collection, the figure was higher, appearing in about 36 percent of the companies. However, the researcher notes that the two collections were analyzed with slightly different settings, so the difference in numbers cannot be definitively attributed to a change over time. What is clear is that the same types of contradictions appear in both eras.
The most frequent contradictions involved third-party data sharing. A common pattern found was a company promising not to sell or share data, while simultaneously describing how it shares identifiers and commercial information with advertising partners. This specific pattern aligns with a legal settlement involving the company Sephora in 2022, where regulators determined that sharing data with advertising partners for targeted ads could legally count as a "sale" under California law, even if the company claimed it did not sell data. The study found that these contradictions often arise not from deliberate deception, but from structural factors. As laws change, companies add new sections to their policies to comply with regulations, such as adding a "we do not sell" statement to meet California requirements. These new sections are often added without reconciling them with older sections that describe existing data-sharing practices, creating an internal conflict.
The study also revealed that the system is not perfect and likely misses many contradictions. The automated filters are designed to be very strict to avoid false alarms, which means they may discard pairs that a human expert would consider contradictory. Furthermore, the system relies on artificial intelligence to judge the contradictions, and these AI models have not yet been validated against human experts. The researcher emphasizes that the findings are candidates for further review rather than final proof of wrongdoing. The study does not claim that companies are intentionally lying; rather, it suggests that the way these documents are constructed—often by different teams adding sections over time—creates a structural environment where reassuring promises and detailed practices can easily contradict one another.
Ultimately, the research provides a new way to look at privacy policies. Instead of just measuring how long they are or how hard they are to read, this study measures whether the text contradicts itself. The findings suggest that for a significant number of companies, the promise of privacy and the description of practice exist in tension within the same document. While the study cannot determine if users are actually being misled, it identifies a specific textual condition where the "net impression" of the policy could be deceptive, even if every individual sentence is technically true. The work serves as a tool to flag these inconsistencies, offering a starting point for regulators, companies, and researchers to examine whether the promises made in privacy policies truly match the practices described within them.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.