17 May 2025 Search

Previously: PET projects or real privacy?, stop putting privacy-enhancing advertising technologies in web browsers

From recent news: Google is reportedly wrapping up work on in-browser privacy-enhancing advertising features. Instead, they’re keeping third-party cookies and even encouraging going back to older user tracking methods like fingerprinting.

Google’s Privacy Sandbox projects were their own special case, and it certainly looks possible that their continued struggles were mostly because of trying to replicate a bunch of anticompetitive tricks from Google’s old ad stack inside the browser. Privacy-enhancing technologies are hard enough without adding in all the anticompetitive stuff too. But by now it looks more and more clear that it wasn’t just a problem with Privacy Sandbox trying to do too much. Most of the hard problems of PETs for advertising are more general. Although in-browser advertising features persist, for practical purposes they’re already dead code. Right now we’re in a period of adjustment, and some of the interesting protocols and code will probably end up being adaptable to other areas, just not advertising.

While PETs for advertising were a bad idea for a lot of reasons, all I’m going to list here are the big problems they couldn’t get over.

PETs without consent didn’t work. The original plan in the early days of Privacy Sandbox was to deploy to users with a simple Got it! dialog. That didn’t work. Regulators in the UK wrote (PDF),

We believe that further user research and testing of the dialogue box, using robust methodologies and a representative sample of users, is critical to resolve these concerns. Also, it is not clear if users will be prompted to revisit their choices, and the frequency of this.

In the real world, PETs will be required to get the same kind of consent that other adtech is. Buried in a clickwrap agreement isn’t going to pass inspection. PETs are catching the same kinds of complaints over lack of consent as any other adtech. And getting consent will be hard, because…

Users are about as creeped out by PETs as by other kinds of tracking. Jereth et al. find that perceived privacy violations for a browser-based system that does not target people individually are similar to the perceived violations for conventional third-party cookies. Co-author Klaus M. Miller presented the research at FTC PrivacyCon (PDF):

So keeping your data safer on your device seems to help in terms of consumer perceptions, but it doesn’t make any difference whether the firm is targeting the consumer at the individual or group level in the perceived privacy perceptions.

Martin et al. find substantial differences between the privacy that users expect and the privacy (ish) features of PETs. In fact, users might actually feel better about old-fashioned web tracking than about the PET kind.

In sum, the use of inferences rather than raw data collected by a primary site is not a privacy solution for users. In most instances, respondents judged the use of raw data such as browsing history, location, search terms, and engagement data to be statistically the same as using inferences based on that same data. Further, for improving services across contexts, consumers judged the use of raw data as more appropriate compared to using inferences based on that same raw data.

PET developers tried to come up with solutions that would work as a default for all web users, but that’s just not realistic considering that the research consistently shows that people are different. About 30% of people prefer cross-context personalized advertising, 30% really don’t want it, and for 40% it depends how you ask. PETs are too lossy for people who want cross-context personalized ads and too creepy for people who don’t. (In addition to this published research, there is also in-house research at a variety of companies, including at some of the companies that had been most enthusiastically promoting PETs.)

PETs never had a credible anti-fraud story. One of the immutable laws of adtech is that you can take any adtech term and put fraud after it, and it’s a thing. PETs are no exception.

  • Anti-fraud lesson of the 1990s: never trust the client

  • Anti-fraud lesson of the 2000s: use machine learning on lots of data to spot patterns of fraud

  • PETs: trust the client to obfuscate the data that your ML would have needed to spot fraud. (how was this even supposed to work?)

If PET developers could count on an overwhelming percentage of users to participate in PETs honestly, then there might not be a problem. A few people would try fraud but they would get lost in the noise created by PET math. But active spoofing of PETs, if they ever caught on, would have the same triad of user motivations that open-source software does: it feels like the right thing to do (since the PETs come from the same big evil companies that people are already protesting), you would have been able to make money doing it, and it’s fun. Any actual data collected by PETs would have been drowned out by fake data generated on principle, for money, or for lulz.

PETs didn’t change the market. The original, optimistic pitch for PETs was that they would displace other surveillance advertising technologies in marketing budgets and VC portfolios. That didn’t happen. The five-year publicity frenzy around Google’s Privacy Sandbox might actually have had the opposite effect. The project’s limitations, well understood by adtech developers and later summarized in an IAB Tech Lab report, encouraged more investment in the kinds of non-cookie, non-PET tracking methods that Mozilla calls unintended identification techniques.

Just as we didn’t see articles written for end users recommending PETs as a privacy tip—because the privacy they provide isn’t the privacy that users want—we also didn’t see anyone in the advertising business saying they were cutting back on other tracking to do PETs instead. Even Google, which was the biggest proponent of PETs for a while, lifted its 2019 ban on fingerprinting as Privacy Sandbox failed to take off. And Google was not just a technology provider for PETs, but also a possible client, with consumer phone and thermostat brands, among others. The most informative report about Privacy Sandbox was the absence of a report: no case study was ever released about Google’s own results using it for the Pixel and Nest brands.

PETs would create hard-to-predict antitrust issues. If users are still creeped out by PETs, and advertisers find PET features too limiting, then the designers of PETs must be splitting the difference and doing something right, right? Well, no. PETs aren’t just about users vs. advertisers, they’re about large-scale platforms vs. smaller companies. PETs introduce noise and obfuscation, which tend to make data interpretation only practical above a certain data set size—for a few large companies, or one. Designers of PETs can tune the level of obfuscation introduced to make their systems practical for any desired minimum size of company.

The math is complicated enough, and competition regulators have enough on their to-do lists, to make it hard to tell when PET competition issues will come up. But they will eventually.

PETs would have made privacy enforcement harder. This year’s most promising development in privacy news in the USA is the Honda case. Companies that had been getting by with non-compliant opt outs and Right to Know forms are finally fixing their stuff. CCPA+CPRA are progressing to their (intended?) true form, as a kind of RCRA for PII. Back in the 1980s, companies that had a bunch of random hazardous materials around decided that it was easier to safely get rid of them than to deal with RCRA paperwork, and something similar is happening for surveillance marketing today.

PETs would have interfered with this trend by making it harder for researchers to spot problematic data usage practices, and helping algorithmic discrimination to persist.

Conclusion: learning from the rise and fall of PETs. In most cases, there should be little or no shame in chasing a software fad. At their best, hyped-up technologies can open up a stale industry to new people by way of hiring frenzies, and create change that would have been harder to do otherwise.all right, the cryptocurrency and AI bubbles might be an exception because of the environmental impact, but the PET fad wasn’t that big. Having been into last year’s trendy thing can feel a little embarrassing, but really, a trend-driven industry has two advantages.

  • A trend can give you a face-saving way to dig up and re-try a previous good project idea that didn’t get funded at the time. (this could still happen with prediction markets)

  • Investing in a trend can be an excuse to fix your dependencies (I once got to work on fixing software builds, making RPMs, and automating a GPL corresponding source release, because Docker containers were a big thing at the time) and produce software that’s useful later (PDF-to-structured-text tools, so hot right now)

In the case of PETs there probably should have been more user research earlier, to understand that the default PETs without consent idea wouldn’t have worked and save development time, but that’s a deeper problem with the relative influence of people who write code and people who do user research within companies, and not just a PET thing.

Not all the development work that went into PETs was wasted, because PETs are still really promising in other areas, just not advertising. For example, electricity markets could benefit from being able to predict demand without revealing when individual utility customers are at home or away. PETs are already valuable for software telemetry—for example, revealing that a certain web page crashed the browser without telling the browser maintainer which users visited which pages—and could end up being more widely used for other products, where the manufacturer and user have a shared interest in facilitating maintenance and improving quality. But advertising is different, mostly because it’s unavoidably adversarial. Every market has honest and dishonest advertisers, and advertising’s main job is to build reputation by doing something that’s practical for a legit advertiser to do and as difficult as possible for a dishonest advertiser. As the shift to a low-trust economy continues, and more software companies see their reputations continue to slide, real ad reform solutions will need to come from somewhere else. More: Sunday Internet optimism