Your phone isn’t secretly listening to you for ad targeting, part 2

Last month, the FTC fined Cox Media Group (CMG) and two other companies to settle allegations that the firms deceived customers about the utility of a marketing service sold by CMG, which claimed it could target ads based on conversations recorded by the microphones on consumers’ phones. Wired initially reported on the fine; from the FTC’s press release:

The Federal Trade Commission will require Cox Media Group (CMG) and two smaller marketing firms to pay a total of $930,000 to settle allegations they deceived customers by falsely claiming to offer an AI-powered service that could target localized ads based on conversations captured from consumers’ smart devices and that consumers had opted into such targeting …

CMG, MindSift and 1010 Digital Works claimed their “Active Listening” branded marketing service listened in on consumers’ conversations overheard by smart devices, in real time, to target advertising, according to the complaints. The three companies claimed the Active Listening service would allow small businesses to advertise to consumers in the small businesses’ desired locations.

According to the complaints, this service did not, in fact, listen in on consumers’ conversations or use voice data at all—nor did the service accurately place ads in customers’ desired locations. Instead, the service the companies provided consisted of reselling—at a significant markup—email lists obtained from other data brokers.

If this case sounds familiar, it might be because I covered CMG’s marketing materials and the various reporting that breathlessly propagated its claims in September 2024 in a piece titled Your phone isn’t secretly listening to you for ad targeting. From that piece:

Advertising agencies often overstate their capabilities. Perhaps the most infamous example of this is Cambridge Analytica, which claimed to be able to target potential voters based on psychographic profiles but in reality was determined by the UK’s ICO to have used “well recognised processes using commonly available technology” which raised “concern internally about the external messaging when set against the reality of their processing.” With respect to voice data collection, CBS News partnered with the Electronic Frontier Foundation to monitor outgoing traffic to potential audio data harvesters when commercially valuable keywords were uttered into a smartphone and didn’t observe any. The BBC engaged in a similar exercise and reached the same conclusion.

The FTC provides more detail on the claims made by CMG, which essentially resold products offered by the other two defendants, in its complaint (note that I have condensed the list of claims):

In 2023, Respondent disseminated or caused to be disseminated claims for its Active Listening marketing services on its website, cmglocalsolutions.com, that included the following:

a. “At a basic level, your smartphone is technically always listening. Any voiceactivated device must eavesdrop around the clock to pick up on ‘wake words’ or the voice commands used to activate various virtual assistant services. Phrases like Hey Siri and OK, Google can only work if a smart device is always listening. So, it makes sense that our devices can’t ‘shut off’ and ignore everything that not ‘wake words.’”
b. “What does your iPhone do with all that info? Smart devices use data of all kinds to create a consumer ‘profile’ of you to show you the most relevant ads.”
c. “Don’t Just Know What They’re Searching For-Know What They’re Talking About”
d. “Voice data goes beyond search engine data, so every casual conversation between two consumers becomes a tool for you to target, retarget, and retain customers. Our technology is on the cutting edge of voice data processing. We can identify buyers based on casual conversations in real time. It may seem like black magic, but it’s not-it’s AI.”
e. “The growing ability to access microphone data on devices like smartphones and tablets enables our technology partner to aggregate and analyze voice data during pre-purchase conversations. The result? Unprecedented understanding of consumer behavior, so we can deliver personalized ads that make your target audience think: wow, they must be a mind reader.” …
i. “By incorporating and analyzing customer data gleaned from conversations happening around smart devices, we can pinpoint where and when customers are most likely to engage with ads.”
j. “Creepy? Sure. Great for marketing? Definitely.”

The FTC’s complaint is unequivocal: CMG’s claims about the efficacy of its Active Listening product were utterly fabricated. From the complaint (emphasis mine):

Respondent’s representations about its Active Listening service were false or misleading. Contrary to Respondent’s statements, smart devices with voice collection capabilities did not transmit voice data to Respondent’s Active Listening service, and Respondent did not use “AI to detect pertinent conversations.” In fact, Respondent’s Active Listening service did not collect or use voice data in any manner. Rather, Respondent’s advertising and marketing services were nothing more than consumer email list buying, an industry practice where data brokers sell lists of email addresses of individuals presumed to have particular interests or demographic characteristics. Respondent resold these lists at a significant markup over the cost of the data.

In other words, as the UK’s ICO said of Cambridge Analytica, CMG’s Active Listening service relied on “well recognised processes using commonly available technology.” The Active Listening episode is a scandal not because of commercial surveillance concerns but because it involves a far more pedestrian misdeed: false advertising.

As I note in my original response to the marketing claims, the notion that consumers’ phones surreptitiously listen to their conversations dates back years and wasn’t catalyzed by this case: Mark Zuckerberg was asked about the ability of his company’s apps to listen to and mine consumers’ real-world conversations from their phone microphones for relevant ad targeting data in a Senate hearing in 2018 (he denied it). The case of CMG is interesting because it seemingly reignited media interest in the persistent myth — bordering on superstition — each time the firm’s marketing deck was circulated. And this media interest was overwhelmingly credulous, taking the claims at face value; in the most recent round of intrigue, which catalyzed my article, the New York Post wrote, “Your long-held suspicions are confirmed, according to a report: Your phone really is listening to you.”

There are two problems with this theory.

First, modern smartphones block app access to the microphone absent consent, and they display an indicator when the microphone is in use. But the most severe practical constraint would simply be the battery’s limitations; the battery would drain quickly if voice data were continuously recorded and transmitted in real time.

And second, and more consequentially, this data would almost certainly be less valuable than the other data available to advertising platforms for targeting. If an advertiser wants to target ads based on keywords, isn’t Search (or, now, a chatbot) the better, more down-funnel option for that? How could an advertiser be sure that the keyword they are targeting from a live conversation wasn’t being used in a non-commercial context? Why would simply uttering a word imply that a person is in-market for a product? There are innumerable advertising platforms that an advertiser can use to target consumers who have expressed commercial intent through observed behaviors. Uttering a keyword in casual conversation almost certainly carries less intent signal than conducting a Search or browsing a vendor’s website.

When I’ve made these two points, either in written form or in conversation, I’ve been met with some form of, “But I’ve been shown ads for things that I had just previously discussed within earshot of my phone. This can’t possibly be a coincidence.”

It isn’t. A person seeing an ad for a product they recently discussed in conversation would be a coincidence if ad platforms targeted ads randomly. But ad platforms target ads based on observed actions — behavioral profiles — that correlate with the same impulses that lead to discussing products live. The omitted variable in this situation is the interest or commercial intent that inspired the person to talk about the product. If a person exclusively saw ads only for things they had recently discussed in conversation, it might suggest their phone was eavesdropping on them. But when a person sees an ad for a product they have a commercial interest in, and that was recently discussed in a conversation, the underlying targeting is likely being done against their interests, as inferred from other behaviors. So here, a person’s interest in a product is a confounder: it creates a non-causal association between the conversation and the ad, while in fact, both the conversation and the ad are consequences of the same underlying commercial intent.

The frequency illusion, or Baader-Meinhof phenomenon, is also useful for interpreting this phenomenon. A person is bombarded by ads throughout their digital life; they may see 19 ads for kayaks in the course of a day and only really notice the 20th after discussing kayaks with a friend. These two ideas aren’t mutually exclusive — again, the person is being shown ads for kayaks because they have commercial intent, which also prompts the conversation — but the last impression is conspicuous because it followed a discussion.

For consumers and practitioners alike, my utility explanation often falls flat. In a conversation on Sam Harris’ podcast from 2018, computer scientist Jaron Lanier recounts claims of data utilization from advertising salespeople in pitch meetings:

It is changing all the time, so it is really different every week. It is kind of—I have been present a couple of times when the various tech companies present themselves to the big advertising conglomerates in the annual negotiation of fees, which is this amazing hidden ritual that I wish more people could see. When you will see the representatives from companies like Facebook and Google present themselves, they are so overreaching and creepy and they probably are exaggerating, but they will talk about being able to just measure your soul. It will just be like, “Oh, we are measuring the movement of your phone, which can tell us how you are walking and then from that we can derive your mood and whether you are having a period if you are a woman.” All kinds of stuff. “We are following where you go, we are following all this stuff, we are following your voice tone when you do voice stuff, we are following your facial expression if we have a camera on you. We are measuring all this stuff. We know all these things about your health that you will never know. We know all these things about your mood, about your mental state, about your state of attention. And we offer all of this amazing encyclopedia of spying on you to you our true customer, if only you’ll give us more money so you can have your piece of controlling the world.”

It is just such a weird thing. In terms of the ground truth of how much is actually measured and how much of that measurement is actually valid, and of the data that is valid that is gathered, how much of it is used in any way that preserves any of that validity, that’s extremely hard to determine. I do not think anybody really knows, and it is constantly in flux. But at any rate, the attitude is get as much as you can and be as creepy about it as possible when you are trying to earn money from your true customers. And the true customers, of course, are the ones that they call the advertisers. I prefer to call them the manipulators.

I was an operator when I first heard this interview; I remember being amused at the thought of an ad platform pitching “gait-based” targeting. These propositions are implausible, but not because of technological limitations; see IDNet: Smartphone-based Gait Recognition with Convolutional Neural Networks, which proposes a model architecture for identification based on accelerometer-measured gait.

Rather, their infeasibility stems from their inferiority to commercial data in both volume and predictive power. Amazon announced at its Upfronts presentation last month that it reaches 90% of US households with an authenticated, logged-in identity. Even if it could trust its validity, would an advertiser ever prefer a keyword uttered in a casual conversation or a person’s gait to that person’s retail purchasing history as a targeting parameter?

The public conception of advertising targeting may be overly baroque. People imagine microphones, cameras, gait recognition, and other byzantine forms of data collection because those mechanisms feel maximally invasive and exploitative. But the most valuable advertising data is often much more prosaic and unremarkable: behavioral history. The CMG episode demonstrates that even marketers selling advertising products understand this intuition.