AI personalization works until it feels like spying.

“Is this personalized, or is it creepy?” Clients ask it about any campaign that uses what the brand knows about a customer. Artificial Intelligence (AI) makes the question louder, because it can tailor a message, an offer or a price to one person at a volume no team could reach by hand. The useful answer is that the line sits where a person finds out how you knew.

When does personalization help, and when does it backfire?

A peer-reviewed study in the Journal of Retailing tested exactly this. Aguirre and colleagues ran a field study on Facebook plus three experiments. When a firm collected data openly, more personalized ads drew higher click-through intentions¹. When it collected the same data covertly, the pattern flipped. The authors trace the flip to feelings of vulnerability, and in the field study click-through rates dropped sharply once people realized their information had been collected without their consent¹.

So the same ad can land warm or cold depending on how the data was gathered. The authors also report an offset: trust-building strategies, such as informational trust cues¹, softened the damage. It is a 2015 study of Facebook ads, so treat it as a pattern to test in your own audience rather than a law.

Picture a hotel hallway. One guest hangs a card on the handle that says “extra pillow,” and housekeeping brings one to the door. That feels like good service, because the guest asked. Another guest comes back to find their suitcase open and their clothes pulled out. The hotel is the same and so is the staff, but this guest hangs the Do Not Disturb sign and keeps the door shut. In marketing, the request card is the data a customer gave you. The searched suitcase is anything gathered somewhere they never agreed to, and the study above is the measurement of that sign going up.

What does AI add to the picture?

AI doesn’t invent new data. It makes more of it usable: a model can guess things a customer never typed in, and it can write a different message for every person. Both raise the stakes on the first question, because a guess about someone is information they never handed over. When a vendor says its tool personalizes, ask which signals it uses and where each one came from.

Which lines do platforms and regulators already draw?

The ad platforms have fenced off the most sensitive ground. Meta removed Detailed Targeting options tied to health, race or ethnicity, political affiliation, religion and sexual orientation, effective January 19, 2022, after experts raised concerns that such options could be used in ways that lead to negative experiences for people in underrepresented groups². Google’s policy lists the same kinds of subjects as sensitive interest categories, and says audiences an advertiser builds from its own data cannot be used if promoting in sensitive interest categories³.

The law adds two lines for anyone reaching people in the European Union (EU). Under the General Data Protection Regulation (GDPR), when personal data is used for direct marketing, the person has the right to object at any time⁴, and that covers profiling tied to the marketing. Once they object, the data shall no longer be processed for such purposes⁴. Separately, the EU AI Act bans AI that uses purposefully manipulative or deceptive techniques⁵ or exploits a person’s vulnerabilities, such as age, disability or social or economic situation, when it distorts behavior and is reasonably likely to cause significant harm. Those bans have applied since February 2, 2025. Where a given campaign falls is a question for the client’s counsel.

What's Real?

• •How the data was gathered changes the response. In one peer-reviewed study, open collection raised click-through intentions¹ for more personalized ads, and covert collection reversed it.
• •The big platforms already restrict sensitive targeting. Meta removed the options in 2022, and Google limits the audiences an advertiser can build for sensitive interest categories³.
• •In the EU, people can say stop. The GDPR gives them the right to object at any time⁴ to direct marketing, profiling included.
• •The hard evidence is thinner than the confidence. The controlled study here dates from 2015, and none of the sources measures AI-written personalization directly, so test it with your own audience.

The “so what” — for anyone serving tech clients

• •Label every personalized element by source: told to us, observed, or inferred. Put that label in the brief. Inferred signals get a second look before they go live.
• •Keep sensitive categories out of both targeting and inference, whatever a platform currently allows. The platforms already restrict them, and a rule can tighten faster than a campaign plan changes.
• •Tell people why they are seeing it. A short “why this ad” cue is cheap, and the study above found trust cues softened the backlash.
• •Measure the wince. Track opt-outs, unsubscribes and complaints next to clicks, and compare against a lightly personalized version. A lift in clicks that comes with a rise in opt-outs is a cost, and the client should see it.
• •Keep a person on the exceptions. Humans review any new data source and any campaign aimed at a vulnerable group, the same last-mile review as brand safety⁶.