Say you’re filling in an optional shopping questionnaire about your interests. You want more useful suggestions, but you wonder whether each answer makes you easier to persuade.
There are several questions inside that worry. What are you telling someone? What else could they work out from your answers? Would using that knowledge change a choice you would otherwise make?
Those are different things. Predicting that you’re likely to buy something doesn’t show that a message made you buy it. And a message can work without its sender knowing your personality.
This module is for working adults thinking about what to share in everyday shopping and political settings. Sharing less personal data doesn’t promise protection from influence.
What you’ll get from this module
In this module
By the end of this module, you’ll be able to tell being profiled apart from being persuaded. You’ll see what someone might guess about you from what you share. You’ll also see why that doesn’t give them control over your choices. Research on online traces, political messages, paid search ads and a campaign aimed at places will explain the difference. You’ll link those findings to everyday sharing and choose one situation. You’ll finish with a choice, or a clear question you still need answered.
Ask Jing
Become a Jingius.
What changes between knowing you and influencing you?
Prediction study: what you reveal can help someone guess something else.
Michal Kosinski, David Stillwell and Thore Graepel’s 2013 study used the Facebook Likes of U.S. volunteers to predict things about them. The researchers checked their predictions against people’s profile details, their answers to questionnaires and their test results.
Patterns of Likes predicted personal details, such as personality. Some were predicted better than others. To test the models, the researchers made predictions about people whose data hadn’t been used to build them. This checked that the predictions worked on new people too.
What people gave and what could be worked out from it weren’t always the same thing. An interest could help guess something a person hadn’t said outright. So leaving one answer blank doesn’t always stop someone guessing it from other data.
Why it matters: This older study tested prediction, not messages designed to change people’s choices. It leaves a further question: does choosing a message using that knowledge make it more persuasive?
Read the study of predictions from digital records.
Survey experiments with random groups: a tailored message needs a fair comparison.
Ben M. Tappin and colleagues looked at that question in research published in 2023. They used earlier tests to pick political messages shaped for different people. They also picked one message expected to work best for everyone.
They then recruited a new group of adults in the United States for online experiments. Each person was given one of the messaging approaches at random. The researchers measured views on policies straight afterwards, not votes in an election.
For the U.S. Citizenship Act, tailoring did better than the shared message chosen by testing, by a margin unlikely to be chance. For a universal basic income, tailoring looked a little better, but the evidence didn’t clearly show that it was.
Another part of the experiment picked between different issues. There, tailoring didn’t clearly beat the best shared message. These comparisons used the same group of people, not three separate groups.
The shared-message approach is worth understanding. The researchers used average reactions in earlier tests to pick one appeal, then showed it to everyone. They learned from people without matching each person to a different message.
Tailoring sometimes added something, and sometimes its advantage wasn’t shown. The unclear comparisons don’t prove the approaches worked equally well. They show why the word “personalised” alone doesn’t tell you how much extra influence a message has.
Why it matters: The comparison matters, too. Beating a badly chosen message isn’t the same as beating one already picked through testing.
Read the political microtargeting experiments.
A field experiment by area: a purchase after an ad can make the ad look better than it was.
Thomas Blake, Chris Nosko and Steven Tadelis looked at paid search ads at eBay in research published in 2015. One test stopped some search ads in chosen areas of the United States during 2012.
These ads came up for searches that didn’t include the eBay name. The researchers compared how sales changed in areas where those ads stopped with areas where they carried on. Areas were picked at random and then paired up; individual shoppers weren’t picked at random.
The paper estimated that the ads cost more than the extra money they brought in. That was on average, and in the short run. The sums used the sales changes from the experiment and published figures for revenue and spending.
Yet new and occasional buyers did respond to the ads. The average didn’t mean nobody was influenced. Frequent customers accounted for much of the spending, with no gain that could be detected to match it.
Someone already planning to buy could click an ad and then buy. Giving the ad credit for that purchase could make it look better than it was.
That’s the difference between finding likely buyers and creating extra purchases. A purchase after an ad tells you which came first. Comparing with areas where the ads stopped helped the researchers estimate what the ads added.
Why it matters: This result is about eBay’s ads, customers and the time studied. It doesn’t show that all advertising loses money.
Read the paid-search field experiment.
A campaign tested on random areas: a sender can influence groups without a profile of each person.
Emily Breza and colleagues studied videos from doctors and nurses over the Thanksgiving and Christmas holidays in the United States in 2020. Their 2021 paper reports a campaign asking people to stay home during the pandemic.
The trial picked counties at random to get more or less of the campaign. It also picked ZIP codes, or postal areas, at random to get ads or to go without. A bigger campaign meant that more postal areas in a county got ads.
The researchers looked at travel and reported infections using data combined across everyone in an area. A bigger campaign cut the distance people travelled before the holidays. The share of people leaving home on the holidays didn’t change by a clear margin.
Areas that got the ads also had fewer reported infections than the others in the two weeks after each holiday. These results are about the whole campaign. They don’t separate people who were persuaded directly from other effects. Shared messages, for example, could reach people who never saw an ad.
The research team didn’t recruit individuals or use data about single people. But Facebook’s system for showing ads still estimated how likely each user was to engage with an ad. And the combined travel data came from each person’s location records.
The sender didn’t need a separate personality profile for every person to run a campaign that worked. That doesn’t mean the platform used no personal data.
Why it matters: This pandemic campaign shows one way messages aimed broadly can change outcomes. It explains why the sender not holding a profile of you doesn’t make you immune to influence.
Read the location-based campaign trial.
What this means for you
You can decide whether to share a particular detail without knowing exactly how easily someone could persuade you. The useful question is what sharing makes possible and whether you’re comfortable with that trade.
Identify the information and what it may enable
Start with the detail you’re being asked for. Separate what your answer says outright from anything else someone might guess by using it with other data.
In the made-up questionnaire, naming an interest gives away that interest directly. The Facebook study shows how recorded interests can also help predict other things. It doesn’t tell us what this imagined questionnaire actually works out.
If you know how the data will be used, bring that into your choice. If you don’t, name what’s missing. A possible use isn’t the same as something you’ve confirmed they do.
Ask what changed, and compared with what
When you hear that targeting worked, ask what changed. Predicting a trait, getting a click, changing someone’s view straight away and creating an extra purchase are different results.
Then look at the comparison. The political experiments compared tailoring with a shared message chosen through testing. The eBay experiment compared sales changes where ads stopped with changes where they carried on. Both asked what the ads or messages added beyond another option.
If a claim counts only purchases after an ad, you still don’t know how many the ad caused. If it includes a fair comparison, keep the setting and the result in mind. An average effect in a study still doesn’t show exactly how one ad changed your own choice.
Weigh the purpose against what sharing reveals
Ask what you want from sharing the detail and what you’re willing to let someone learn about you. Profiling means using what is known about you to guess your traits, interests or other details. Useful tailoring may depend on data you would otherwise keep to yourself.
In the made-up questionnaire, you might value suggestions based on the interests you gave. The benefit may be worth it, and you may be at ease with what sharing makes possible. Then answering can be a choice you make on purpose. If you aren’t comfortable, leaving an optional answer blank can also be deliberate.
You need a real choice for this judgement. Where you can’t choose whether to give the data, write down that limit instead of treating the situation as voluntary.
If an important fact about the purpose or use is missing, name it before you treat your choice as settled. A clear question can be more useful than forcing a yes or no.
Name what your decision controls
Holding back one answer controls that one piece. It doesn’t wipe out other signals that are out there, or what has been learned from groups of people. A message that works broadly can still reach you.
You can decide that sharing a detail isn’t worth it. That doesn’t mean holding it back makes you impossible to persuade. You can also accept useful tailoring without assuming every tailored message controls what you do.
The research doesn’t measure how much this choice about sharing protects one person from persuasion. Use the exercise to explain your purpose, your choice and the limits of what you know.
How about your daily life?
Where are you deciding whether to share something about your interests or preferences? You might be looking at an optional shopping questionnaire or a request for data alongside a political message.
Write down a situation and what you want to understand. One useful idea is enough.
| Situation | What I want to understand |
|---|---|
| ____________________ | ____________________ |
| ____________________ | ____________________ |
My chosen situation: ____________________
Choose an everyday sharing choice where you have a real choice. Carry it into the exercise below. You can change your choice if it doesn’t fit.
Make one deliberate sharing decision
Start with your chosen situation. If you came straight here, choose one everyday shopping or political request for data that you can really accept or turn down.
This exercise is about ordinary choices adults make about sharing. It doesn’t give safety advice for children or controlling relationships. Nor does it cover serious medical choices, or close targeting of one named person.
- Name what you’re being asked to share. Write the detail being asked for and your reason for thinking about sharing it. If you have no real choice, write down that limit and stop this exercise, which is about sharing by choice.
Information and purpose: ____________________ - Separate what you know from what someone might infer. Describe the uses of your answer that you know about. If someone might guess something else from it, mark that as uncertain unless you have data to back it up.
Known use, possible inference and missing information: ____________________ - Decide whether the trade-off is worthwhile. Weigh your purpose against what you’re willing to let someone learn about you. Include any useful tailoring you might lose by holding the answer back. If an important fact is missing, write the question you still need answered.
Share, withhold or leave unresolved—and why: ____________________ - State what your decision controls. Name the data you can choose to give or hold back. Other signals, what has been learned from groups and broad messages can still help influence you. Don’t turn your choice into a promise of protection.
What this decision controls and what it leaves uncertain: ____________________
Ask Jing
Become a Jingius.
Your next step
You’re finished when you have a reasoned choice about sharing, or a clear question you still need answered. Add an honest note of its limits.
A note on the evidence
These studies answer different questions. These results can’t be added up into one score for how much you are influenced. Predictions, quick changes of view, sales and campaign results are different things.
The targeting experiments include both clear and unclear comparisons. The field studies are about particular settings. The campaign also includes how the platform showed ads, and effects that may have reached people who saw none. None measures how much holding back one detail protects you from persuasion.
How we prepared this module
This module links a study of predictions from what people do online, tests of political messages, and two field studies of ads. A made-up questionnaire shows a trade-off in sharing. The final exercise helps you state a choice and its limits.
Sources
- Michal Kosinski, David Stillwell, and Thore Graepel. (2013-03-11 online; 2013-04-09 issue). Private traits and attributes are predictable from digital records of human behavior. Proceedings of the National Academy of Sciences, 110(15), 5802–5805. Published article. Fig. 1 and Results, pp.5803–5804.
- Ben M. Tappin, Chloe Wittenberg, Luke B. Hewitt, Adam J. Berinsky, and David G. Rand. (2023-06-12). Quantifying the potential persuasive returns to political microtargeting. Proceedings of the National Academy of Sciences, 120(25), e2216261120. Published article. Alternative messaging strategies, p.3; Study 1 results, p.4; Study 2, p.6; Methods, p.9.
- Thomas Blake, Chris Nosko, and Steven Tadelis. (2015-01). Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment. Econometrica, 83(1), 155–174. Accepted manuscript, 2014-08-12; abstract, §3.1 printed p.8, Introduction printed p.5 and §4 printed p.16.
- Emily Breza, Fatima Cody Stanford, Marcella Alsan, Burak Alsan, Abhijit Banerjee, Arun G. Chandrasekhar, Sarah Eichmeyer, Traci Glushko, Paul Goldsmith-Pinkham, Kelly Holland, Emily Hoppe, Mohit Karnani, Sarah Liegl, Tristan Loisel, Lucy Ogbu-Nwobodo, Benjamin A. Olken, Carlos Torres, Pierre-Luc Vautrey, Erica T. Warner, Susan Wootton, and Esther Duflo. (2021-08-19). Effects of a large-scale social media advertising campaign on holiday travel and COVID-19 infections: a cluster randomized controlled trial. Nature Medicine, 27, 1622–1628. Published article. Abstract; Results; Methods, “Trial design, eligibility, randomization and recruitment” and “Details about Facebook ad campaigns.”
