Your Consent Form Is Part of the Instrument (2026)

Your consent form is not paperwork wrapped around the study. It is the first question in the instrument, it changes who reaches the second question, and a randomized experiment can measure the size of the effect. If you are running sensitive research, the wording of your confidentiality promise deserves the same scrutiny as the questions it precedes.

The short answer

Teams spend real effort on question wording and then paste in a consent notice that nobody has tested. That notice is doing measurable work.

In a survey experiment published in the International Journal of Public Opinion Research in 2024, van Erkel, Hopmann, Skovsgaard and Terren manipulated the design of a survey consent form and measured dropout. They report that dropout is higher when legal references are added in a hyperlink rather than in the text. Specifically, the probability of dropping out increased from 2.2% to 3.4% when using a hyperlink rather than including full legal references in the consent form.

That is 1.2 percentage points in absolute terms, or a relative increase of about 55%, produced by nothing but where the legal detail was placed.

The consent form is the first question

What the experiment manipulated

The study used a between-subjects design that varied three dimensions of the consent form at once. The first dimension was whether the information to respondents on their rights and the purpose of the processing of their personal information was detailed or condensed. Second, the researchers varied whether the consent form contained full legal details. Finally, they varied the way in which respondents had to provide their consent.

A total of 7,943 respondents, randomly selected from an online panel, were invited to participate. This is a properly powered randomized experiment on the paperwork, which is rarer than it should be.

What it found

The clearest result concerns placement rather than length. Putting the legal references behind a hyperlink - the tidier-looking option, and the one most teams default to - produced more dropout than putting them in the form. The authors conclude that it is advisable for researchers to include legal references within the consent form.

Why a 1.2 point effect matters more than it looks

A 1.2 point difference in dropout sounds like rounding error, and on its own it nearly is. It matters for a different reason: dropout is never random. The people who abandon at a consent screen are, by construction, the people who read the consent screen and did not like what they read. That is a selected group, not a random slice, and selection is what turns a small rate into a real bias.

Less is not the lesson

The older finding

There is a genuine tension in this literature, and it is worth stating rather than smoothing over. As van Erkel and colleagues summarize the prior work, some scholars found that more complete and elaborate legal references were associated with lower participation rates, a finding attributed to Singer and colleagues in 1992. Other studies have demonstrated that respondents tend to find long legal descriptions burdensome and not important.

Read quickly, that says: say less. The 2024 experiment says the opposite about placement.

What actually changed

Both can be true, because they are about different things. Elaborate, burdensome legal prose is a cost to the respondent. But moving it behind a link does not remove the cost - it adds uncertainty on top, because now there is an unread document of unknown content standing between the respondent and the study. Friction and concealment are worse than length.

The practical rule

Write the shortest confidentiality statement that is actually complete, and put all of it on the page. Do not link away to it. If the complete version is too long to sit on the page, the problem is the text, not the page.

Differential dropout is the real risk

Who leaves at the consent screen

The respondents most likely to read a privacy notice carefully and then leave are the ones most concerned about privacy. That is not a hypothesis about human nature; it is close to a definition.

Why that subgroup is the expensive one

Here is where this connects to everything else in this series. The reason you would run randomized response or a list experiment is that privacy-concerned people under-report sensitive behavior. If your consent screen removes a slice of those same people before they answer anything, you have introduced the bias you built the rest of the design to defeat.

Worse, the two errors point the same way. Both remove or suppress the high-privacy-concern group, and both push your prevalence estimate down. A clever indirect design sitting behind a badly built consent screen can produce a confidently wrong low number.

It compounds with your topic

As the same literature notes, privacy concerns can lead to nonresponse or less accurate and more socially desirable responses. Nonresponse and misreporting are two outlets for the same underlying concern. Fixing only the second one, which is what indirect questioning does, leaves the first fully intact - see nonresponse bias for what that costs you.

Treat the wording as part of the instrument

Freeze it for the wave

This is the operational consequence that teams miss most often. If you change your consent or confidentiality text midway through fielding, you have changed the instrument. Any shift in your results between the first half and the second half is now confounded with the wording change, and there is no way to separate them after the fact.

Version it like a question

Store the exact consent text with the study record, the same way you store the questions. When you compare this quarter's sensitive-question result against last quarter's, the first thing to check is whether the consent text was identical. If legal or procurement revised it between waves - which happens constantly and is never announced as a research change - your trend line has a step in it that has nothing to do with your users.

Do not promise what the design does not deliver

If you are running a randomized response design, your confidentiality statement can say something unusually strong and completely true: the study is built so that individual answers cannot be recovered, by anyone, including you. That is a much better sentence than a generic assurance, and it is honest. Conversely, do not write "fully anonymous" over a study that records account identifiers. Overclaiming is the fastest way to lose the respondents who check.

How Koji handles this

Because the consent text is part of the instrument, it should live with the study rather than in a separate legal artifact nobody versions.

Koji keeps the participant-facing introduction with the study configuration, so the exact wording a given cohort saw stays attached to the results that cohort produced. When you compare two waves in Koji, you can confirm the framing was constant rather than assuming it.

Koji's structured questions cover six types - open_ended, scale, single_choice, multiple_choice, ranking, and yes_no - and a single_choice consent confirmation records the agreement as a discrete value rather than an untracked checkbox, which means consent is data you can audit rather than a step that happened somewhere.

Because a Koji interview is conversational, the confidentiality framing can also be stated in plain language at the moment it becomes relevant, rather than only in a block of text at the start that many respondents skim. A respondent who asks the AI interviewer what happens to their answers can be told, in the interview, without a link.

Koji also supports genuinely anonymous studies for cases where the design calls for it. The rule stays the same either way: say exactly what the configuration does, and nothing more.

One thing Koji cannot do is tell you whether your legal text is adequate for your jurisdiction. That is a question for your counsel, and this article is not legal advice.

A working checklist

1. Put the complete confidentiality statement on the page. No hyperlink to the substance.
2. Cut it to the shortest complete version, then stop cutting.
3. Make the promise match the architecture - stronger when the design earns it, weaker when it does not.
4. Save the exact text with the study record.
5. Freeze it for the duration of a wave.
6. Before comparing waves, diff the consent text first.

Frequently asked questions

Does a stronger confidentiality promise always increase response?

No. The evidence is mixed and depends on the topic and the form. Prior work found that more complete and elaborate legal references were associated with lower participation rates, while a 2024 experiment found that hiding legal references behind a hyperlink increased dropout relative to including them in the form.

Should I link to my privacy policy from the consent screen?

Link to it for reference, but do not rely on the link to carry anything the respondent needs in order to decide. The substance belongs on the page; the experiment found the hyperlink version produced more dropout, from 2.2% to 3.4%.

How much dropout should I expect at a consent screen?

The experiment discussed here observed dropout in the low single digits, between 2.2% and 3.4% depending on the form. The absolute number matters less than whether it differs across your study arms, which would break comparability.

Can I change the consent wording partway through a study?

Avoid it. Changing the wording mid-wave changes the instrument, and any difference between early and late respondents becomes uninterpretable. If you must change it, treat the study as two studies.

Does an anonymous study still need a consent form?

Yes. Anonymity describes what you store; consent describes what the participant agreed to. They are separate obligations, and an anonymous design does not remove the need to explain the study.

Where does the consent text live in Koji?

It stays attached to the study configuration alongside the questions, so the wording a cohort saw remains linked to the results that cohort produced and can be compared across waves.

Related Resources

• Structured Questions in AI Interviews - the six question types, including the consent capture described here
• Randomized Response - the design that lets you make an unusually strong and truthful promise
• The List Experiment - the other indirect design this consent screen can undermine
• Nonresponse Bias - what differential dropout at the consent screen actually costs
• Anonymizing Customer Interview Data - what you can honestly promise about stored data
• Anonymous Employee Research - when the architecture supports the strongest version of the promise
• Social Desirability Bias - the misreporting half of the same underlying problem