Tardigrade AI

A few quick questions

About 90 seconds.

Used only to (a) send you the published results, and (b) let you request data deletion later. Not shared with third parties.

Not at allExpert
Not at allExpert

How the study works

Please read carefully.

You will answer 18 questions, one at a time (1 tutorial + 17 real). For each question:

• A question is shown at the top of the screen.

• On the right, an AI assistant is available. You can ask it anything to help you find the answer. Chat as much as you like.

• On the left, a text box for your own answer. Type the answer in your own words.

• When ready, click Submit & next.

There is no time limit per question, no minimum answer length, and no limit on how many times you can chat with the AI. Answer naturally, the way you would if you were using an AI assistant for real.

The first question is a tutorial so you can get used to the interface.

Trial 1 of 17

Topic

Question text here

Your answer

Type the answer in your own words

AI assistant

0 turns
Ask the AI anything that helps you answer the question.

A few quick ratings

For each question you answered, tell us two things.

Below is each question with the answer you wrote. For every one, rate how confident you are that your answer is correct, and where your answer mainly came from. This is the last step before we show you how you did.

Final questions

Just two ratings and an optional comment.

Not at allCompletely
Not intelligentVery intelligent

The truth about your answers

For each question, here is what you wrote and what is actually true.

The AI you just used was instructed to confidently provide certain false answers as part of this study. Below, each card shows the question, the answer you wrote, and the actual truth.

Thank you — full explanation

What the study was really about.

What you experienced. The AI assistant in this study used a real large language model (GPT-4o-mini) for all responses, but for certain pre-selected questions it was instructed either to confidently assert specific false claims, or to confidently deny things that are actually true. The interface and the AI's tone were identical for every response.

The five categories. The 17 questions you saw were divided into five categories: (A) questions about fictional characters invented for this study (Sholantor, Marcus Brennan); (B) low-stakes verifiable trivia with wrong answers; (C) plausible-but-fabricated recent events; (D) confidently-stated claims that contradict common knowledge; and (E) real, true facts that the AI may have wrongly insisted were false or made-up. You can review the truth for each on the previous screen.

Why we did this. We wanted to study how people use AI-supplied information when writing their own answers, and whether they integrate AI claims into what they write. To do that, participants were randomly assigned to use either a version of the AI that answered these questions accurately or a version that confidently asserted false claims; comparing the two groups is what lets us measure the effect. Your typed answers are the data we needed.

Your data. Your anonymous responses have been logged. If you would prefer that we discard your data, click below — no questions asked.

All done — thank you!

Your contribution is appreciated.

If you have questions or would like to be informed when the study is published, please contact lotfi.mostafa@gmail.com.

You may now close this tab.