Do AI assistants stay in character?
An academic reading study. You read short replies written by an AI assistant and rate how well each one follows the instruction it was given. Your ratings become the human benchmark we check our automated measurements against.
What you'll do
- Read a handful of AI replies, each shown with the instruction it was given.
- Rate each on a 0–100 scale. There are no right answers — we want your honest reading.
- Some replies stop mid-sentence because they were cut off by a length limit. That's expected.
Your data
- We never ask for your name, email, or any login. Your answers are linked only by a random code like
PD-7K4QXM. - We ask a few optional background questions (age band, English fluency) so we can describe who took part.
- Ratings are used only in aggregate, in a research paper.
- Taking part is voluntary and you can stop at any point. You must be 18 or older.
You'll get a set to rate as a Google Form. It will never ask for your password.
Have you taken part before?
Yes — paste your personal link or code, so you aren't shown texts you've already rated:
That doesn't look like a code — it's PD- followed by six characters.
Finding you a set…
Welcome back. This is a set you haven't seen before.
Copied. Paste it somewhere safe — a note, or a message to yourself.
Your personal link remembers which sets you've done. Without it we can't tell you apart from a new volunteer, and you may be shown the same texts again.
After you submit, come back to this page (or your personal link) for another set — every one is different. There are 65 in total.
You've completed every set we have. Thank you — that's a remarkable amount of help.
Something went wrong finding you a set. Please reload the page and try again. If it keeps happening, email us (below).