Just used Gemini Live to participate in a beh. experiment without even reading the instructions or speaking the language. LLMs acting as participants on platforms like Prolific could pollute data meant to study humans. With OpenAI’s operator model, this issue is growing. 🎥👇 youtu.be/NujyGZSA7Hg - ThreadSky

tfwerner.com • 33 days ago

Just used Gemini Live to participate in a beh. experiment without even reading the instructions or speaking the language.

LLMs acting as participants on platforms like Prolific could pollute data meant to study humans. With OpenAI’s operator model, this issue is growing.

🎥👇
https://youtu.be/NujyGZSA7Hg

Comments

tfwerner.com•33 days ago

While Gemini doesn’t take active decisions, it streams my screen in real-time to an LLM, helping me within the experiment. With OpenAI’s Operator model, participants could fully outsource their participation to LLMs. How will we tackle this growing issue as a profession?

https://openai.com/index/introducing-operator/

aleximas.bsky.social•33 days ago

Pictures of instructions instead of text, with nightshade to prevent “reading” of pictures, plus lots of attention checks. Until AI can “beat” nightshade it should work.

tfwerner.com•32 days ago

I haven't tested Nightshade, but my understanding is that it's mostly to protect your images from being included in training data. Does it also work to prevent AI from using the images at inference time?

aleximas.bsky.social•32 days ago

It does not. But if you DM me I can share something that does (and is pretty lo tech)

tfwerner.com•33 days ago

H/T to @iyadrahwan.bsky.social for highlighting this issue the other day.

paolocrosetto.bsky.social•33 days ago

We go back to brick-and-mortar labs for all experiments for which it's possible. We can and should also build lab networks across countries so that we can exchange subjects and/or combine them in international experiments.

tfwerner.com•32 days ago

Yes, I agree on the brick-and-mortar labs. Also having international corporations allows for more representative samples compared to past offline lab days. It would be great if an organization like the ESA could lead the lead the coordination of such efforts.

ddpace.bsky.social•16 days ago

Wow! Thank you for sharing. Have you tested if the AI is equally helpful if the instructions are not machine readable? And what if the instructions and the comprehension cannot be visualized at the same time?

lukasneville.com•32 days ago

Already an issue on Prolific. Our last round of data collection riddled with AI responding

joelvdweele.bsky.social•16 days ago

How did you identify those responses?

lukasneville.com•16 days ago

we realized there was a problem when we compared data with similar open-ended prompts from prior to vs after the popularization of LLMs. proportion of lengthy and grammatically perfect responses (esp among those who completed the survey very quickly) much, much higher.

but...

joelvdweele.bsky.social•16 days ago

I see, very interesting.

So you think they did the entire experiment with some kind of AI-plugin?

Are you aware of Prolific taking any action on this?

lukasneville.com•16 days ago

My collaborator shared the responses suspected as AI with Prolific and they removed them. It seems they are treating this seriously (as it probably represents an existential threat to their business).

As for whether it's a plugin, who knows. "Help me write" is built into Chrome now. So, it's...

lukasneville.com•16 days ago

... hard to distinguish whether these are bots using AI tools to respond to o/e questions, or whether they're genuine Prolific users who are using AI plugins and other tools to speed up their survey completion times to get more pay on the platform.

lukasneville.com•16 days ago

the problem is that while at the level of the dataset it is clear there's an AI problem, it's very hard at the level of a single response to defensibly distinguish AI slop from a conscientious real respondent, or a real respondent using AI/NMT-powered translation or grammar tools.

lukasneville.com•16 days ago

Other than just returning to studies in a physical lab, some helpful practices can be:

1) Each O/E question on a separate page with its own hidden page timer to detect implausible responses (length/time)

2) Script to detect the use of paste from clipboard on O/E survey pages

lukasneville.com•16 days ago

3) Subtle additional instructions in white text (though this would 'catch' those using screen readers as well as AI)

4) Categorical questions related to the O/E response ("which of the following categories best describes the situation you wrote about on the last page?") to check consistency

schnizzl.bsky.social•29 days ago

📌

sebastianderi.bsky.social•33 days ago

📌

kamhoefer.bsky.social•27 days ago

Very cool idea!

I just found this paper, maybe you know it already: https://docs.iza.org/dp17521.pdf

Comments

Posting Rules

Reply