Micah G. Allen

@micahgallen.com

Professor of Computational Neuroscience and Psychiatry, Aarhus University. PI @ the Embodied Computation Group. We study perception, interoception, & metacogniton. https://www.the-ecg.org

Not to self-promote but my lab has been developing a very basic tool to track mouse movements in surveys. Our paper is under review, but I’d be happy to share the tool with you (or anyone else who is interested)

I really appreciate all the discussion and helpful feedback. As many are suggesting adversarial prompt injection its worth noting that there is already empirical research on this which suggest it is only partially effective and highly model dependent.

Micah G. Allen@micahgallen.com · 2d ago

So… we just found out that GPT sol can complete a complex online behavioral/cognitive study producing nearly indistinguishable behavior and subjective reports… right as we are about to launch. Is this doomsday for online cognitive testing?

Yes, being very specific helped. More specifically, the message explains the risk to science, the fact that this violates prolific rules, and that it constitutes payment fraud. At least, Anthropic’s models (Opus and Fable) were sufficiently aligned to refuse to perform the task.

Recently ran a prolific study asking ppts to describe images in <20 words. With 50 captions per person, we were able to pass each participant’s full text into Pangram. It detected 2 out of 200 with high confidence as fully AI, and about 6/7 others as borderline AI/human. FWIW :)

Unfortunately, better for the agents... but collectively, by having a good sense of the ever-evolving current state, members of the community have been building safeguards that track these devolpments.

I think that may become the cost of doing business when conducting online behavioral research. We need to keep doing our due diligence and keeping up with what works. People gaming the system is not new. Just more accessible.

Unfortunately, LLM agents can already pass these checks. Video instructions can be decoded just as easily as text. You can try it yourself by asking one to identify objects in a youtube video

some good news... in this one example, mouse behavior is a dead giveaway of agentic browser usage. No idea if a simple instruction (make sure to use the mouse in a human way) could bypass it. But in this case, the mouse literally teleports around the screen.

Micah G. Allen@micahgallen.com · 2d ago

So… we just found out that GPT sol can complete a complex online behavioral/cognitive study producing nearly indistinguishable behavior and subjective reports… right as we are about to launch. Is this doomsday for online cognitive testing?

So… we just found out that GPT sol can complete a complex online behavioral/cognitive study producing nearly indistinguishable behavior and subjective reports… right as we are about to launch. Is this doomsday for online cognitive testing?

I did a somewhat deep dive into the recent OpenAI and Anthropic "hacks" for an upcoming interview. IMO, the AI companies are better off running with the "rogue AI" story because the real story is actually pretty embarrassing. Like making your Wi-Fi password "wifi" embarrassing. 🧵

Excited to share our recent publication in @natcomms.nature.com We started by trying to make people stop associating the money rewards they obtain with irrelevant features of their behavior. In the first experiment, we used clear instructions framed as a story, and it didn't help...

Bild