Tobias Werner

Considering outsourcing my ARC Raiders inventory management to Claude Cowork

bsky.app/profile/tfwe...

AI may upend online studies critical to social science Sophisticated bots risk contaminating surveys, games, and other approaches designed to shed light on human behavior

LLMs and agents are reshaping the landscape of social science research.

Our preprint on "LLM Pollution" is featured in @ScienceMagazine's new piece on how Al is upending online studies.

www.science.org/content/arti...

Our dept is hiring an associate professor in microeconomics / competition theory.

It is a wonderful department - do get in touch.

The job ad is in German but we teach more and more in English, so we‘re flexible.

www.hu-berlin.de/en/universit...

The Streuselkuchen is alright in this club! ;)

Huge shout-out to our PhD students presenting their research at the @unisouthampton.bsky.social ’s Doctoral Research Poster Showcase this week! 🎓 #PhDLife #ResearchExcellence #PhD #DoctoralResearch #Southampton #AcademicShowcase

Our department hosted a great seminar this week by my colleague @tfwerner.com titled "Experimental Evidence That Conversational AI Can Steer Consumer Behavior Without Detection". Eye-opening insights into conversational AI and consumer choices! #AIResearch #BehavioralEconomics #AcademicResearch

Delegate Pricing Decisions to an Algorithm? Experimental Evidence We analyze the delegation of pricing by participants, representing firms, to a collusive, self-learning algorithm in a repeated Bertrand experiment. In the baseline treatment, participants set prices ...

💡 Takeaway: Humans-in-the-loop are key.

When people are involved, even a ``collusive'' AI can lead to more competition and lower prices .

Full paper: arxiv.org/abs/2510.27636

🔧 Why? (2/2): In RECOMMENDATION, adoption is high, but adherence is low.

Participants frequently override the AI's advice, resulting in longer price wars than in the baseline.

🔧 Why? (1/2): In OUTSOURCING, non-adopters often use a cyclic-deviation strategy (repeatedly undercutting the AI).

This is not profitable, but it punishes adopters & discourages delegation. Our evidence suggests that this is driven by spite towards AI, not myopia.

📉 Result 2 (Prices): Pro-competitive effects of AI!

We thought the collusive AI would raise prices.

However, by the end, prices are significantly lower in both AI treatments than in the human-only BASELINE.

📈 Result 1 (Adoption): Participants delegate a lot.

But control matters. Adoption is significantly higher when participants can override the AI (RECOMMENDATION).

We ran a lab experiment with 3 conditions:
1️⃣ BASELINE: Humans only.
2️⃣ OUTSOURCING: Full delegation to a collusive AI.
3️⃣ RECOMMENDATION: AI advice, but humans can override it .

Most work on AI collusion simulates AI vs. AI and simply assumes firms use them. This overlooks the (human) decision to adopt.

So we ask: Will firms delegate to a collusive AI? And how does this strategic choice change market outcomes?

Delegate Pricing Decisions to an Algorithm? Experimental Evidence We analyze the delegation of pricing by participants, representing firms, to a collusive, self-learning algorithm in a repeated Bertrand experiment. In the baseline treatment, participants set prices ...

📣 New paper! "Delegate Pricing Decisions to an Algorithm? Experimental Evidence" with @normann.bsky.social, Nina Rulié, & @ostypa.bsky.social

We study what happens when humans delegate to a collusive algorithm, and AI pricing adoption becomes a strategic choice!

arXiv: arxiv.org/abs/2510.27636

The fact that fries/chips are just normally sold at Chinese restaurants in Ireland should be studied. But still: spice bag is the best fast food invention since ???

And yes: I thought before leaving Berlin I had to go full Berlin and get a bleached buzzcut

First day in the office!

Excited to join the University of Southampton as an Assistant Professor/Lecturer

I like the new robot 'personality' feature of ChatGPT. Although it's probably just a tweak to a system message, it's much more to the point and less wordy.

Just weird that, out of the blue, it sometimes ends with things like, 'I am synthesized and not embodied.'

We have our first games ever in Ireland next month. The event will be at the University of Galway as part of CERIS’s Annual Workshop on September 26th. Virtual participation is possible and coauthorship to a meta paper is granted.

Register here: www.surveymonkey.ca/r/Replicatio...

With the release of Gemini Live / OpenAls Operator, we started working o
perspective piece on how to tackle the growing threat these systems pose to online behavioural research and its validity.

We hope this helps move the conversation forward.

Paper: arxiv.org/abs/2508.01390

Absolutely! That's part of the solution, but online studies also offer key advantages, like more representative samples + larger N. Hopefully, we'll see more cross-lab collaborations.

If you run studies online or use platforms like Prolific, MTurk, or CloudResearch, this impacts you!

We hope this paper encourages discussion, collaboration, and improved safeguards.

🔗 arxiv.org/abs/2508.01390

At the end of the day, this isn't just a researcher's problem.

Platforms that advertise "100% human" samples must be held accountable.

Researchers need tools and transparency to protect their work, and in some cases, it may be time to go back to the physical lab!

We propose a layered defence strategy, but let's be clear: this is already an arms race.

Some measures:
- reCAPTCHA and Cloudflare
- Multimodal instructions (images, audio)
- Input restrictions (no paste, voice answers)
- Behavioural logging
- Platform enforcement

LLM Pollution scrambles your data:
🧠 Dampens variance
📈 Inflates effects
📊 Mocks WEIRD norms
🔍 Obscures who your participants really are
… all while staying nearly undetectable!

In a recent Prolific pilot, 45% of participants copied/pasted open-ended items or showed signs of AI-generated content, with responses starting like "As an artificial entity…"

This is contamination, not noise!

We map three invasion paths for LLMs:

1. Partial Mediation: LLM assists with rephrasing or answering.
2. Full Delegation: Agents like OpenAI Operator handle everything.
3. Spillover: Humans act differently due to bot expectations.

(See figure below.)

This isn't sci-fi. Tools like OpenAI's Operator and open-source browser agents can now read surveys, click consent, and answer your questions, all without a human.

LLM Pollution is not hypothetical. It is happening now.

We map three variants in the paper 👇

Recognising, Anticipating, and Mitigating LLM Pollution of Online Behavioural Research Online behavioural research faces an emerging threat as participants increasingly turn to large language models (LLMs) for advice, translation, or task delegation: LLM Pollution. We identify three int...

🚨New paper alert!

"Recognising, Anticipating & Mitigating LLM Pollution of Online Behavioural Research"

Online experiments are being polluted by LLMs. We map the threat and fixes🧵

w/ Raluca Rilla, @hiromu1996.bsky.social, @iyadrahwan.bsky.social &
Anne-Marie Nussberger

arxiv.org/abs/2508.01390