Shitstorm surrounding the University of Zurich: Researchers pretended to be rape victims in an online experiment

The University of Zurich is currently facing fierce hostility on the discussion platform Reddit. "Sad, such sloppy, unethical research," writes one user, "incredibly disappointing," says another. Other comments include: "crazy," "they should have known better," and "I had to gag."

The shitstorm was triggered by a study conducted by university researchers on Reddit. Reddit is an online platform where users can find discussion forums on a wide variety of topics. In one of the largest of these forums, a subreddit titled "Change my View," users post their opinions and ask others to persuade them to change their minds with sound arguments.

Users post views like: "Modern medicine is better than natural remedies, and it's dangerous to claim otherwise." Or: "Trump's tariff war isn't stupid or incompetent; he wants to strengthen the dollar." The author and all readers have the opportunity to reward those comments that have managed to change their opinion – with the mathematical symbol for change: a delta.

This makes the subreddit a perfect field for research: Based on the assigned deltas, it is possible to measure which comments were most successful in changing attitudes.

Contributions from AI chatbots are explicitly prohibited in the subreddit. The idea of ​​the discussion forum is for real people to interact respectfully and help each other question their viewpoints – without the aid of technology.

But now the administrators of the subreddit published a statement over the weekend in which they informed their users about an unauthorized experiment at the University of Zurich: The researchers had placed AI-generated comments under users' posts in order to influence their opinions.

The researchers allegedly faked several identities. In one post, the AI ​​pretended to be a man who had been raped by a 22-year-old when he was 15. In another case, it pretended to be a Black man arguing against the Black Lives Matter movement.

There were also alleged posts from a therapist specializing in abuse and from someone accusing a religious group of causing the "deaths of hundreds of innocent traders, farmers, and villagers." The posts have since been deleted, with information provided by the subreddit's administrators.

Although the researchers also published posts in which the AI ​​argued without such false identities, they found in a pre-publication of their key findings that the personalization was particularly persuasive. This was measured by the responses people left to the AI-generated posts: The personalized discussion posts garnered a particularly high number of deltas—also significantly more than the responses from other Reddit users.

The study design raises questions. The basic principle for scientific work is that anyone conducting experiments with humans must obtain their consent and may not lie to them during the experiment. Both principles are not absolute, but can be relaxed in special cases. For example, if a subject cannot be researched otherwise, but the research results have high value for society. In such cases, potential harms and benefits must be weighed against each other.

Now, the authors of the study argue that what they have researched could not have been discovered otherwise. The persuasive power of AI tools has only been demonstrated in laboratory experiments, i.e., in "controlled, artificial environments." With this study, they can now present the first large-scale field experiment with language models.

The authors have so far remained anonymous. For crisis communications, they refer to the university's media office. The university has not yet provided any further information. Also unusual is that the authors have already posted parts of their study online. On Reddit, they write that a university ethics committee approved their project because the risks are small, but the study provides important insights. However, according to Reddit moderators, the personalized posts were created without consulting the ethics committee.

The subreddit's administrators reject the researchers' arguments. They point to a study by the chat GPT developer Open AI. This study successfully compared the argumentative persuasiveness of language models with those in the "Change My View" subreddit without misleading anyone.

To do this, Open AI used questions from the subreddit and generated appropriate answers using AI. These answers were then presented to test subjects alongside real answers from Reddit users. In this experiment, too, the chatbots' answers often outperformed human arguments.

Experiments with people who believe in conspiracy theories have already demonstrated the argumentative strengths of AI chatbots. It's hardly surprising that this works even better when you believe a real person is speaking to you, or even someone affected by the conspiracy.

Martin Steiger, a lawyer and expert on digital law, is also unconvinced by the study authors' arguments: "The end does not justify the means. The study design clearly conflicts with ethical principles," he says when asked. "This study is a significant faux pas by the University of Zurich."

Now, a dispute has erupted over what should be done with the study results. The subreddit's administrators are calling on the University of Zurich not to publish the study and to publicly apologize to its users. In their response on Reddit, the study authors defend themselves and continue to insist on publishing the study.

The university apparently supports them in this: The subreddit's administrators cite a statement from the Ethics Committee of the Faculty of Arts at the University of Zurich, according to which the case was thoroughly investigated and the professor responsible received a warning. Nevertheless, the University of Zurich considers it excessive to prevent the publication – the benefits are great, the risks minimal.

However, it's unclear whether a specialist publisher could even be found for the publication, as they also have their own ethical guidelines for studies.

Behavioral economist Ernst Fehr from the University of Zurich argues to the NZZ that economic journals would probably reject this study because there is a strict no-deception rule in economics, according to which one may not lie to or deceive study participants.

However, it's unclear whether this argument actually holds. Economists insist on the no-deception rule so stubbornly primarily because they want to protect study participants' trust in their experimenters. If participants are lied to too often, this trust will be destroyed and experiments will eventually stop working, so the argument goes. But in this case, the subjects weren't even aware they were participating in an experiment. Therefore, it would only have a very indirect impact on other behavioral experiments.

Whether the study is published or shredded, the damage has been done. Those who interacted with the AI ​​responses on Reddit feel betrayed. Those who learned about the case through other channels now have new reason to question the trustworthiness of science and the ethical principles of the University of Zurich.