An unnamed research group at University of Zurich is in the news for unethical research: According to this article by Science and the corresponding Reddit report, a research group experimented with the use of LLMs posing as humans on a subreddit (forum). The research question was to see if and how LLMs were more successful than humans to change people’s opinions on a subject. The research was performed without the consent of any of the involved parties.
Informed consent is an ethical (in general) and legal (by funders) requirement that no research be performed on human subjects without their understanding (“informed”) and agreement (“consent”). The researchers clearly violated this requirement and should have their funding withdrawn and be sent to the doghouse. The university responded and argued that the researchers felt “the bot […] did little harm”. According to the Reddit moderator, the bot played the card of a (fictional) victim of rape and I’ll leave to the reader to have an opinion now.
Beyond informed consent, there are further ethical considerations, most notably beneficence, which are often given short shrift. They shouldn’t. In a research context, beneficence requires that the benefits of the research significantly outweigh the harms. A consequence is that a researcher has to weigh whether other research would not be more appropriate than the one on the mind of a researcher. Here the laziness or lack of imagination of the University of Zurich researchers shows. Off the bat I can imagine any number of other ways to similar data to answer the question of how effective LLM-based bots can be at swaying people’s opinions.
For one, I’m inundated by clearly LLM-generated consulting work requests on LinkedIn, clearly LLM-generated job applications to my professor email address, and clearly LLM-generated sales attempts to my business email address. Many colleagues and friends report about similar experiences. Clearly the data the researchers are seeking already exists in various forms and guises.
It remains a question whether the researchers would be able to gain access to this data and whether it would be ethical to use such data. (Some of these natural experiments might be questionable, for example, if they were violating the GDPR rules). However, the significantly higher beneficence of using this data over the Reddit-generated data is obvious to me.
Did I mention the many LLM-generated emails students send me trying to convince me of giving them homework extensions due to “personal circumstances beyond their control”?
Leave a Reply