At a glance
- What it is: Artificial Intelligence in deliberation: The AI penalty and the emergence of a new deliberative divide
- Who: Andreas Jungherr (University of Bamberg, Germany) and Adrian Rauchfleisch (National Taiwan University).
- Where: Government Information Quarterly, September 2025, open access. doi.org/10.1016/j.giq.2025.102079
- Type: preregistered survey experiment, representative sample of Germany (n = 1,850).
First reading: what it does and what it finds
It helps to start with the type of study, because it frames everything else. It is not a pilot where people use an AI and then give their opinion. It is a survey experiment: a representative sample of 1,850 adults in Germany was given descriptions of specific tasks in an online deliberation (summarizing contributions, organizing arguments, aggregating opinions, checking facts, moderating, translating), and each task was randomly attributed to “AI tools” or to “human moderators.” The description was identical; the only thing that changed was who did it. Each person saw six tasks, three human and three AI, and answered two things: how willing they would be to participate and what quality they expected from that forum.
The design is deliberately informational rather than experiential. The authors defend it: they wanted to isolate the effect of learning that AI is involved, before any contact. It is the prior filter, the moment when someone decides whether or not to enter the room. And there they found what they call an AI penalty. Announcing that the task is done by an AI lowers willingness to participate (from 47% to 40% of people above the midpoint of the scale) and lowers expected quality (from 57% to 48%). The effects are modest in magnitude but statistically clear and consistent task by task.
The most interesting detail is where the penalty hits hardest: precisely in the tasks where academics see the greatest promise for AI, those of processing information and mapping arguments. For fact-checking, by contrast, people rated AI similarly to a human. The authors themselves highlight this as a paradox, because fact-checking is, in their view, where today’s AI is least reliable. In the open-ended responses, 45% of those who declined to participate mentioned distrust or fear of AI, and 13% said they preferred dealing with humans.
The finding that gives the paper its title comes next. Who penalizes more? Not those with less education or fewer resources. In fact, income, education, having young children, and prior experience with participation did not predict willingness to participate. What did matter was, on the one hand, political interest and enjoyment of thinking through complex problems (what the literature calls need for cognition), and on the other, specific attitudes toward AI. Those who see social benefits in it, those who use it, and those who tend to attribute human traits to it barely penalize it at all; those who perceive it as risky penalize it more. And a fine nuance: general attitudes toward technology did not move the needle, only attitudes toward AI in particular. Hence the idea of a new deliberative divide: on top of the usual inequalities, which follow education level or social background, comes another one that follows what each person thinks of AI.
Second reading: from Latin America
The first thing is to place the study. It is Germany, with a single national sample, and the authors themselves warn that cultural affinity with AI varies across countries and that the penalty could vary too. So none of this carries over automatically to the region. What travels well is not the number, but the mechanism: when people learn that an AI will mediate a public conversation, that alone changes their willingness to join in.
This connects directly with something we have already read here. In the Doble Lectura on politicians and public officials we saw how governments in the region are starting to use AI to organize citizens’ comments in consultations and participatory processes. That study looked at those who adopt the tool within the state; this one looks at the other side of the counter, at the citizens who decide whether to participate. And it adds an uncomfortable warning for the Latin American participatory tradition, the one of participatory budgeting and public consultations: if announcing that AI is involved drives away the most skeptical, the process may end up listening to a less representative version of the public. Exactly the opposite of what a consultation seeks.
A tension also emerges that the paper leaves on the table: transparency has a cost. Disclosing that an AI helped systematize opinions is the right thing to do, but according to this study that notice pays a toll in participation. Not disclosing it avoids the toll at the price of hiding something people evidently would want to know. There is no clean way out, and the region will have to choose with its eyes open.
One open question remains that the study does not answer and that I want to flag as my own hypothesis, not as a finding of the authors. If the penalty depends on how much people distrust AI, and in other readings we have seen signs of greater optimism about AI in poorer countries, the penalty may be smaller in Latin America than in Germany. That would be good news, but it could also mean the opposite: less penalty because of less distrust, and not because there are more reasons to trust. The paper did not study the region, so this is my reading and it stays as a question for research, not as a conclusion.
The fine print
- People did not use any AI: they responded to descriptions. The authors present this as a strength for capturing the prior filter, and they have a point, but they acknowledge that experiencing the tool could change reactions. In fact, their result clashes with other work in which, after using it, people preferred AI facilitation.
- It is a single German sample, and the design did not separate AI that organizes content from AI that generates it, so the penalty cannot be attributed to any particular type of tool.
- The authors declare no conflict of interest; it is independent academic research, funded by European and Taiwanese public funds, not by an AI company.
Paper keywords: Deliberation, Artificial intelligence, Survey experiment, Political behavior, Participation
Automated reading. This text was generated by Claude, an Anthropic model, from the original source, without line-by-line human review. It may contain errors or debatable interpretations; to check any point, see the original source.