acceptodds
Under review as a conference paper at ICLR 2027

Does Anthropomorphic Language Impact Public Perceptions of AI?

Abstract

Public discourse about artificial intelligence (AI) often uses anthropomorphic language: language that attributes human capabilities and characteristics to AI systems. This practice has been criticized for setting misleading expectations, inflating claims, and fueling hype around AI, which may distort public understanding of AI and impact policy priorities. We study the effects of anthropomorphic framing by comparing changes in participants’ perceptions of AI (N=815) when reading passages with and without anthropomorphic language, designed to reflect realistic public-facing AI discourse. We further examine whether these effects differ across two types of AI technologies—large language models and recommendation systems—and measure changes in perceptions of AI across several dimensions that are prominent in current public discourse. In a separate condition using a text that explicitly discusses the dangers of AI, we show that individuals' views of AI can shift in response to reading a text; yet in the main conditions of the experiment, where we compare anthropomorphic and non-anthropomorphic descriptions, we find that whether the text uses anthropomorphic language does not substantially affect participants' perceptions of AI. Our results indicate that any immediate effects on opinions of AI are modest, although they leave open the possibility that anthropomorphic language could have an effect in naturalistic settings, or over gradual, continued exposure.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
anthropomorphism, language, human study

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CD9sOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents a human study on anthropomorphic language when describing AI and people's perception.

Critical strengths
  • anthropomorphic portryal of AI is widespread and worth consideration
Critical weaknesses
  • Is this the right venue?
Questions for authors
  • My chief concern is that perhaps an ML conference is not the right venue? Something like CSCW/Facct/Nature Human Behavior might be more fitting?

  • Is recommender systems the right A/B testing against AI? I feel like the anthro description of recommender systems in Table 1 is unconvincing: it is just kind of hard to imagine it as a human in the first place, so it might be a weak baseline?

  • Would participants have strong prior registered opinion about the questions surveyed?

  • The results and analysis is too short (for the main paper), only one illustrative figure 2 without statistics, demographic breakdowns, anything deeper than the headline result.

3 comments
AuthorsOpenReview ↗

Response to Reviewer CD9s

Thank you for your feedback! Below we respond to the points raised:

Scope of paper for ICLR

Anthropomorphism has been discussed widely in the ML community and the broader field, and our work provides empirical grounding on the question in order to inform how AI systems are designed, evaluated, and discussed, which we believe is of direct interest to the ICLR community. We chose to submit this paper here given that the CFP listed “societal considerations including fairness, safety, privacy”, and explicitly encouraged authors to “take the opportunity to be ambitious” highlighting “integration with … humanistic fields outside the historical domain of ‘core ML’”.

Choice of recommender systems

We chose recommender systems as a type of AI system that is perceived differently from LLMs to examine whether the effects of anthropomorphism extend beyond LLMs, so we do not intend for recommender systems to be an equally anthropomorphized baseline. We examine whether anthropomorphic framing affects either technology and recognize that the magnitude of any effect may differ across technologies, although our results show that the effects are negligible in both cases in our setting.

Participant priors on survey questions

We control for any previous exposure or prior beliefs on the topics by measuring the shift in each participant’s pre and post responses, rather than only comparing post responses across treatments. From our raw pre-survey data (included in Appendix J), we also see that pre-survey responses were similar across conditions and showed relatively neutral stances. Questions on a 1-4 scale had pre-survey responses ranging from 1.8-3.2, questions on a 1-3 scale had responses from 1.7-2.3, and questions on a 0-1 scale had responses from 0.1-0.4, so we did not observe any ceiling effects.

Analysis details

We included main findings across four experiments in pages 7-9 of the paper, specifically the effects of the Doomsday packet, effects of anthropomorphic vs non-anthropomorphic packets, differences in effects across technologies, and changes in writing style, and will aim to include more details of statistical analysis in the next version. Additional figures and raw results are in Appendices J-P due to space constraints, and demographic breakdowns are in Appendix B.

I would like to thank the authors for their clarifications. I still have reservations about venue fit as well as whether the depth of insights meet the standard. The AC could eventually weigh in on this.

Thank you for the response. We appreciate your continuous engagement and input on our work.

Reviewer_pWsWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

At present, public discussions and descriptions of AI can sometimes be anthropomorphic, which might raise the concern of misleading the public understanding of AI tools, with further consequences in policies. Facing such concern, the paper is mainly to investigate whether different levels of anthropomorphic descriptions of AI can influence people's perceptions of AI. Such goal is to be achieved by a survey described in the paper. To begin with, they recruited 900 respondents (815 retained after exclusions) on Prolific, who were divided into five groups. Four such groups were due to the binary choices of technology concerned and being anthropomorphic or not when providing downstream articles. The additional group is for a test group who were provided with articles describing severe risks of AI. During the procedure, each respondent went through pre-survey, packets of articles and finally post-survey. Both surveys were concerned about five groups of questions on views of AI, while the packets of articles provided different versions of texts describing AI. There is also a writing task in each survey for respondents to describe some given interactions between a human and a chatbot. Using Bayes factors, the paper finds that anthropomorphic approach may have limited immediate influence on perceptions of AI.

Critical strengths

Clear presentation

  • The paper is clearly structured and well written. In the main sections, the paper starts with introducing the main components of the survey, including the procedure (pre-survey, briefing packets, post-survey) and the grouping due to technology and use of anthropomorphism. Then further details of each stage are provided, with examples of briefing packets. The five groups of survey questions are also respectively covered. Finally, the main statistical comparisons used were also introduced to support the conclusions.
  • Figures have been appropriately used to further help with understanding the experimental steps and results.
Critical weaknesses

Insufficient validation of the level of anthropomorphism in the texts

  • Since the main goal of the paper is to investigate whether (non-)anthropomorphic versions of content would change people's perceptions of AI, the assumption should be that the two versions have enough gap in the level of anthropomorphism. As a result, more clarification or justification is needed to show that the current approach to creating such two versions can produce a meaningful difference in the level of anthropomorphism as judged by readers. A potential support can also be a pilot study.
  • The group of Doomsday packet seems more to provides evidence that the reading materials can change some measured perceptions, and might not be able to answer ths above question.

More explanation of the statistical calculations would be helpful

  • The main conclusions are based on the statistical analysis, while more clarifications on how the Bayes factor and related terms are calculated could be helpful, especially in the main paper.
  • Mentioning main assumptions behind the above calculations could also better help readers to assess the conclusions.
Questions for authors

For clarification

  • Could the authors provide further justification or pilot study-based validation that the current two levels of (non-)anthropomorphism differ meaningfully as judged by respondents?
  • Could the authors provide further details or steps in the calculation of Bayes factors and related quantities, as well as their assumptions?
4 comments
AuthorsOpenReview ↗

Response to Reviewer pWsW

Thank you for your feedback! Below we respond to the points raised:

Gap in level of anthropomorphism between packets

We did not survey general audiences to validate differences in anthropomorphism between packets because lay participants may not have a well-defined notion of or tend to notice anthropomorphic language in the context of AI, making their judgments potentially noisy or unreliable as a measure of the intended linguistic manipulation. Instead, the authors independently assessed the differences between anthropomorphic and non-anthropomorphic packets. We additionally used AnthroScore (Cheng et al., 2024), a metric of anthropomorphism defined as the log-ratio of the probabilities that the AI system would be replaced by human versus non-human pronouns within a text. We found that, on average, anthropomorphic packets have a 1.16 point higher AnthroScore than non-anthropomorphic packets. For comparison, Cheng et al. report a difference of ~0.3 points between language model papers and papers from other computer science subfields, suggesting that our experimental conditions differ substantially in anthropomorphic framing, as intended. We will be sure to clarify our validation processes in the paper.

Bayes factor clarification

Due to space constraints, background on the Bayes factor analysis method and the assumptions made in our analysis are in Appendix M. Appendix O additionally shows the sensitivity analysis done to verify that the results of our analysis are not sensitive to the assumptions made on the prior. We recognize that descriptions of how Bayes factors are calculated and the assumptions made would be useful, and we will include them in the main text in the next version.

Thank you for the response!

Level of anthropomorphism

  • The above mentioned AnthroScore seems to be a newly reported result that has not been included in the paper.
  • A 1.16 point difference, compared to a difference of ~0.3 points between language model papers and papers from other computer science subfields, might not justify how substantial the difference in anthropomorphism is. One perspective to clarify is statistical uncertainty. Another perspective is how and why the rather specific score difference between language model papers and papers from other computer science subfields can serve as a convincing baseline.
  • Are there packet-level scores?
  • More fundamentally, anthropomorphism could vary in degree and form. According to the reviewer's understanding, the reported experiment only examines a particular set of wording contrasts. Such contrasts might not represent the broader range of anthropomorphic language. As a result, the authors might want to describe their conclusions as evidence of limited immediate effects for the specific settings, rather than a more general conclusion.

Thank you for the response. Yes, the AnthroScores were not included in the paper; we will add them to the revised paper. We used AnthroScore as an additional measure on top of our manual checks to verify the difference in level of anthropomorphism across conditions. The 1.16 difference alone does not establish that the difference is substantial, given the lack of an established baseline or threshold for interpreting AnthroScores. More broadly, defining and quantifying degrees of AI anthropomorphism remains an open problem, with multiple recently proposed methods in addition to AnthroScore (Xu et al., 2026; Lonke et al., 2025).

We agree that anthropomorphism can take different forms and vary in degree, and that our operationalization may capture only a subset of these variations. Our findings provide evidence for limited immediate effects of anthropomorphic language in this specific setting. We will clarify this in the paper, particularly when introducing our definition and operationalization of anthropomorphism.

References:

  1. Xu, Y., Chi, C. G., Gursoy, D., & Cai, R. (2026). “Rethinking AI Anthropomorphism: A Holistic Conceptualization and Scale Across AI Systems and Service Contexts.” Technology in Society, 85, 103189.
  2. Lonke, D., Bloem, J., & Sommerauer, P. (2025). “AnthroSet: A Challenge Dataset for Anthropomorphic Language Detection.” Proceedings of the OMMM Workshop, 27–39.

Thank you for the clarification!

Reviewer_2Z5UOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a controlled experiment that studies the effects of anthropomorphic language over several dimensions of human perception of two technologies: LLMs and recommendation systems. Prolific human participants filled out a survey measuring their baseline views about LLMs or recommendation systems, read a briefing about the topic (the intervention) containing either anthropomorphic language or neutral language, and then filled out a post-intervention survey measuring their views after the reading. The tested dimensions include 5 topics: harms related to AI use, loss of control over AI, workforce replacement by AI, effectiveness of AI safety testing, and overall impact of AI on society. The study measure the change of perception of participants in these topics after controlled exposure to anthropomorphic (or neutral) language, with an extra condition of risk-focused content (doomsday scenarios) that measure the effect of content in contrast to the core conditions which test the effect of anthropomorphic style of writing. The authors found that anthropomorphic language does not substantially affect participants’ perceptions of AI. Conversely, explicitly risk-focused reading materials changed participants’ views, suggesting that participants are more sensitive to the substance of the text than to the anthropomorphic language style.

Critical strengths
  1. The study has practical significance. It undermines a widespread claim about the use of anthropomorphic language, which little prior work has directly examined.

  2. The paper is clearly written; the experiment and results are explained in plain language and are easy to follow.

Critical weaknesses
  1. This paper is out of scope for this venue, although it touches on the effects of science communication relevant to AI researchers and practitioners. This weakness means two things: a) the ICLR reviewer pool might not contain the relevant expertise to evaluate this paper, and b) the impact of this paper might be more substantial at a different venue. It is possible that HCI conferences or a relevant sub-Nature journal (e.g., Nature Machine Intelligence or Nature Communications) would be more appropriate.

  2. Participants could have been exposed to similar articles before, meaning the effect is limited to the change from their baseline but does not account for previous encounters with anthropomorphism language. Given the null result, one concern is that participants may already have been exposed to the manipulation outside the experiment.

Questions for authors
  1. How did you decide on 900 participants (200 per core condition, 100 for doomsday)? Did you perform a pilot experiment with power analysis?

  2. You mention the materials vary in their level of anthropomorphism. How does the magnitude of anthropomorphism affect the perception?

  3. While Bayes factors could be a standard psychology tool, they are not a standard tool for the audience of this venue and deserve additional explanation in the main text. In particular, the plot labels and alternating between BF_{10} and BF_{01} is confusing. Stating the x-axis function would make the results more readable for a larger audience.

  4. Why did you choose this venue?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2Z5U

Thank you for your feedback! Below we respond to the points raised:

Scope of paper for ICLR

Anthropomorphism has been discussed widely in the ML community and the broader field, and our work provides empirical grounding on the question in order to inform how AI systems are designed, evaluated, and discussed, which we believe is of direct interest to the ICLR community. We chose to submit this paper here given that the CFP listed “societal considerations including fairness, safety, privacy”, and explicitly encouraged authors to “take the opportunity to be ambitious” highlighting “integration with … humanistic fields outside the historical domain of ‘core ML’”.

Prior exposure to similar articles

We control for any previous exposure to articles or prior beliefs on the survey topics by measuring the shift in each participant’s pre and post responses, rather than only comparing post responses across treatments. Given our sampling procedure and random assignment to conditions, it is highly unlikely that there would be systematic differences in prior exposure to anthropomorphism between groups.

Sample size

Yes, we conducted a pilot study with an initial set of survey questions that had one question per category (harms related to AI use, loss of control over AI, workforce replacement by AI, effectiveness of AI safety testing, and overall impact of AI on society). This was done with 75 participants per condition and the power analysis suggested ~200 participants. Based on these results, we expanded the set of survey questions to include multiple questions per category and conducted the full study.

Level of anthropomorphism of packets

The packets in our study only show two levels of anthropomorphism, that is, with anthropomorphic language either enhanced or removed wherever possible per our methodology. We do not observe whether varying levels of anthropomorphism leads to additional variation in perceptions. Relatedly, our method for creating the briefing packets only discerns between two levels.

Bayes factor

We include background on bayes factor analysis in Appendix M due to space constraints, but agree that this information would be useful and will move it back in the main text. We appreciate the suggestions on clarifying the plot labels and will apply them in the next version.

Thank you for the clarifications. The pilot study and power analysis are reassuring, and I appreciate the willingness to improve the presentation of the Bayes factor analysis.

However, I am not fully convinced by the response regarding prior exposure. Measuring pre-intervention beliefs does not control for prior exposure to anthropomorphic language, which could reduce participants’ sensitivity to the experimental manipulation and attenuate the measured pre-post change. While randomization makes systematic differences in prior exposure across conditions unlikely, it does not address a potential saturation effect across the sample. Specifically, the null result may reflect participants having already been exposed to and influenced by anthropomorphic framing, leaving little room for additional change, rather than the absence of an effect of such language. The current design cannot distinguish between these explanations.

I still do not think this is the best venue: I find the question interesting and \emph{somewhat} relevant to the AI community, although I am not entirely convinced that ICLR is the most appropriate venue for this contribution (not to mention we do not have the right expertise to assess the quality of this work).

I will keep my score.

Thank you for the response. Our previous response referred to controlling for participants’ prior beliefs on the five survey topics, not on prior exposure to anthropomorphism of AI. We appreciate your point about effects being saturated due to the existing prevalence of anthropomorphic language in AI discourse. Indeed, one possible explanation of our null result is that the level of anthropomorphism in AI discourse is so prevalent that it overwhelms any effects that could be observed in our experiment, and we will clarify this in the paper. However, we do not believe that this effect needs to be controlled for because the current discourse on anthropomorphic language assumes that anthropomorphism is already the dominant mode of referring to AI systems (Inie et al., 2026; Cheng et al., 2025). Thus, what is relevant to the debate is whether attempts to either reduce or increase the use of anthropomorphism are likely to change individuals’ perceptions in this context.

For this reason, in the paper we emphasize that our findings concern the immediate effects of exposure to anthropomorphism and may not reflect effects on long time scales or across large populations (lines 25-28, 82-87, 466-471). Our results suggest that a single exposure to anthropomorphic framing does not necessarily produce detectable changes in users' perceptions, even when the language differs substantially in its degree of anthropomorphism. Future work to investigate how repeated or sustained exposure shapes perceptions over time would be interesting, including whether prior exposure moderates the effects of subsequent anthropomorphic framing.

We believe that establishing these boundaries on the immediate effects of anthropomorphic language is a meaningful first step toward understanding how users perceive AI systems and the effects of language and discourse on these perceptions, an important yet poorly understood area.

References:

  1. Inie, N., Zukerman, P., & Bender, E. M. (2026). “De-anthropomorphizing ‘AI’: From Wishful Mnemonics to Accurate Nomenclature.” First Monday, 31(2).
  2. Cheng, M., Blodgett, S. L., DeVrio, A., Egede, L., & Olteanu, A. (2025). “Dehumanizing Machines: Mitigating Anthropomorphic Behaviors in Text Generation Systems.” Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 25923–25948.

Thank you for the clarifications. I agree it is a first step, and as I mentioned I do not have the right expertise in this particular subdomain to judge with high certainty. I will maintain my score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.