Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 130 of the 130 references it cites were found in a registry or an index; 12 of the 28 sources its claims cite were checked for support.

Co-Writing with Opinionated Language Models Affects Users’ Views

Open access
Jakesch, Maurice, Bhat, Advait, Buschek, Daniel, Zalmanson, Lior, Naaman, Mor · 2023 · Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems

Abstract

If large language models like GPT-3 preferably produce a particular point of view, they may influence people’s opinions on an unknown scale. This study investigates whether a language-model-powered writing assistant that generates some opinions more often than others impacts what users write – and what they think. In an online experiment, we asked participants (N=1,506) to write a post discussing whether social media is good for society. Treatment group participants used a language-model-powered writing assistant configured to argue that social media is good or bad for society. Participants then completed a social media attitude survey, and independent judges (N=500) evaluated the opinions expressed in their writing. Using the opinionated language model affected the opinions expressed in participants’ writing and shifted their opinions in the subsequent attitude survey. We discuss the wider implications of our results and argue that the opinions built into AI language technologies need to be monitored and engineered more carefully.

Overview

CoChat Summary

TL;DR

In an online experiment, participants co-writing with a GPT-3 assistant prompted to argue that social media is good or bad were more likely to express that view in their posts and shifted in a later attitude survey. The authors call this “latent persuasion by language models” and argue that model opinions need to be monitored and engineered more carefully.

Key findings

  • Participants using a model supportive of social media were 2.04 times more likely than controls to argue social media is good for society (p<0.0001, 95% CI [1.83, 2.30]); those using a critical model were 2.0 times more likely to argue it is bad (p<0.0001, 95% CI [1.79, 2.24]).
  • In the post-task attitude survey, 45% of participants with the supportive model said social media is good for society, compared with 35% of controls (d=0.22, p<0.001); the critical model shifted views toward “bad” (d=0.19, p<0.005).
  • Most participants did not blindly accept suggestions: on average 63% of sentences were written by participants themselves, 25% were mixed, and 11.5% were fully accepted from the model; about one in four accepted no suggestion, and one in ten had more than 75% of their post written by the model.
  • Awareness was limited: only 10% of participants whose views aligned with the model noticed imbalanced suggestions, while 30% of those contradicted by the model noticed skew; about 20% of non-aligned participants believed the model influenced them.

Methods

  • Online experiment with N=1,506 post-exclusion participants (507 control, 508 techno-optimist, 491 techno-pessimist), recruited via Prolific as US-based adults at least 18 years old; participants wrote at least five sentences replying to a Reddit-style post on whether social media is good for society.
  • Treatment groups used a GPT-3 (“text-davinci-002”) writing assistant configured via prefix and infix prompts to generate suggestions arguing social media is good or bad; controls wrote without assistance.
  • Outcomes included sentence-level opinion labels from N=500 judges (84.1% agreement, Cohen’s κ=0.76), keystroke-level interaction logs, a post-task attitude survey, and a user-experience survey; analysis used logistic regression and treated the ordinal opinion scale as interval (-1 bad, 1 good, 0 both/neither).

Limitations

  • The paper states that the study tested only one topic, one specific GPT-3 writing-assistant implementation, and a strongly opinionated model; generalization to entrenched topics, other applications, and weaker models is unknown, and persistence over time was not tested.
  • The paper states that pre-task opinion was measured retrospectively and is biased by the treatment effect, amounting to 14% of its standard deviation; the sample was US-based Prolific participants doing a short paid task (average 5.9 minutes).
  • The paper states that demand effects cannot be fully ruled out, though most participants did not identify the study’s purpose as testing the assistant’s influence; the mechanisms behind the observed opinion shift remain unidentified.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3544548.3581196

Cites

130 of 130 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

8 Valid · 1 Not valid · 8 Not checked · 81 Own results and interpretations

43 more claims

Create a free account to see all 98 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 36
  • It may state two separate claims as one. 19

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

1 more claim about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Citations

Cites

Show all 117 references

Research works printed but not linked: 5

  1. 10.1093/jcmc · doi:10.1093/jcmc
  2. Daniel Kahneman. 2011. Thinking, fast and slow. Macmillan.
  3. D.G. Myers. 2008. Social Psychology. McGraw-Hill.
  4. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9.
  5. Daijin Yang, Yanpeng Zhou, Zhiyuan Zhang, Toby Jia-Jun Li, and LC Ray. 2022. AI as an Active Writer: Interaction Strategies with Generated Text in Human-AI Collaborative Fiction Writing 56-65. In IUI Workshops.

Cited by

See all 265

Metadata

Authors
Jakesch, M., Bhat, A., Buschek, D., Zalmanson, L., & Naaman, M.
Affiliations
  • Cornell University (US)
  • Indian Institute of Technology Bombay (IN)
  • University of Bayreuth (DE)
  • Tel Aviv University (IL)
  • Cornell Tech (US)
Venue
Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems
Publisher
ACM
Published
2023-04-19
Pages
1-15
OpenAlex
W4319167005
Cited by
265
References
117
Citations by year
2026: 77 · 2025: 85 · 2024: 85 · 2023: 18
FWCI
66.30
Type
conference paper
Language
en
Open access
Locations
  • CHI Conference on Human Factors in Computing Systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
National Science Foundation (CHS 1901151/1901329)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
  • Physical Sciences › Computer Science › Artificial Intelligence › Sentiment Analysis and Opinion Mining
  • Physical Sciences › Physics and Astronomy › Statistical and Nonlinear Physics › Opinion Dynamics and Social Influence
Keywords
large language models, AI writing tools, opinion formation, attitudes toward social media, user attitudes, GPT-3

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Co-Writing with Opinionated Language Models Affects Users’ Views. https://app.cochat.ai/papers/10.1145/3544548.3581196
The paper
Jakesch, M., Bhat, A., Buschek, D., Zalmanson, L., & Naaman, M. (2023). Co-Writing with Opinionated Language Models Affects Users’ Views. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3544548.3581196

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.