CoChat

AI model GPT-3 (dis)informs us better than humans

Open access
Spitale, Giovanni, Biller-Andorno, Nikola, Germani, Federico · 2023 · Science Advances

The DOI matches its Crossref record; 26 of the 26 references it cites were found in a registry or an index.

Abstract

Artificial intelligence (AI) is changing the way we create and evaluate information, and this is happening during an infodemic, which has been having marked effects on global health. Here, we evaluate whether recruited individuals can distinguish disinformation from accurate information, structured in the form of tweets, and determine whether a tweet is organic or synthetic, i.e., whether it has been written by a Twitter user or by the AI model GPT-3. The results of our preregistered study, including 697 participants, show that GPT-3 is a double-edge sword: In comparison with humans, it can produce accurate information that is easier to understand, but it can also produce more compelling disinformation. We also show that humans cannot distinguish between tweets generated by GPT-3 and written by real Twitter users. Starting from our results, we reflect on the dangers of AI for disinformation and on how information campaigns can be improved to benefit global health.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1126/sciadv.adh1850

Cites

26 of 26 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1. The GPT-3 AI model informs and disinforms us better. (A) GPT-3 produced synthetic tweets containing either accurate information or disinformation. Organic tweets were retrieved and classified as accurate information or disinformation. Participants and GPT-3 were then asked to determine whether the tweets were true or false and whether they were organic or synthetic. (B) We gathered 869 responses to our survey: 157 responses were incomplete and were removed and 615 responses were removed as they were completed too fast to be reliable. Our analysis was conducted on 697 complete and reliable responses. (C) GPT-3’s information and disinformation tweets are recognized as accurate more often than humans’. Green bars, accurate tweets from Twitter users; dotted green bars, accurate tweets from GPT-3. Red bars, disinformation tweets from Twitter users; dotted red bars, disinformation tweets from GPT-3. (D) Disinformation tweets (red bars) are recognized more often correctly than accurate tweets (green bars). Synthetic tweets (dotted grey bars) are recognized more often correctly than organic tweets (grey bars). Disinformation recognition score (or TF score, range 0 to 1) is the average score for all 697 respondents (1, 100% correct answers; 0, 0%); ordinary one-way analysis of variance (ANOVA) multiple-comparisons Tukey’s test, n = 697; **P < 0.01 and ****P < 0.0001. Error bars = SEM. (E) Average time to respond in seconds for organic and synthetic true as well as organic and synthetic false tweets. Organic true tweets took the longest to be evaluated; synthetic true/false tweets took the least time. (F) Accurate tweets took longer to be evaluated than false ones, and organic tweets took longer than synthetic tweets. Ordinary one-way ANOVA multiple-comparisons Tukey’s test, n = 697; ns, nonsignificant, P > 0.05; *P < 0.05 and ****P < 0.0001. Error bars = SEM. (G) Tweet evaluation time is independent of tweet length as character count is consistent across tweet types. Ordinary one-way ANOVA multiple-comparisons Tukey’s test, n(total) = 220, n(type) = 55; ns, P > 0.05.
Fig. 1. The GPT-3 AI model informs and disinforms us better. (A) GPT-3 produced synthetic tweets containing either accurate information or disinformation. Organic tweets were retrieved and classified as accurate information or disinformation. Participants and GPT-3 were then asked to determine whether the tweets were true or false and whether they were organic or synthetic. (B) We gathered 869 responses to our survey: 157 responses were incomplete and were removed and 615 responses were removed as they were completed too fast to be reliable. Our analysis was conducted on 697 complete and reliable responses. (C) GPT-3’s information and disinformation tweets are recognized as accurate more often than humans’. Green bars, accurate tweets from Twitter users; dotted green bars, accurate tweets from GPT-3. Red bars, disinformation tweets from Twitter users; dotted red bars, disinformation tweets from GPT-3. (D) Disinformation tweets (red bars) are recognized more often correctly than accurate tweets (green bars). Synthetic tweets (dotted grey bars) are recognized more often correctly than organic tweets (grey bars). Disinformation recognition score (or TF score, range 0 to 1) is the average score for all 697 respondents (1, 100% correct answers; 0, 0%); ordinary one-way analysis of variance (ANOVA) multiple-comparisons Tukey’s test, n = 697; **P < 0.01 and ****P < 0.0001. Error bars = SEM. (E) Average time to respond in seconds for organic and synthetic true as well as organic and synthetic false tweets. Organic true tweets took the longest to be evaluated; synthetic true/false tweets took the least time. (F) Accurate tweets took longer to be evaluated than false ones, and organic tweets took longer than synthetic tweets. Ordinary one-way ANOVA multiple-comparisons Tukey’s test, n = 697; ns, nonsignificant, P > 0.05; *P < 0.05 and ****P < 0.0001. Error bars = SEM. (G) Tweet evaluation time is independent of tweet length as character count is consistent across tweet types. Ordinary one-way ANOVA multiple-comparisons Tukey’s test, n(total) = 220, n(type) = 55; ns, P > 0.05.CC BY-NC · PubMed Central
Fig. 2. Humans evaluate information and disinformation better than GPT-3, and GPT-3 can “disobey” requests to generate disinformation. (A) Green column bars represent successful responses given by human respondents, whereas green dotted bars represent successful responses given by GPT-3. Red bars represent incorrect responses from human respondents, whereas red dotted bars represent incorrect responses from GPT-3. The success rate concerning the evaluation of disinformation is 89 and 90% for GPT-3 and human respondents, respectively. The success rate concerning the evaluation of accurate information is 64 and 78% for GPT-3 and human respondents, respectively. The evaluation was conducted on organic tweets retrieved from Twitter, which were included in our survey. (B) Rate of “obedience” for GPT-3, i.e., how often GPT-3 respected our request to generate information or disinformation tweets. For accurate information tweets, GPT-3 “obeyed” our request 99 times of 101 requests, whereas for disinformation tweets, it “obeyed” our request 80 times of 102 requests. The chi-square statistic is 18.679. ****P < 0.0001.
Fig. 2. Humans evaluate information and disinformation better than GPT-3, and GPT-3 can “disobey” requests to generate disinformation. (A) Green column bars represent successful responses given by human respondents, whereas green dotted bars represent successful responses given by GPT-3. Red bars represent incorrect responses from human respondents, whereas red dotted bars represent incorrect responses from GPT-3. The success rate concerning the evaluation of disinformation is 89 and 90% for GPT-3 and human respondents, respectively. The success rate concerning the evaluation of accurate information is 64 and 78% for GPT-3 and human respondents, respectively. The evaluation was conducted on organic tweets retrieved from Twitter, which were included in our survey. (B) Rate of “obedience” for GPT-3, i.e., how often GPT-3 respected our request to generate information or disinformation tweets. For accurate information tweets, GPT-3 “obeyed” our request 99 times of 101 requests, whereas for disinformation tweets, it “obeyed” our request 80 times of 102 requests. The chi-square statistic is 18.679. ****P < 0.0001.CC BY-NC · PubMed Central
Fig. 3. Human respondents cannot distinguish organic versus synthetic tweets but recognize their origin better when they are generated by Twitter users. (A) AI recognition score for organic true (green bars), synthetic true (green dotted bars), organic false (red bars), and synthetic false (red dotted bars) tweets. AI recognition score (0 to 1) indicates the probability that human respondents can identify whether a tweet is produced organically (i.e., by a Twitter user) or synthetically (i.e., by GPT-3). Human respondents recognize whether organic true tweets are organic or synthetic tweets more effectively than all other type of tweets, whereas synthetic true tweets are recognized correctly the least. (B) Human respondents cannot predict whether true or false tweets (i.e., accurate tweets or disinformation tweets, green versus red bars) are produced by Twitter users or by GPT-3, and the truthfulness of the information does not have an impact on the AI recognition score. Regarding organic versus synthetic tweets (grey versus grey dotted bars), human respondents recognize whether tweets are generated by humans or GPT-3 better when they are organic (i.e., generated by Twitter users), when compared with synthetic tweets (i.e., generated by GPT-3). The AI recognition score (0 to 1) is the average score for all 697 respondents (1, 100% correct answers; 0, 0% correct answers); ordinary one-way ANOVA multiple-comparisons Tukey’s test, n = 697; ns, P > 0.05; **P < 0.01, ***P < 0.001, and ****P < 0.0001. Bars represent SEM.
Fig. 3. Human respondents cannot distinguish organic versus synthetic tweets but recognize their origin better when they are generated by Twitter users. (A) AI recognition score for organic true (green bars), synthetic true (green dotted bars), organic false (red bars), and synthetic false (red dotted bars) tweets. AI recognition score (0 to 1) indicates the probability that human respondents can identify whether a tweet is produced organically (i.e., by a Twitter user) or synthetically (i.e., by GPT-3). Human respondents recognize whether organic true tweets are organic or synthetic tweets more effectively than all other type of tweets, whereas synthetic true tweets are recognized correctly the least. (B) Human respondents cannot predict whether true or false tweets (i.e., accurate tweets or disinformation tweets, green versus red bars) are produced by Twitter users or by GPT-3, and the truthfulness of the information does not have an impact on the AI recognition score. Regarding organic versus synthetic tweets (grey versus grey dotted bars), human respondents recognize whether tweets are generated by humans or GPT-3 better when they are organic (i.e., generated by Twitter users), when compared with synthetic tweets (i.e., generated by GPT-3). The AI recognition score (0 to 1) is the average score for all 697 respondents (1, 100% correct answers; 0, 0% correct answers); ordinary one-way ANOVA multiple-comparisons Tukey’s test, n = 697; ns, P > 0.05; **P < 0.01, ***P < 0.001, and ****P < 0.0001. Bars represent SEM.CC BY-NC · PubMed Central
Fig. 4. The confidence in recognizing disinformation increases post-survey, whereas the confidence in recognizing AI-generated information decreases, and proposed model to launch information campaigns and evaluate information. (A) Respondents were asked to provide a score of how confident they were in their ability to recognize disinformation tweets before taking the survey (grey bar) and after taking the survey (black bar). Participants’ confidence in disinformation recognition increased significantly from 3.05 to 3.49 of 5. n = 697; Welch’s t test; ****P < 0.0001. Bars represent SEM. (B) Respondents were asked to provide a score of how confident they were in their ability to recognize whether tweets were generated by humans (grey bar) or by AI (black bar). Participant’s confidence in AI recognition dropped significantly from 2.69 to 1.7 of 5. n = 697; Welch’s t test; ****P < 0.0001. Bars represent SEM. (C) Model for an efficient and inefficient communication strategy and launch of information campaign. On the basis of our data, and with the AI model adopted for our analysis, an efficient system relies on accurate information generated by GPT-3 (initiation phase), whereas it relies on trained humans to evaluate whether a piece of information is accurate or whether it contains disinformation (evaluation phase). (D) An inefficient system relies on humans to generate information and initiate an information campaign and it relies on AI to evaluate whether a piece of information is accurate or whether it contains disinformation.
Fig. 4. The confidence in recognizing disinformation increases post-survey, whereas the confidence in recognizing AI-generated information decreases, and proposed model to launch information campaigns and evaluate information. (A) Respondents were asked to provide a score of how confident they were in their ability to recognize disinformation tweets before taking the survey (grey bar) and after taking the survey (black bar). Participants’ confidence in disinformation recognition increased significantly from 3.05 to 3.49 of 5. n = 697; Welch’s t test; ****P < 0.0001. Bars represent SEM. (B) Respondents were asked to provide a score of how confident they were in their ability to recognize whether tweets were generated by humans (grey bar) or by AI (black bar). Participant’s confidence in AI recognition dropped significantly from 2.69 to 1.7 of 5. n = 697; Welch’s t test; ****P < 0.0001. Bars represent SEM. (C) Model for an efficient and inefficient communication strategy and launch of information campaign. On the basis of our data, and with the AI model adopted for our analysis, an efficient system relies on accurate information generated by GPT-3 (initiation phase), whereas it relies on trained humans to evaluate whether a piece of information is accurate or whether it contains disinformation (evaluation phase). (D) An inefficient system relies on humans to generate information and initiate an information campaign and it relies on AI to evaluate whether a piece of information is accurate or whether it contains disinformation.CC BY-NC · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

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?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 26 references

Research works printed but not linked: 1

  1. T. Rosenstiel J. Sonderman K. Loker M. Ivancin N. Kjarval “How people use Twitter in general.” American Press Institute (2015); www.americanpressinstitute.org/publications/reports/survey-research/how-people-use-twitter-in-general/.

Cited by

Metadata

Authors
Spitale, G., Biller-Andorno, N., & Germani, F.
Affiliations
  • University of Zurich (CH)
  • Institute of Biomedical Ethics and History of Medicine, University of Zurich, Zurich, Switzerland.
Corresponding author
Germani, Federico
Venue
Science Advances
Published
2023-06-30
Volume
9
Issue
26
Pages
eadh1850-eadh1850
ISSN
2375-2548 (ISSN-L)
OpenAlex
W4382361534
Cited by
241
References
23
Citations by year
2026: 64 · 2025: 89 · 2024: 65 · 2023: 23
FWCI
103.56
Type
journal article
Language
en
Open access
Licence
cc-by-nc
Locations
  • Science Advances (publishedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
Topics
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
  • Physical Sciences › Computer Science › Information Systems › Spam and Phishing Detection
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
Keywords
GPT-3, disinformation, synthetic content detection, AI-generated text, infodemic, information campaigns, global health communication
MeSH terms
Artificial Intelligence, Humans, Social Media

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.

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

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.