CoChat
Sign Up
Cited by 1,197Computational and Text Analysis MethodsTopic ModelingSentiment Analysis and Opinion Mining

ChatGPT outperforms crowd workers for text-annotation tasks

Open access
Gilardi, Fabrizio, Alizadeh, Meysam, Kubli, Maël·2023·Proceedings of the National Academy of Sciences
DOI 10.1073/pnas.2305016120PMC PMC10372638
Open this paper in CoChat — free account

Abstract

Many NLP applications require manual text annotations for a variety of tasks, notably to train classifiers or evaluate the performance of unsupervised models. Depending on the size and degree of complexity, the tasks may be conducted by crowd workers on platforms such as MTurk as well as trained annotators, such as research assistants. Using four samples of tweets and news articles ( n = 6,183), we show that ChatGPT outperforms crowd workers for several annotation tasks, including relevance, stance, topics, and frame detection. Across the four datasets, the zero-shot accuracy of ChatGPT exceeds that of crowd workers by about 25 percentage points on average, while ChatGPT’s intercoder agreement exceeds that of both crowd workers and trained annotators for all tasks. Moreover, the per-annotation cost of ChatGPT is less than $0.003—about thirty times cheaper than MTurk. These results demonstrate the potential of large language models to drastically increase the efficiency of text classification.

AI snapshot

Written from the paper’s full text.

TL;DR

Using four datasets of tweets and news articles (n = 6,183), the paper finds that zero-shot ChatGPT annotations outperform MTurk crowd workers on several text-annotation tasks and achieve higher intercoder agreement than both crowd workers and trained annotators. ChatGPT is also far cheaper, at less than $0.003 per annotation, about thirty times cheaper than MTurk.

Key findings

  • Across four datasets, ChatGPT’s zero-shot accuracy exceeds MTurk’s by about 25 percentage points on average, and it outperforms MTurk for most tasks.
  • ChatGPT’s intercoder agreement averages 91% at temperature = 1 and 97% at temperature = 0.2, compared with about 56% for MTurk and 79% for trained annotators.
  • Relevance-task accuracy rates were 70% for content moderation tweets, 81% for content moderation news articles, 83% for US Congress tweets, and 59% for 2023 content moderation tweets.
  • Per-annotation cost for ChatGPT is less than $0.003, about thirty times cheaper than MTurk. Correlations reported: intercoder agreement and accuracy r = 0.36; ChatGPT accuracy and trained-annotator intercoder agreement r = 0.46; ChatGPT overperformance vs. MTurk and trained-annotator intercoder agreement r = −0.37.

Methods

  • Design: compare zero-shot ChatGPT annotations against MTurk crowd workers and trained annotators on the same codebooks and tasks, using trained annotators as the gold standard.
  • Data: four datasets totaling n = 6,183, including a random sample of 2,382 content-moderation tweets from 2020–2021; 1,856 US Congress tweets from 2017–2022; 1,606 newspaper articles on content moderation from 2020–2021; and 500 tweets from January 2023, of which 339 were in English.
  • Analysis: tasks included relevance, topic detection, stance, general frame detection, and policy frame detection. ChatGPT used the “gpt-3.5-turbo” API with temperature 1 and 0.2, two responses per temperature per tweet, and a new chat session for every tweet. MTurk workers were restricted to “MTurk Masters,” >90% HIT approval rate with at least 50 approved HITs, and US location. Metrics were accuracy against trained annotators and intercoder agreement.

Limitations

  • The paper states that further research is needed to understand how ChatGPT and other LLMs perform in a broader range of contexts, and it lists future questions including multiple languages, few-shot learning, semiautomated labeling, chain-of-thought prompting, and comparison across LLM types.
  • The paper states that in the 2023 sample, ChatGPT struggled with misclassifying tweets about specific user suspensions in the relevance task due to a lack of examples in the prompt, underscoring the importance of high-quality prompts.
  • Evident from the text: the evaluation uses one ChatGPT model version (“gpt-3.5-turbo”) and does not report testing in languages other than English; the paper itself lists performance across multiple languages as a future question.

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

Integrity signals

Registry

Crossref · 10.1073/pnas.2305016120

Cites

15 of 15 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 1 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    ChatGPT's zero-shot accuracy is higher than MTurk for most tasks across the four datasets.

    “Regarding accuracy, Fig. 1 shows that ChatGPT outperforms MTurk for most tasks across the four datasets.”
    • Also stated in: § Abstract · plain
      “We find that across the four datasets, ChatGPT’s zero-shot accuracy is higher than that of MTurk for most tasks.”
    Quote in the paperSubstantial claim
  • Findingown data

    ChatGPT's intercoder agreement exceeds that of both crowd workers and trained annotators for all tasks.

    “ChatGPT’s intercoder agreement exceeds that of both crowd workers and trained annotators for all tasks.”
    • Also stated in: § Abstract · plain
      “For all tasks, ChatGPT’s intercoder agreement exceeds that of both MTurk and trained annotators.”
    Quote in the paperSubstantial claim
  • Findingown data

    ChatGPT's accuracy is positively correlated with the intercoder agreement of trained annotators (Pearson's r = 0.46).

    “ChatGPT’s accuracy is positively correlated with the intercoder agreement of trained annotators (Pearson’s r = 0.46), suggesting better performance for easier tasks.”
    Quote in the paperNot a substantial claim
  • Findingown data

    ChatGPT's outperformance of MTurk is negatively correlated with the intercoder agreement of trained annotators (Pearson's r = -0.37).

    “Conversely, ChatGPT’s outperformance of MTurk is negatively correlated with the intercoder agreement of trained annotators (Pearson’s r = -0.37), potentially indicating stronger overperformance for more complex tasks.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Across the four datasets, ChatGPT's zero-shot accuracy exceeds that of crowd workers by about 25 percentage points on average.

    “Across the four datasets, the zero-shot accuracy of ChatGPT exceeds that of crowd workers by about 25 percentage points on average”
    • Also stated in: § Results · plain
      “On average, ChatGPT’s accuracy exceeds that of MTurk by about 25 percentage points.”
    Quote in the paperSubstantial claim
  • May combine two claims. 8
  • Classified as probably not a substantial claim. 10

  • May combine two claims.

    Findingreasoning

    In the 2023 sample, ChatGPT performed much better than MTurk in the second task but struggled with misclassifying tweets about specific user suspensions in the relevance task due to a lack of examples in the prompt.

    “In the 2023 sample, ChatGPT performed much better than MTurk in the second task but struggled with misclassifying tweets about specific user suspensions in the relevance task due to a lack of examples in the prompt.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On average, intercoder agreement is about 56% for MTurk, 79% for trained annotators, 91% for ChatGPT with temperature = 1, and 97% for ChatGPT with temperature = 0.2.

    “On average, intercoder agreement is about 56% for MTurk, 79% for trained annotators, 91% for ChatGPT with temperature = 1, and 97% for ChatGPT with temperature = 0.2.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Accuracy rates for relevance tasks are 70% for content moderation tweets, 81% for content moderation news articles, 83% for US Congress tweets, and 59% for 2023 content moderation tweets.

    “Accuracy rates for relevance tasks, with two classes (relevant/irrelevant) are 70% for content moderation tweets, 81% for content moderation news articles, 83% for US Congress tweets, and 59% for 2023 content moderation tweets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    ChatGPT outperforms crowd workers for several annotation tasks, including relevance, stance, topics, and frame detection, using four samples of tweets and news articles (n = 6,183).

    “Using four samples of tweets and news articles (n = 6,183), we show that ChatGPT outperforms crowd workers for several annotation tasks, including relevance, stance, topics, and frame detection.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    ChatGPT's per-annotation cost is less than $0.003, about thirty times cheaper than MTurk.

    “Moreover, the per-annotation cost of ChatGPT is less than $0.003—about thirty times cheaper than MTurk.”
    • Also stated in: § Abstract · plain
      “ChatGPT’s per-annotation cost is about $0.003 or a third of a cent—about thirty times cheaper than MTurk, with higher quality.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    The evidence is consistent across different types of texts and time periods.

    “The evidence is consistent across different types of texts and time periods.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The sample size was determined by the number of texts needed to build a training set for a machine learning classifier.

    “The sample size was determined by the number of texts needed to build a training set for a machine learning classifier.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationown data

    A lower temperature value may be preferable for annotation tasks, as it seems to increase consistency without decreasing accuracy.

    “This suggests that a lower temperature value may be preferable for annotation tasks, as it seems to increase consistency without decreasing accuracy.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    No worker could annotate more than 20% of the tweets for a given task.

    “Moreover, we ensured that no worker could annotate more than 20% of the tweets for a given task.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The authors intentionally avoided adding any ChatGPT-specific prompts to ensure comparability between ChatGPT and MTurk crowd workers.

    “We intentionally avoided adding any ChatGPT-specific prompts to ensure comparability between ChatGPT and MTurk crowd workers.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Some tasks involve a large number of classes and exhibit lower levels of intercoder agreement, which indicates a higher degree of annotation difficulty.

    “Moreover, some of the tasks involve a large number of classes and exhibit lower levels of intercoder agreement, which indicates a higher degree of annotation difficulty (11).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [11] P. S. Bayerl, K. I. Paul, What determines inter-coder agreement in manual annotations? A meta-analytic investigation Comput. Linguist. 37, 699– 725 (2011).

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Each tweet was annotated by two different crowd workers.

    “As with the trained human annotators, each tweet was annotated by two different crowd workers.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The trained annotators annotated the tweets independently task by task.

    “The coders annotated the tweets independently task by task.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Average accuracy was computed using trained human annotations as the gold standard and considering only texts that both annotators agreed upon.

    “First, we computed average accuracy (i.e., percentage of correct predictions), that is, the number of correctly classified instances over the total number of cases to be classified, using trained human annotations as our gold standard and considering only texts that both annotators agreed upon.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    To ensure annotation quality, access to MTurk tasks was restricted to workers classified as 'MTurk Masters' by Amazon, with an HIT approval rate greater than 90% with at least 50 approved HITs, and located in the United States.

    “To ensure annotation quality, we restricted access to the tasks to workers who are classified as “MTurk Masters” by Amazon, who have an HIT (Human Intelligence Task) approval rate greater than 90% with at least 50 approved HITs, and who are located in the United States.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    A new chat session was created for every tweet to ensure that the ChatGPT results are not influenced by the history of annotations.

    “We created a new chat session for every tweet to ensure that the ChatGPT results are not influenced by the history of annotations.”
    Quote in the paperNot a substantial claim
16 of 16

  • The paper relies on this without support.

    The trained annotators' annotations constitute a valid gold standard for correctness.

    “using trained human annotations as our gold standard”
    “using our trained annotators as a benchmark”

  • The paper relies on this without support.

    The 2023 tweets were not included in ChatGPT's training data.

    “a new sample of tweets posted in 2023 to address the concern that ChatGPT might be relying on memorization for texts potentially included in the model’s training dataset.”
2 of 2

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?

  • LLMs have been shown to perform well on diverse text-analysis tasks, often matching or exceeding state-of-the-art models and crowd workers, even with zero-shot prompts.

    “LLMs have been shown to perform well on diverse text-analysis tasks 36 , often matching or exceeding state-of-the-art models and crowd workers 37,38 , even with zero-shot prompts 39 .”

    From Quantifying the prevalence and impact of overreaching causal claims in social science (2026) · cites this paper as [38]

    The Abstract reports that zero-shot ChatGPT outperformed MTurk crowd workers on several text-annotation tasks. It does not report a comparison with state-of-the-art models, so it supports the crowd-worker portion of the cited claim but not the model-comparison portion. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • GPT-4 is employed for automated evaluation of generated hypotheses and provided background across the four metrics.

    “we employ GPT-4 for automated evaluation, where both the generated hypotheses and the provided background are evaluated across these aspects.”

    From Large Language Models as Biomedical Hypothesis Generators: A Comprehensive Evaluation (2024) · cites this paper as Gilardi et al., 2023

    The Materials and Methods (ChatGPT) section describes automated text annotation using gpt-3.5-turbo, providing a general precedent for LLM-based assessment. It does not describe GPT-4 evaluating generated hypotheses or provided background across four metrics; its Evaluation Metrics section specifies accuracy and intercoder agreement instead. The citing authors’ choice to employ GPT-4 is their own method, not a finding of this source. (LLM, conf=0.81)

    Experimental — accuracy not yet measured

  • Recent advancements in large language models mitigate many limitations of human coders by leveraging computational efficiency, consistency, and scalability.

    “These models mitigate many of the above mentioned limitations associated with human coders by leveraging their computational efficiency, consistency, and scalability [34–36].”

    From The quantum technology job market: data driven analysis of 3641 job posts (2026) · cites this paper as [34]

  • LLMs can match or outperform crowd workers on some text-classification tasks.

    “LLM-assisted annotation shows a similar tension: LLMs can match or outperform crowd workers on some text-classification tasks, yet may introduce systematic bias in qualitative coding (Ashwin et al., 2025;Gilardi et al., 2023).”

    From Beyond Human-Likeness: Mapping the Scientific Critique Profiles of LLMs and Human Reviewers (2026) · cites this paper as Gilardi et al., 2023

  • Prior studies have reported that LLM evaluations can be competitive with crowdworker annotations.

    “Prior studies have also reported that LLM evaluations can be competitive with crowdworker annotations (Gilardi et al., 2023).”

    From MAPLE: Multi-Aspect Panels of LLM Evaluators for Open-Ended Questions (2026) · cites this paper as Gilardi et al., 2023

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. M. Desmond, E. Duesterwald, K. Brimijoin, M. Brachman, Q. Pan, Semi-automateddatalabeling, in NeurIPS 2020 Competition and Demonstration Track, (PMLR, 2021), pp. 156–169.

Cited by

Sorted by citations

Related

Authors
Gilardi, F., Alizadeh, M., & Kubli, M.
Affiliations
Harvard University (US)
Corresponding author
Gilardi, Fabrizio
ORCID
Venue
Proceedings of the National Academy of Sciences
Publisher
National Academy of Sciences
Published
2023-07-18
Volume
120
Issue
30
Pages
e2305016120-e2305016120
ISSN
  • 0027-8424 (ISSN-L)
  • 1091-6490
DOI
10.1073/pnas.2305016120
PMC
PMC10372638
OpenAlex
W4384662964
Cited by
1,197
References
12
Citations by year
2026: 440 · 2025: 356 · 2024: 274 · 2023: 125 · 2022: 1 · 2021: 1
FWCI
554.15
Type
journal article
Language
en
Open access
Licence
Locations
  • Proceedings of the National Academy of Sciences (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Zurich Open Repository and Archive (University of Zurich) (acceptedVersion, cc-by-nc-nd)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • Zurich Open Repository and Archive (University of Zurich) (submittedVersion, other-oa)
  • arXiv (Cornell University) (cc-by)
  • Zurich Open Repository and Archive (University of Zurich)
Funders
EC | European Research Council (883121)
Topics
  • Social Sciences › General Social Sciences › Computational and Text Analysis Methods
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Sentiment Analysis and Opinion Mining
Keywords
ChatGPT, large language models, textual annotation, crowd workers, Amazon Mechanical Turk, automatic annotation, inter-coder agreement, annotation cost, stance detection, topic extraction, frame detection, tweets, news articles, text classification

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

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