Data from: Durably reducing conspiracy beliefs through dialogues with AI

Costello, Thomas 1 ; Pennycook, Gordon 2 ; Rand, David 1

Published Aug 29, 2024; Updated Sep 29, 2026 on Dryad. https://doi.org/10.5061/dryad.v6wwpzh4h

Data files

Aug 29, 2024 version files 71.40 MB

Data_8.28.24.zip

71.36 MB
README.md

40.45 KB

Sep 29, 2026 version files 322.43 KB

ConspiracyAI-Correction-Data.zip

309.43 KB
README.md

13 KB

Abstract

Conspiracy theory beliefs are notoriously persistent. Influential hypotheses propose that they fulfill important psychological needs, thus resisting counterevidence. Yet previous failures in correcting conspiracy beliefs may be due to counterevidence being insufficiently compelling and tailored. To evaluate this possibility, we leveraged developments in generative artificial intelligence in two experiments whose primary analyses included 2044 conspiracy believers with complete pretreatment- and post-treatment-belief data after personalized evidence-based dialogues with GPT-4 Turbo. The intervention reduced conspiracy belief by ~20%. The effect remained 2 months later, generalized across a wide range of conspiracy theories, and occurred even among participants with deeply entrenched beliefs. Although the dialogues focused on a single conspiracy, they nonetheless diminished belief in unrelated conspiracies and shifted conspiracy-related behavioral intentions. These findings suggest that many conspiracy theory believers can revise their views if presented with sufficiently compelling evidence.

Thomas H. Costello, Gordon Pennycook, and David G. Rand. Related article: Costello et al. (2024), Science 385, eadq1814, https://doi.org/10.1126/science.adq1814

Existing Dryad record: https://doi.org/10.5061/dryad.v6wwpzh4h

Questions can be directed to Thomas H. Costello through the Dryad record contact information.

This is the data-only correction archive. It contains six deidentified analysis/flow CSVs and 28 tabular metadata/diagnostic CSVs. It contains no raw-format or recombined survey exports, frozen R objects, executable code, notebooks, model caches, plots, article text, or Supplementary Materials.

The complete package—including reconstruction inputs, frozen artifacts, code, caches, and publication materials—is available in the related Zenodo record, https://doi.org/10.5281/zenodo.22754984. Those materials are needed to reproduce the complete analyses.

Where to start

FILE_GUIDE.md gives the purpose of every included data/provenance file.
data/manifests/public_data_inventory.csv gives dimensions, unit of
observation, and purpose for the six authoritative deidentified tables.
data/manifests/public_data_dictionary.csv has 91 rows covering every
column in those six tables: data type, plain-language definition, units or categories, observed values/range, exact missing count and meaning, study availability, provenance, analysis use, and definition status.
MANIFEST.csv records file size and SHA-256; SHA256SUMS.txt is a command-line checksum list.

Study and file scope

Participants were assigned to an AI dialogue about a focal conspiracy belief or to an unrelated control dialogue. The data cover Study 1 and two survey waves pooled as Study 2. The corrected screened records are 748 in Study 1 and 1,346 in Study 2. The six files in data/deidentified/ are authoritative for the corrected analyses and flow summaries. The frozen cluster labels and aggregate AI-strategy counts needed by those public tables are included as CSVs; underlying reconstruction inputs and frozen R objects are in the separate complete Zenodo package. Use public_cluster_labels_deidentified.csv for corrected cluster labels.

Key codes in the public dictionary include
StudyNumber 1 = Study 1, 2 = Study 2a, 3 = Study 2b
ExperimentalCondition Treatment = focal-belief AI dialogue and Control = unrelated dialogue. Focal-belief and BCTI mean variables use 0–100 scales
Importance uses 0–8
-999 is the substantive “not applicable—I did not supply a theory” response, not missing.
public_response_id and public_participant_id are release-specific
pseudonyms
salts and crosswalks are withheld.
In the aggregated AI-strategy counts, None/Low/Moderate/High correspond to 0/1/2/3, and absent combinations have count zero.

Missing and blank values

The six data/deidentified/*.csv files use literal NA for unavailable values and contain no ambiguous empty data cells. The 91-row dictionary documents each variable's missing count and context. Depending on the field, NA may mean nonresponse, an unreached stage, absent linked follow-up or frozen derived value, structural inapplicability, unrecorded control subtype, or no applicable exclusion category. Zero is a valid observed value.

The included data/manifests/deidentification_manifest.csv summarizes field categories and public actions. Detailed transformation reports for raw-format and recombined reconstruction files belong to the separate Zenodo package, not this Dryad archive.

File formats and software

CSV files can be read with R, Python, or a spreadsheet program; preserve identifier strings and distinguish the literal missing-value token NA from zero. The complete Zenodo package supplies executable analysis and reconstruction instructions,renv.lock for the focused R analysis, recorded package versions, and cached model results. Reading these data does not require refitting causal forests or calling an external AI service.

*_counts.csv files contain counts or descriptive summaries; *_predictions.csv contain fitted values and confidence limits; *_model_results.csv contain coefficients and inferential statistics; *_model_metadata.csv document formulas, samples, inference, and sources.
In model outputs, estimate is the fitted coefficient, std_error its standard error, p_value its p value, and conf_low/conf_high (or conf.low/conf.high) the lower/upper 95% confidence limits. Observation counts are labeled n_obs or model_n; residual degrees of freedom are residual_df or df. Consult each file's formula, term, and inference fields rather than assuming the same model across all outputs.

In public data columns, pre and post denote pretreatment and posttreatment; followUp and followUp2 denote the Study 1 10-day and 2-month follow-ups. R1, R2, and R3 refer to dialogue rounds.

Human-participant privacy

These are deidentified, not anonymous, human-participant data. The authors completed Dryad's Human Subjects Information declaration. Direct contact, platform, network, precise-location, detailed timestamp, browser-identification, credential, and authorization fields were blanked or replaced; identifier salts/crosswalks are withheld.
data/deidentified/public_ai_strategy_rating_counts.csv contains aggregate strategy counts rather than conversation-level ratings. The raw-format and recombined surveys and the frozen conversation-level rating matrices are not included here; they are in the separate complete Zenodo package.
The included deidentification manifest summarizes the public-data actions. These controls reduce risk but cannot guarantee anonymity because unusual analytic combinations may remain distinctive. Do not attempt reidentification or linkage to external data.

Rights and citation

Data and associated data documentation in this archive are under
CC0 1.0 Universal, consistent with the existing Dryad license; see LICENSE. The Zenodo complete package preserves MIT for code and separate publisher-defined rights for publication materials. Please cite the article and Dryad record when using these data, even though CC0 does not require attribution.

Archive inventory

The deposit contains ConspiracyAI-Correction-Data.zip. Extract it to obtain the ConspiracyAI-Correction-Data/ folder. The archive has 39 files: 34 data/provenance CSVs, this README, a file guide, a CC0 notice, and two integrity lists. The original 2024 version remains in Dryad’s version history.

File-level guide

Every data/provenance file in this archive is listed below. MANIFEST.csv supplies byte counts and SHA-256 hashes. The two inventory/checksum files describe the archive itself and are not additional study observations.

Authoritative deidentified tables

Path under `data/deidentified/`	Purpose
`public_primary_analysis_deidentified.csv`	Screened participant-level analysis input (2,094 records, 54 variables).
`public_participant_screening_flow_deidentified.csv`	Response-level screening and attrition path, including screened-out test records.
`public_duplicate_attempt_selection_deidentified.csv`	Attempts and selections within repeated-participant groups.
`public_screening_flow_summary_deidentified.csv`	Aggregate reporting flow, row-level diagnostic flow, and separately labeled nonexclusive export diagnostic.
`public_cluster_labels_deidentified.csv`	Archived cluster labels joined by public response ID; authoritative corrected cluster labels.
`public_ai_strategy_rating_counts.csv`	Aggregate study × strategy × round × rating counts, without participant-level rating links.

Tabular metadata and derived diagnostics

All paths below are under data/manifests/. Figure tables are supporting machine-readable outputs, not image files.

File	Purpose
`public_data_dictionary.csv`	Definitions, coding, missingness, provenance, and uses for all 91 columns of the six public tables.
`public_data_inventory.csv`	Units, dimensions, and purposes of the six public tables.
`deidentification_manifest.csv`	Field categories, public actions, examples, and rationales for deidentification.
`participant_flow_assertions.csv`	Expected participant-flow counts and their source labels.
`primary_cluster_provenance.csv`	Provenance of the primary cluster-label construction and linkage.
`ai_strategy_rating_bridge.csv`	Field/value crosswalk for frozen strategy ratings.
`ai_strategy_rating_summary.csv`	Valid and missing counts plus mean/SD by strategy and round.
`prompt_pair_agreement_counts.csv`	Pairwise prompt-agreement counts and source provenance.
`figure_s2_counts.csv`	Study 1 screening-flow counts for Figure S2.
`figure_s3_counts.csv`	Dialogue word-count summaries by speaker, condition, and round for Figure S3.
`figure_s4_counts.csv`	Study 2 screening-flow counts for Figure S4.
`figure_s5_counts.csv`	Figure S5 complete-case counts by plausibility and condition.
`figure_s5_model_results.csv`	Figure S5 model coefficients and intervals.
`figure_s5_predictions.csv`	Figure S5 plotted fitted values and intervals.
`figure_s6_model_metadata.csv`	Figure S6 model specification and source metadata.
`figure_s6_model_results.csv`	Figure S6 model coefficients and intervals.
`figure_s6_plot_counts.csv`	Figure S6 observed item-response counts by condition, veracity, and belief.
`figure_s6_predictions.csv`	Figure S6 fitted values and intervals.
`figure_s7_cluster_counts.csv`	Figure S7 cluster sizes, shares, order, and labels.
`figure_s7_model_metadata.csv`	Figure S7 model specification and source metadata.
`figure_s7_model_results.csv`	Figure S7 cluster-model coefficients and HC2 intervals.
`figure_s7_plot_counts.csv`	Figure S7 plotted participant counts by cluster and condition.
`figure_s7_predictions.csv`	Figure S7 plotted fitted values and HC2 intervals.
`figure_s8_cluster_counts.csv`	Figure S8 cluster sizes/shares, archived subset counts, order, and labels.
`figure_s8_model_metadata.csv`	Figure S8 model specification and source metadata.
`figure_s8_model_results.csv`	Figure S8 cluster-model coefficients and HC2 intervals.
`figure_s8_plot_counts.csv`	Figure S8 plotted participant counts by cluster and condition.
`figure_s8_predictions.csv`	Figure S8 plotted fitted values and HC2 intervals.

Raw-format and recombined survey files, frozen reconstruction artifacts, and their detailed audit reports are available only in the separate complete Zenodo package, https://doi.org/10.5281/zenodo.22754984.

Human subjects data

Yes. This deposit contains deidentified data from human participants. Before participating, respondents were shown an MIT consent screen stating that participation was voluntary, that potentially identifying information would be removed before data were shared outside the research team, and that data with identifying information removed could be shared with other researchers and made available through online data repositories for future analysis. Respondents indicated agreement and that they were at least 18 years old by advancing past the consent screen. The studies were deemed minimal risk and exempt by the MIT Committee on the Use of Humans as Experimental Subjects (protocol E-5539). The deposited files exclude direct and platform identifiers, dates and timestamps, IP address and location/device metadata, free text, credentials, and detailed demographic combinations. Internal survey identifiers were replaced with release-specific salted join keys; source identifiers, the salt, and crosswalks are withheld. AI strategy ratings are released only as aggregate counts. We describe the files as deidentified rather than guaranteed anonymous because residual attribute-linkage risk cannot be ruled out.