AI Polished Petitions Without Improving Early Support

A large natural experiment on Change.org found a sharp divide between better-looking text and better early support. After an in-platform AI writing assistant became fully available in the United States, Great Britain and Canada—but not yet Australia—petitions became longer, more lexically varied, more complex and more similar to one another. Access did not improve either of the study’s two early engagement thresholds, and the share reaching 10 signatures declined.

That distinction matters for human agency. AI can reduce the friction of drafting. It cannot, by itself, supply the relationships, specificity, trust or follow-through that turn a statement into collective action.

Research references: [1] [2]

A rollout created an unusual comparison

Change.org introduced an AI-generated drafting tool built on an OpenAI-provided language model. The full rollout reached the United States, Great Britain and Canada 11 weeks before Australia. Researchers used that staggered access from October 2 through December 15, 2023, in a difference-in-differences design.

The full archive contained 1,517,433 petitions posted from January 1, 2022, through December 31, 2024. The main causal analysis was narrower: 361,635 English-language petitions from the four countries that met location, timing and minimum-length rules. Its aggregate estimates compared 237,020 pre-access petitions with 42,111 petitions from the 11-week unequal-access period, excluding a 26-week A/B-testing phase because individual exposure during that phase was unknown.

This is stronger than a simple before-and-after comparison. The authors report parallel pre-rollout trends and multiple robustness checks, including synthetic controls. But the treatment was access to the tool, not verified use by every writer. The estimates therefore describe a platform-level intervention, not a universal effect of AI-authored petitions or proof about what happened to each individual user.

Research references: [2]

The writing changed quickly

Across the unequal-access period, the estimated median petition length rose by 53.63 words from a pre-AI baseline of about 110 words—a 49 percent increase. Median Flesch–Kincaid grade level rose by 1.49 grades, meaning the text became more complex, not necessarily easier to understand. Lexical diversity rose by 0.049 units, about 6 percent from its baseline.

A separate descriptive rating exercise found higher average writing-quality and persuasiveness scores among petitions created with access to the tool. That exercise does not establish that the tool caused every individual improvement, but it supports the narrower observation that the post-access text looked more polished.

The writing also became more alike. Mean pairwise text similarity within country-week groups increased by 0.056 from a pre-AI baseline of 0.240, about 23 percent. The authors added this homogeneity analysis beyond the preregistered core design.

That tradeoff deserves attention. Assistance can help more people clear the blank page while also nudging many voices toward a shared style.

Research references: [2] [4]

Better prose did not improve the measured support thresholds

The study used two deliberately modest outcomes: whether a petition received at least one comment within 30 days, and whether it had reached at least 10 signatures by the time the data were collected. Only the comment outcome was time-boxed because signature timestamps were unavailable.

Access was associated with a 2.51-percentage-point decline in the comment threshold, but the 95 percent confidence interval crossed zero, so that estimate was not statistically distinguishable from no effect. The share reaching 10 signatures declined by 5.33 percentage points, with a 95 percent confidence interval from −8.21 to −2.45 percentage points.

The responsible conclusion is specific: the study found no evidence that access improved either measured early outcome, and it found evidence of a decline in reaching 10 signatures. Those measures do not capture ultimate policy change, long-term mobilization or every way a petition might matter.

Research references: [2]

The mechanism remains unresolved

The paper discusses several possibilities. More generic language might remove useful specificity. Readers might discount text they suspect was machine-generated. Writers might feel less ownership and invest less effort in sharing. Or stronger prose might leave the surrounding social network unchanged.

These are hypotheses, not findings. The researchers could not observe off-platform amplification, private sharing, news coverage or the full set of in-platform promotion decisions. Their public data also could not identify the causal mechanism connecting access to outcomes.

An analysis of repeat petition writers pointed in the same direction but was not a second randomized or difference-in-differences experiment. Across 4,611 returning writers, those whose second petition came after access wrote substantially longer second petitions; their adjusted odds of receiving a comment were lower than for their first petitions. The authors explicitly discuss regression to the mean and present the analysis as supporting consistency rather than clean individual-level causation.

Research references: [2]

Design for agency, not just fluency

The practical lesson is not that civic platforms should remove writing assistance. It is that drafting support should preserve the human work that makes a petition credible and actionable.

A useful system can treat generated prose as a scaffold. It can prompt writers to add the local facts, affected people, requested remedy and personal reason that make a campaign specific. It can show alternatives, preserve easy editing and make it clear that the writer—not the model—owns the final claim.

Support can also extend beyond the sentence. Prompts for trusted first readers, evidence, respectful outreach and transparent updates may connect a polished draft to the community around it. Those are design implications, not interventions tested by this study.

Evaluation should also reach beyond surface quality. Readability scores, word counts and persuasive ratings are intermediate measures. Product teams should separately examine whether a tool broadens participation, preserves distinctive voices and helps people follow through.

What this study does—and does not—show

The paper was published in *Nature Human Behaviour* on September 30, 2026. The authors preregistered the main difference-in-differences analysis, shared petition metadata and analysis code through OSF, and withheld full petition text and user names from the public release for privacy, agency and copyright reasons.

Cornell’s institutional review office deemed the main public-data analysis not human-subjects research; the labeling task was exempt under protocol IRB0150383. The authors reported no specific funding for the work and declared no competing interests.

The study does not show that AI made every petition worse, that longer writing caused lower support, or that the tool reduced final political impact. It shows something narrower and more useful: changing the text is not the same as changing the social outcome.

Human agency is not merely the ability to produce competent language. It includes deciding what matters, standing behind a claim, persuading particular people and sustaining action after the draft is finished. AI can help people find words. Communities still determine whether those words become consequential.

: Corpus, Gilbert, Koenecke and Naaman, “Introducing AI to an online petition platform changed outputs but not outcomes,” *Nature Human Behaviour*, published September 30, 2026. https://doi.org/10.1038/s41562-026-02580-8 : Author manuscript and supplementary material, revised July 29, 2026. https://arxiv.org/abs/2511.13949 : Study metadata and code repository. https://osf.io/q3hj5/?view_only=3fc5b89d05eb4741957af431b107cf2a : Preregistered difference-in-differences analysis. https://aspredicted.org/xtcd-tvqk.pdf

Research references: [1] [3] [4] [2]