Did AI-written petitions persuade more people to act?
A peer-reviewed natural experiment covering 1.5 million Change.org petitions found that access to an embedded AI writer changed language and increased similarity, but did not improve the engagement outcomes the researchers measured.
The short version
- What happened
- Researchers collected 1.5 million Change.org petitions; the main causal analysis used 277,496 English-language petitions during a staggered rollout across the United States, Great Britain, Canada and Australia.
- Who it affects
- Petition writers may receive faster drafting help, but a more polished petition did not translate into stronger early engagement in this study.
- What remains unproven
- The causal treatment is access to the embedded tool, not verified use or the proportion of generated text in each petition.
Keep in mind
The causal treatment is access to the embedded tool, not verified use or the proportion of generated text in each petition.
Australia serves as the delayed-access comparison during the main ten-week window; country-specific events or platform changes may still affect estimates.
Comments and ten-signature thresholds are early platform outcomes, not evidence of policy influence, campaign durability or democratic legitimacy.
Source analysis · 2 linked sources. Primary analysed source date: 30 September 2026. Earliest linked supporting source: 17 November 2025.
Read the full report · 5 min The evidence, methods, implications and limitations
The rollout created a rare real-world comparison
A peer-reviewed Nature Human Behaviour study published on 30 September examines what happened when Change.org embedded a generative-AI writing assistant into petition creation. The researchers collected 1.5 million petitions posted from 2022 to 2024. Their main difference-in-differences analysis focused on 277,496 English-language petitions and exploited a ten-week period in which users in the United States, Great Britain and Canada had access to the feature while Australia did not yet have the same access.
This design is stronger than simply comparing texts that an AI detector labels as generated. The staggered rollout creates treated and comparison trends, and the authors tested pre-intervention parallel trends, alternative time windows, synthetic controls and other specifications. They also examined 4,611 returning writers who posted before and after access changed. Even so, the treatment is access to the feature—not proof that every writer used it, accepted its suggestions or relied on it to the same degree. Off-platform AI use may also blur the comparison.[1][2]
Petitions became longer, more complex and more alike
Access to the AI tool was associated with a median increase of 55.15 words, about 45%, and a 1.51-grade rise in Flesch–Kincaid reading level, about 19%. Lexical diversity also rose. At the same time, mean pairwise similarity between petition embeddings increased by 0.055, around 9%, indicating that petitions became more alike at platform level even while individual texts used a broader vocabulary. The paper describes a shift from short, concrete verbs towards more formal terms such as implement and establish.
That combination matters for civic speech. A tool can make each item look more polished while narrowing the range of styles, local expressions and ways people frame a grievance. Homogeneity is not automatically harmful, and the study does not show that particular viewpoints were suppressed. But a shared model and interface can exert editorial influence at scale without a conventional editor. Platforms therefore need to measure linguistic diversity, specificity and user control, not only grammatical quality or completion speed.[1]
Better-looking text did not produce better measured outcomes
The language changes did not increase the share of petitions receiving at least one comment within 30 days, and the weekly volume of people creating petitions did not rise. In the main analysis, the share reaching ten signatures fell by 4.65 percentage points. Returning writers produced longer second petitions when AI was available but also recorded worse outcomes. These results challenge the assumption that lowering the effort needed to produce polished text will automatically broaden participation or improve mobilization.
The outcomes are deliberately modest indicators. Ten signatures and one comment show early engagement, not whether a petition reached a decision-maker, changed a policy, secured media coverage or built a lasting campaign. The design also cannot determine why outcomes failed to improve. Readers may distrust generic language; generated drafts may omit community knowledge; or writers may invest less effort in promotion after drafting. The study raises those mechanisms but does not directly test them.[1][2]
What platforms should test—and what would change our assessment
A civic platform considering an AI writing tool should preserve a clear human-only route, disclose when suggestions are generated and let writers reject or substantially revise them. Evaluation should include edit distance, time spent, writer ownership, linguistic diversity, sharing activity, repeat participation and real-world outcomes. A feature should not become the default merely because it produces longer or more formal text. Product teams also need rollback criteria if authenticity, trust or participation declines.
Our assessment would strengthen if later research records actual tool use and revisions, randomises access, measures distribution and policy effects, and repeats the analysis across languages and platforms. It would weaken if the result disappears under a different comparison period or if unmeasured ranking and promotion changes explain the outcome shift. For now, the evidence supports a specific conclusion: embedded AI access changed the form of civic writing much more clearly than it improved the early outcomes measured here.[1][2]