Paper Breakdown: When AI Writes the Petition, Nobody Signs It Faster

Executive summary

• The paper: Corpus, Gilbert, Koenecke and Naaman, "Introducing AI to an online petition platform changed outputs but not outcomes," Nature Human Behaviour (published September 30, 2026).
• The question: when a civic platform gives users an AI writing assistant, does the advocacy become more effective?
• The answer: no. The tool changed the text substantially, making petitions longer and more uniform, but engagement did not improve, and on one measure it declined.
• The lesson for leaders: AI can multiply the volume and polish of messages. It does not, by itself, produce persuasion, trust or mobilization.

Why this study matters

Political campaigns, NGOs and corporate communication teams are adopting generative AI to draft statements, emails, posts and appeals at scale. The implicit assumption is that better-written messages will perform better. Until now, most evidence on that assumption has come from laboratory tasks or short online experiments.

This study is different. It examines a real platform, real users and real outcomes, over a period when an AI tool was introduced without the users being part of a research project. That makes it one of the most direct tests so far of what happens to civic communication when writing is partly automated.

The design

A natural experiment

• Platform: Change.org, one of the world's largest online petition sites.
• Intervention: an in-platform "write with AI" tool that helped users draft petitions.
• Rollout: the tool reached the United States, the United Kingdom and Canada first, and Australia about 11 weeks later, during the last quarter of 2023.
• Data: roughly 1.5 million petitions.

The method

The staggered rollout allowed a difference-in-differences design. The researchers compared how petitions changed in countries that received the tool against a country that had not yet received it, under the assumption that all four followed similar trends beforehand. A second analysis followed repeat petition writers, comparing the petitions they wrote before and after the tool appeared. Both approaches pointed in the same direction.

What they measured

• Text features: length of titles and bodies, lexical diversity, readability and similarity between petitions.
• Outcomes: early signs of traction, such as whether a petition received comments and reached a minimum number of signatures.

The findings

1. The writing changed. Petitions written after the tool arrived were longer and used different vocabulary. Titles began to feature verbs such as "implement," "mandate" and "urge," which had been rare before.
2. The writing converged. Petitions became more similar to one another, especially in their titles. The authors describe this as homogenization.
3. The outcomes did not improve. Petitions did not gain more support. According to Cornell's summary of the study, the share of petitions receiving at least one comment within 30 days fell by about 5 percent relative to the pre-tool baseline.

The authors offer possible explanations, while noting that they remain hypotheses: AI-assisted petitions may be less specific; readers may discount text that feels generic; and writers may feel less ownership of a plea they did not fully write, and share it less.

The theoretical frame

The results connect to two lines of research.

The first is work on connective action. Bennett and Segerberg (2012) argued that digital mobilization often spreads through personalized expression shared across personal networks rather than through formal organizations. If that is how petitions travel, a more polished but less personal text may weaken the very mechanism that moves it.

The second is emerging evidence on AI and homogenization. In an experiment on short-story writing, Doshi and Hauser (2024) found that access to generative AI improved individual stories but made the collection of stories more similar. The Change.org study suggests the same trade-off at platform scale: each message may look better, while the overall landscape becomes more uniform and less distinctive.

Limitations worth noting

• The design depends on the assumption of parallel trends across countries before the rollout.
• The data cover English-speaking countries and a single rollout in 2023, early in the adoption of AI writing tools.
• Outcomes were measured through a small number of engagement indicators.
• The study measures access to the tool, not exactly how each writer used it.

Lessons for senior leadership

1. Do not confuse output with impact. More messages, produced faster, are not the same as more persuasion. Measure results, not volume.
2. Protect specificity. Generic language is the default output of a general-purpose model. Concrete details, names, places and stakes are what make a message credible.
3. Watch for sameness. If every organization uses similar tools, messages converge. Distinctiveness becomes a competitive advantage.
4. Keep ownership human. People share what they believe in and helped create. Use AI to draft, but make sure the final message is recognizably yours.
5. Test before scaling. Before automating a communication channel, run a controlled comparison against human-written messages.

The study's conclusion is a useful corrective for any team that equates efficiency with effectiveness. Artificial intelligence can change what we say. Whether people listen still depends on trust, relevance and the sense that a real person is behind the message.

References

Bennett, W. L., & Segerberg, A. (2012). The logic of connective action: Digital media and the personalization of contentious politics. Information, Communication & Society, 15(5), 739–768. https://doi.org/10.1080/1369118X.2012.670661

Corpus, I., Gilbert, E., Koenecke, A., & Naaman, M. (2026). Introducing AI to an online petition platform changed outputs but not outcomes. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02580-8

Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290