Growth/Revise Tone
|
Improve Tone
Structured Task to support newcomers learning about and improving the tone of Wikipedia articles.
|
The Revise Tone Structured Task helps newcomers identify and revise non-neutral language in Wikipedia articles. It is designed to surface in the Suggested Edits feed on the Newcomer Homepage, alongside other Structured Tasks like Add a Link and Add an Image.
This Structured Task builds on the Growth team’s broader strategy to lower editing barriers while supporting contributor development. It aims to increase constructive edits from new editors and deepen their understanding of Wikipedia’s expectations for tone and neutrality.
Unlike Add a Link or Add an Image, Revise Tone is less structured and provides a more open-ended editing experience. It introduces more advanced skills and supports newcomer progression by helping them internalize the core principles of encyclopedic writing.
Current status
[edit]- : Early planning and initial designs
- : Machine Learning & technical planning. Presenting early ideas at Wikimania 2025, in Nairobi.
- Next: Gather community feedback and conduct usability testing and model evaluation
User goals
[edit]Newer volunteers often struggle to begin editing successfully, especially on mobile where screen space is limited and attention is fragmented. Some are discouraged by the trial and error process many go through while learning to contribute constructively, while others have yet to encounter a compelling reason to try. This task helps address those challenges by:
- Surfacing relevant, achievable editing opportunities
- Encouraging edits that are less likely to be reverted
- Supporting newcomer growth through practice with advanced editing skills like tone refinement
- Building awareness of Wikipedia’s content and style policies
Tone Check model
[edit]The Revise Tone Structured Task will be powered by a machine learning model trained on community-driven signals. This is the same model that powers the Tone Check Edit check.
The model works by being fine-tuned on examples of Wikipedia revisions. It learns from instances where experienced editors have applied a specific template ("peacock") to flag tone violations, as well as instances where that template was removed. This process teaches the BERT model to identify patterns associated with appropriate and inappropriate tones based on Wikipedia's editorial standards.
Another layer of editor-powered logic that will help guide this Structured Task is that rather than scanning all articles, we will focus on a subset of articles where editor signals suggest potential tone concerns. Selection criteria for which articles will be considered is still under discussion, but may include maintenance templates, lower content quality rating, limited edit history, or article topics that often contain promotional language (e.g., business-related articles).
Although both Tone Check and Revise Tone features use the same machine learning model, their purposes and audiences differ: - Edit Check: Provides real-time feedback to contributors during the editing process. Its goal is to help editors recognize issues such as tone, citation, or formatting concerns before saving their edits. - Revise Tone Structured Task: Surfaces targeted opportunities for newer contributors who are looking for ways to get involved.
How it works
[edit]This task applies the Tone Check model, a Machine Learning system trained to identify tone issues such as promotional language and puffery. Suggestions are also informed by community-maintained templates (e.g., {{POV}}, {{Peacock}}) and data from reverted edits to help pinpoint problematic phrasing.
When a tone issue is detected, the Edit Check UI will highlight the relevant sentence and offer in-context guidance to help newcomers revise the text to align with Wikipedia’s expectations for neutral, encyclopedic tone. This task will not propose specific language, but simple provide policy guidance and highlight areas of text that may need changes.
- Highlighting of problematic tone, with policy guidance
- Designed mobile-first, with plans to support desktop
- Delivered through the Suggested Edits feed on the Newcomer Homepage
- Less guided than other Structured Tasks, offering newcomers a progressive way to move to incrementally more challenging tasks
- This task will be configurable via Community Configuration, so it can be customized based on local consensus and needs
Designs
[edit]We plan to build on familiar UI elements, interaction patterns, and workflows from Growth features and Edit Check, prioritizing orchestration of known components over new designs.
The task will use the Suggested Edits framework, offering suggestions from the Newcomer homepage. Newcomers will receive brief onboarding, then transition into the Edit Check interface. The primary new design Growth is developing is the onboarding experience.
We are testing three onboarding approaches:
- Quiz-style onboarding: Offers a low-stakes way to practice the Revise Tone concept before editing in VisualEditor, helping reduce newcomer anxiety about making mistakes.
- Video-style onboarding: Uses a short simulation to preview what newcomers will encounter in VisualEditor, drawing from past findings that video guidance was the most engaging form of help content.
- Skip onboarding (A/B test): Compares the no-onboarding experience against the two approaches above. This model provides “in the moment” guidance once an editor encounters their first suggestion, rather than requiring learning upfront.
-
Quiz-style onboarding -
Video-style onboarding -
Revise Tone: Accept or Decline suggestion
An early Revise Tone design prototype is available to test via Figma.
We will gather early feedback on the designs through community discussions on our pilot wikis and user testing with newcomers. Additional feedback is welcome on our Talk page.
Usability test
[edit]After an internal round of reviews, we decided to test the quiz-style onboarding. We were curious whether it might better support newcomers by actively engaging with them, and inviting participation, rather than asking them to passively absorb information.
During the week of August 25, 2025, we used Userlytics.com to conduct five tests of the mobile onboarding prototype with people who had between 10 and 100 edits on Wikipedia. Our testing goals were:
- Observe if people identify tone issues without guidance
- Do people understand "improve tone" (now "revise tone") as a concept?
- Gauge people's reactions to the quiz-based onboarding experience
- Do people feel equipped to revise tone after completing the onboarding?
Summary of findings
[edit]- Neutrality was not intuitive. Only 1 out of 5 participants identified the tone issue "great."
- "Improve tone" was ambiguous for some participants, while the general suggested edit task was well understood. Participants described the task as:
- Simplifying and clarifying text
- Avoiding bias/promotional tone
- Almost no participants opened the info drawer
- All participants completed the tutorial:
- 3 out of 5 found it helpful
- 2 out of 5 thought the tutorial was the actual task
- 3 out of 5 participants felt better prepared to identify tone issues after the quiz-based onboarding:
- "It definitely helped me prepare for the task"
- "If I were totally new to Wikipedia and wanted to contribute, this would be super helpful because if I hadn't received this tutorial, I would probably not be sure what's wrong with these paragraphs or these sentences because grammatically they're correct"
Recommendations
[edit]- Label the tutorial more clearly
- Clarify language and feature purpose
Hypotheses and Success Criteria
[edit]Machine Learning Hypothesis (WE1.1.8)
[edit]If we apply the Tone Check model to published articles, we will learn whether we can identify ≥10,000 tone issues (each with a probability score of 0.8+) to build a high-quality (≥70% accuracy) suggestion pool.
Growth Hypothesis (WE1.1.2)
[edit]If we deliver an initial beta version of the Revise Tone Structured Task, we can evaluate whether the Edit Check framework can technically support proactive suggestions launched from the Suggested Edits feed.
Success criteria (for early work in Q1)
[edit]- Technical feasibility and extensibility
- Implemented efficiently using the Edit Check framework
- Design supports expansion to future Structured Tasks
- Model accuracy and appropriateness
- Model suggestions reach at least 70% accuracy based on a human review (by experienced editors)
- Recommendations align with Wikipedia’s neutrality standards
- Task clarity and usability
- Newcomers understand the task and feel confident completing it
- Community acceptance
- Input collected through discussions and Ambassador feedback
- General support from experienced contributors
Community feedback
[edit]Outreach efforts at Growth pilot wikis will seek input from communities during early planning and beta testing phase to understand support, concerns, and configuration preferences.
Our pilot wikis we will work closely with during this time will be Arabic, English, French, and Spanish Wikipedias.
If you have feedback, please chime in on the Talk page.
Related projects
[edit]This work aligns with the longer-term Contributor Strategy and fits within the WMF 2025-2026 Annual Plan's Contributor Experience Objective, specifically the Wiki Experiences 1.1 Key Result.
Related research
[edit]- 56% of new content edits by newcomers included peacock terms
- Edits with peacock words were 46.7% more likely to be reverted
- 22% of these edits were reverted
Impact of Newcomer Tasks and Structured Tasks
[edit]- “Newcomer Tasks” increase first article edit probability by +11.6% (Experiment analysis, November 2020)
- “Add a Link” improves first edit rate by +16.6%, and retention by +16.2% (Experiment analysis, December 2021)
- "Add a Link" on English Wikipedia increases constructive activation by +33.7% (Add-a-link Experiment on English Wikipedia)
- “Add an Image” boosts first edit rate by +17.0%, and retention by +24.3% (Experiment analysis, March 2024)