Readers/2024 Reader and Donor Experiences/Content Discovery/Wikimania 2024, "Written by AI" How do editors and machines collaborate to create content

We want to make it easier for people to learn from Wikipedia. This will require the usage of new technologies and ways of delivering content we have not previously explored - from generating article recommendations to using machine learning to remix or summarize various pieces of information. As we begin using these technologies, we will need to consider the different ways this content can be under the purview of editors and communities, with easy workflows and entry points for editing and moderation.

During the 2024 Wikimania, representatives of the Web team at the Wikimedia Foundation organized a session that discussed ways that AI/machine-generated remixing of our existing content can be used to make Wikipedia more accessible and easier to learn from, with the goal of helping a new generation of readers discover and learn from reliable, encyclopedic information on Wikipedia.  

Session summary

This session took part as a workshop with two parts - a presentation on why this work has potential and a work session with communities to brainstorm ways editors can be involved in controlling the output of this type of content.  

The presentation portion was focused on exploring the ways in which machine-generated content or suggestions can help readers and the ways communities and editors can be involved in controlling and contributing to this work.  We discussed the challenges that readers face on the current site, such as content presentation, availability, discoverability, and credibility of the content and ways that we can help with these challenges through using our existing content, with the help of AI.  We then presented a few ideas for ways machine generated content, content remixing, and recommendations can help. Central to this discussion was the idea that it’s crucial to us to maintain human oversight of any machine generated tool in order to keep Wikipedia’s quality and trustworthiness.

Results summary

The workshop portion of the session allowed editors to discuss ways to edit or otherwise control machine-generated content.  The results are available below. Overall:

1. Participant editors thought that the features presented had potential to improve the quality of the learning experience on Wikipedia
2. Participants agreed that some community/editorial control over machine generated content and suggestions would be preferable
3. Participants flagged a preference for solutions which did not require a lot of upfront work from editors
   1. For most groups, this meant adding the ability to edit or otherwise flag issues with machine-generated/remixed content once it was published or generated automatically on the wikis
4. Participants flagged a need for clearly marking which features AI was directly involved in

Presentation

[Slides]

Feature Ideas

[Images of each feature]

Results and Notes

Group 1, Read Next Feature

Do editors need to play a role in this feature?

Not by default - if there’s a see also, default to the see also.

• Top is machine generated, reading list is “see also.”

• If yes, when should they play this role?
  • Ability to curate

• If no, what guidelines should we provide to ensure that readers are not confused in the origin of what they are reading
  • Needs to show a “machine generated” label
• If editors should be involved, please design the steps (workflow)
  • When there is a list, show a button to edit the “see also” list.

Group 2, automatic article summaries

• Lead sections are already a summary, so is this an improvement? Might not be better than the human-created version.
  • One summary might highlight the wrong information
• Concerned the AI wouldn’t be able to distinguish between what’s important and what isn’t.
• If I’m a reader and I see a mistake in the AI summary - what can I do? ‘Report this as incorrect’
  • Concern that this would get misused to be about the article subject rather than the AI content.
  • Could use the feedback as a prompt into the AI. But open to abuse so a bad idea to enable public prompting. Maybe it posts to talk page, or other review list, but this could get very messy.
  • Or some communication if there is an error, resort to the main article.
  • -> Report reviewed by experienced editor to edit the prompt or content of the result.
    • There should always be a human check here.
• Editors adjusting the summary, but then you have to regenerate it each time some article content changes.
• Could have AI generated summary, then ‘lock’ it and make it editable to experienced editors to correct mistakes.
• Q: Would the AI summary be per-user or the same for every user?
• AI-generated speech of the AI summary would also be nice, and other options to change how it displays.

Workflow:

• Reader clicks ‘Report as incorrect’
• Writes an explanation of the problem
  • Categorisation - what kind of thing is wrong? Factual problem, style problem, etc. Helps patrollers prioritise
• ‘Your feedback has been received’
• Patroller sees this user feedback. Can approve/reject the feedback, or edit the article.
  • -> Insert negative prompt to the AI to handle this feedback?

Group 3, explore tab

Notes from the recommendation graph group

- All AI generated contents, even only remixing, should be clearly isolated from the original content

- If these AI tools use the user behavior as input (e.g. to customize recommendation), besides being compliant with privacy regulations it should also be easily possible to disable this

- Assuming that the editor community approves this feature, editors would not have any new role because these features showcase existing content which already has roles and procedures for moderation of content

An extra feature could be added to suggest possible missing connections, warning the user to take it with a grain of salt and giving them the ability to discard it.

Actors: Admin/Steward (chooses whether to enable the feature), Reader, Editor

Pre-step: involve the community before admin enables this feature

Pre-step: the editor opts in

1. Editor goes to article page

2. Editor click on button to get this suggestion

3. Editor sees the suggestion

4. The editor would have to manually evaluate if and how to add it

5. The reader sees the added info

Group 4, summaries

Group 4 recommended an idea in which readers could flag issues with AI-generated summaries, which would then be sent to a queue patrollers could check

Next steps

We plan on using the results of this exercise as the basis for our first proposals for patrolling and editing mechanisms for any future AI or machine assisted or generated features