2019/Grants/Wikipedia deployment of credibility signals app

From WikiConference North America
< 2019‎ | Grants
Revision as of 17:35, 5 April 2020 by Sandro (talk | contribs) (minor tweaks on a read-through)
Jump to navigation Jump to search


Title:

Wikipedia deployment of credibility signals app

Name:

Sandro Hawke

Wikimedia username:

Sandro_Hawke

E-mail address:

sandro@w3.org

Resume:

Geographical impact:

global

Type of project:

Technology

What is your idea?

We propose to provide the Wikipedia community with a web app which helps people collaborate in making credibility decisions. The app will help people articulate their knowledge about why and how much they believe information sources, while it also helps them use such articulations from others to improve their own decisions about what sources to believe and to cite. The underlying technologies are being developed in an open standards process at W3C and are being designed to be capable of rapidly growing to global scale. The vision is that after being refined and demonstrated at Wikipedia, the technology can spread across the internet for use by everyone who cares to separate fact from fiction for themselves and their community.

A few key points about this plan, not covered elsewhere in the proposal:

What are the signals of credibility?

We don't yet know which observable features of an information source and the world around it are the best indicators of whether the source is trustworthy. We also don't know if such features, when they are identified, can be generalized across topics and cultures and to what degree they are stable over time.

And yet, people are forced to make credibility decisions constantly in daily life (and in editing Wikipedia). To make these decisions, they end up performing their own personal observations, often in an ad hoc and unconscious way, and usually without documenting their process or using any scientifically validated method.

We propose to move this process several steps toward reliability and scientific validation. Our proposed tool will help people record the signals they find themselves using, while it guides the user based on some consensus advice from experts about how valid those signals might be.

This process of analyzing credibility signals has been sporadic in recent years, but has now re-emerged in the W3C Credible Web Community Group (see credweb.org). The overall strategy is laid out in Technological Approaches to Improving Credibility Assessment on the Web and a small set of partially analyzed signals has recently been documented in Reviewed Credibility Signals. We expect this process to continue and grow in a positive feedback loop with end-users.

With the added urgency of the COVID-19 pandemic, the group has most recently focused on perhaps the most fundamental credibility signal: how credible do other people find the source to be. This signal reveals the recursive nature of any kind of collaboration around trust, where if you’re going to rely on someone in making decisions about who to trust, you need to first decide who you can safely rely on. This sounds challenging, but again, is the kind of reasoning people do unconsciously every day. We’re proposing to add computerized assistance to the process, making it easier and more robust.

Why should people trust what others say?

The app does not do crowdsourcing, not in the usual sense. If Alice posts that Bob is highly credible, the app will in no way suggest to a particular user that Bob is credible unless it can first establish that that particular user trusts Alice, or (in some cases) statistically ought to trust Alice. The app merely reflects trust that people already have for each other, while suggesting sources that are statistically more likely to be trustworthy, based on its limited knowledge.

In other words, a user should trust what others have to say using the app because either (1) that user has specifically indicated they consider that other person credible, (2) the app is suggesting the other person is credible, based on evidence it clearly presents (such as which of their trusted people trust this person). Our intent is to create a cycle where people give the app feedback about its guesses, tweaking the source claims or the algorithms to improve its recommendations for themselves and others over time.

Why should people trust the app?

It's open source and open data (decentralized). The actual data people record in the app will be stored in their own web space. Outside of Wikipedia, that might be a social media account or a Google doc. For the Wikipedia community, we anticipate using Wikipedia user pages. This means when people record some credibility observation in the app, it will be written to one of their user sub-pages as a natural language sentence. It might be as simple as, “I consider https://www.washingtonpost.com/ to be credible” or more complex, like “The news website with its main page at URL http://thestar.com provides a corrections policy at URL https://www.thestar.com/about/statementofprinciples.html and evidence of the policy being implemented is visible at URL https://www.thestar.com/opinion/corrections.html” (example from https://credweb.org/reviewed-signals-20200224/#example-4).

As a result of this architecture, the app can be easily inspected for any security flaws or back doors. It can also be forked or re-implemented without any changes or effort on the part of the user community; there is no need for users to redo their work when switching to a competing app, or even to export and re-import data. The data simply remains in the user's web space, available to various apps much like it is available to various people.

How will you get people to use the app?

This is, of course, a fundamental issue with any new technology, product, or service. We can break it down into (1) How will people learn about it? (2) What will motivate them to try it? (3) How involved will they get during their initial experience? (4) Will they keep using it? and (5) Will they share and promote it to other people?

  • (1) How will people learn about it?

In general, we expect people to learn about the app from person-to-person contact, as people cite the app in discussing why something should be believed or disbelieved. Being a web app, the app will provide URLs that can be shared to make a point.

Disagreements about credibility observations may also spark the kind of discussion which is hard to miss. It will be important to try to frame these disagreements in a healthy way for users.

For seeding the initial user base, we intend to reach out to various wikipedians working on credibility-related projects to give the app a try. To make it useful on day 1, we can seed it with Perennial sources, credibility data from outside of Wikipedia (eg from NewsQ), and with data which proxies for credibility between wikipedians, like the stream of “thanks” events. Outside Wikipedia, we can seed it with data from social media, especially Twitter.

  • (2) What will motivate them to try it?

It seems likely that users will hear about the app with a framing that provokes curiosity. Trust is highly salient to people; they want to know who they can trust and if they themselves are perceived as trustworthy. Since the app is likely to come up in that context, many will find it hard not to take a quick glance.

There is some risk of people feeling like their fragile beliefs are being threatened when someone cites the app; we intend for it to be framed more around curiosity and neutral insight to help avoid this reaction.

  • (3) How involved will they get during their initial experience?

App design will be critical here, turning their “quick glance” into a chance to explore, contemplate, and begin to contribute. Drawing again on curiosity and the salience of trust, the plan here is to lead users from the issue that brought them onto the wider set of credibility data around related topics that may be of even more interest.

A bit like the drive to edit Wikipedia, when users see credibility data that seems wrong to them, the app will be welcoming them to improve it, as a service to themselves (if they choose to keep using the app) and to their community.

  • (4) Will they keep using it?

Users will come back because (1) they find the app useful in figuring out what to trust, (2) they want to help guide other people about what to trust, to avoid people they care about being hurt by misinformation, (3) to engage in the social feedback loop of seeing how others are weighing in on credibility topics they care about, and (4) potentially as a news source, offering a feed of items customized by assessments made within the user's credibility network.

  • (5) Will they share and promote it to other people?

Yes, we expect people will want to show the information in the app to others, and in doing so they will be showing the app. If they come to believe the app is helpful to them, they will presumably also advise others to use.

One general barrier to person-to-person spread of apps is that it can seem crass or commercial. In this case, the app being open source and open data, with a clearly important purpose, should help sidestep this point of resistance.

Why is it important?

For Wikipedia, this idea promises to help in the fight against misinformation, making it easier for wikipedians to collaborate among themselves and with the broader world in identifying credible and non-credible sources.

For the world at large, the stakes are much higher, as this approach has the potential to turn the tide against misinformation across all technology platforms.

Is your project already in progress?

We are developing the relevant concepts and tools (as seen at https://credweb.org) but have not begun deployment in the Wikipedia community or tooling to work with Wikipedia data feeds. Many elements of the app have been demonstrated, but not all of them, and not in an integrated whole.

How is it relevant to credibility and Wikipedia? (max 500 words)

There are many connections between this Credibility Signals work and Wikipedia:

  • Wikipedia has always needed to be able to separate fact from fiction. While it does this very well, these tools might make the task easier. Specifically, this can rapidly highlight which sources have unacceptably low credibility and help with sorting out why particular sources are viewed as credible or not credible. While this happens, data is accumulating (in public, under user control) about how credibility decisions are being made. That data can help guide future development and deployment of the technology.
  • Wikipedia has always needed to reduce harm done by careless and malicious users. It does this very well, but again, these tools might make the task easier, assisting in tracking and management of the reputation of users, which can be used in modifying their privileges.
  • Because of its great expertise in these fields, the Wikipedia community is an excellent proving ground for these technologies. Flaws in the technologies that might eventually lead to failure in the broader media ecosystem are likely to be spotted very quickly by wikipedians, giving time to improve the designs before wider deployment.

What is the ultimate impact of this project?

If successful, this project will show a clear way that people can collaborate online in protecting themselves and their communities from misinformation. This method can be adopted by communities and platforms around the world to greatly reduce misinformation and other online harms.

Could it scale?

Yes, this plan is phenomenally scalable. If it becomes fully established as a decentralized ecosystem, as designed, it will operate and grow with zero effort or support from us or Wikimedia.

It is based on existing social practices, where each individual manages their own credibility assessment process (deciding what to believe), using what they can glean from their surroundings, including their social network. This process scales linearly with the number of individuals, with each individual deciding how much of their own resources to devote to each assessment they make. Adding computers and networking to this existing human process should greatly improve the efficiency and accuracy of this process, without altering this scaling behavior.

In its approach to decentralization, this design avoids any central bottleneck. Every individual and organization is free to deploy as much human and computing resources as they choose, without needing approval or support from us or anyone else. This allows the kind of scaling to billions of users that we see in the web and email, which are similarly decentralized. If the system provides sufficient value to users, as we expect, this approach might grow to global scale in a matter of months.

Why are you the people to do it?

This funding request is to help support my time in leading and organizing this project and doing elements of the work for which I am unable to find volunteers or other funding. I bring experience and expertise in all the necessary challenge areas, including credibility signals, community development, web application development, decentralized systems, and consensus process.

What is the impact of your idea on diversity and inclusiveness of the Wikimedia movement?

This project has no direct connection to diversity or inclusiveness. We are committed to working to address any indirect impacts which might arise.

What are the challenges associated with this project and how you will overcome them?

In general, we are reducing risk in this ambitious project by minimizing complexity and using a progression of small prototypes and experiments.

Challenges include:

  • Getting people to look at credibility data. Approach: make it salient and visually appealing. For example, see credibility network demo at https://credweb.org/viewer/ which has elements that are compelling and fun; it becomes salient when we let people add in the sources they care about and get to see how others judge those sources. We can bootstrap with existing Wikipedia data feeds of likes and reverts as an initial proxy for credibility between wikipedians and draw on existing source credibility work for data on external sources.
  • Getting people to author credibility data. Once people are engaged in the data as a consumer, we hypothesize they will be motivated to engage as a producer to "correct" the data, to express what they believe or know. Additionally, a culture of contributing data to help the world, already common among wikipedians, should help. There are a range of ways to simplify or even gamify the contribution step, if necessary.
  • Harmful participants. Since we propose to primarily and initially use credibility data hosted on Wikipedia user pages, to some degree the existing community safety mechanisms will still apply. We would like to demonstrate, however, that such mechanisms can be largely replaced by credibility data itself. In theory, people observed to do harm can be identified and have their actions demoted like non-credible content.
  • Getting people to trust the system. Approach: transparency and feedback. Make it clear which individuals are the source of each bit of data, with clear provenance and change tracking. Have the interface promote a virtuous cycle of improving the data and improving one's own credibility. This is similar to Wikipedia's own mechanisms for being trustworthy (to people who know how it works).

How much money are you requesting?

10k USD for the Wikipedia aspects (outlined here) of the Credibility Signals work

How will you spend the money?

To support my time on this work

How long will your project take?

Up to 12 months, in three phases:

  • Phase 1 - up to four months - refine deployment plan, identify partners, settle issues within credweb CG
  • Phase 2 - about 2 months - active development of tools; release
  • Phase 3 - up to six months - revise and improve, based on user experience

Have you worked on projects for previous grants before?

Yes, my work has been primarily grant funded for many years. Some highlights with web pages maintained by others:

Retrieved from "https://wikiconference.org/index.php?title=2019/Grants/Wikipedia_deployment_of_credibility_signals_app&oldid=15750"