Agenda 2025-06-10 trustnet

[TzviyaSiegman]

trustnet

Speakers: David Karger and Farnaz Jahanbakhsh with Jeffrey Yasskin and Rick Byers

• Overview of trustnet
• Responses to questionnaire
• Discussion

Meeting information
Minutes

[TzviyaSiegman]

See responses to questionnaire

[jyasskin]

@karger is going to give the overview, which will be similar to his TedX talk

He and @farnazj have published several articles on the system:

• Leveraging Structured Trusted-Peer Assessments to Combat Misinformation
• A Browser Extension for in-place Signaling and Assessment of Misinformation
• Our Browser Extension Lets Readers Change the Headlines on News Articles, and You Won't Believe What They Did!
• Exploring Lightweight Interventions at Posting Time to Reduce the Sharing of Misinformation on Social Media
• Exploring the Use of Personalized AI for Identifying Misinformation on Social Media

And you can try it out at https://trustnet.csail.mit.edu/.

[lrosenthol]

Trustnet is based on the "web of trust" model, similar to that popularized by PGP (back in the day). We've learned since then about the pros and cons of that model, so much so that there has been an ongoing "Rebooting the Web of Trust" event (RWOT, https://www.weboftrust.info/) for over a decade, which hasn't actually been successful in doing so. Can you speak to how you see your approach is different than those (if at all), and why your approach would be successful?

Relatedly, with social networks, users have already established their "social graphs" which represents people (or organizations) that they are interested in in some way. These graphs are similar, though not 100% equivalent to "webs of trust" or to the network that you wish to create - but yet users already have invested time in creating them. Can you give some insight on how you see them interacting, if at all?

[lrosenthol]

One thing I don't see mentioned is the mechanism for identity. Is it email addresses? It is something I get when I register with the system? Other?

The example site seems to require registration, which would seem to then limit one to a centralized approach along with the possibility for people squatting on desirable IDs...

[karger]

One thing I don't see mentioned is the mechanism for identity. Is it email addresses? It is something I get when I register with the system? Other?

The example site seems to require registration, which would seem to then limit one to a centralized approach along with the possibility for people squatting on desirable IDs...

Yes, the Trustnet prototype manages its own identity model. This is not the right option for true deployment, but was a useful shortcut that let us quickly develop a system to explore how a database of trust and assessments could be used cope with misinformation. The focus of Trustnet was to explore/demonstrate the user experience, and we took a number of shortcuts to provide that experience. Trustnet can help us identify requirements for a suitable standard for identity and assessment.

[karger]

Trustnet is based on the "web of trust"( model, similar to that popularized by PGP (back in the day). We've learned since then about the pros and cons of that model, so much so that there has been an ongoing "Rebooting the Web of Trust" event (RWOT, https://www.weboftrust.info/) for over a decade, which hasn't actually been successful in doing so. Can you speak to how you see your approach is different than those (if at all), and why your approach would be successful?

It's an exaggeration to say Trustnet uses a Web of Trust (WoT) The idea of a WoT is to start with individual trust relationships and then let trust propagate through the network, creating a kind of transitive closure. Trustnet doesn't do that: the only trust relationships used to determine credibility of content for a user are that user's own direct trust relationships. We recognized that trust could be propagated---that for example you might want some content to be treated as accurate because it was so labeled by someone who is trusted by someone you trust---but we chose not to implement that in our initial version because we did not want to to tackle that important/interesting/difficult question of exactly how to do that trust propagation.

Instead the key characterization of Trustnet is simply that it allows every individual to make their own independent decisions about who they want to trust, and uses those personalized trust decisions to deliver personalized assessments of credibility to individual users.

This makes Trustnet orthogonal to the WoT. The original PGP WoT and the RWOT project focus on using a trust network to certify identities. It is an identity system. As discussed in my previous response, Trustnet is not an exploration of identity systems, but rather an exploration of how personalized choices of trust in identities (however defined and certified) and assessment by those identities can be used to help every individual create their own personalized misinformation filters.

Another distinction is that WoT is seeking a (physically) decentralized architecture, presumably to ensure that no one identity has the power to take over the identity system. Trustnet is agnostic on the question of physical decentralization. Insteat make a case for individual autonomy, a kind of logical decentralization, in choices of who to trust. That autonomy can be provided on a centralized platform (and the Trustnet prototype is in fact centralized). Of course a centralized platform would have great power to manipulate/ignore people's trust relationships if it were malicious, so there might be good reason to decentralize it, but that is a design question we chose not to explore as we wanted to focus on the user experience.

Relatedly, with social networks, users have already established their "social graphs" which represents people (or organizations) that they are interested in in some way. These graphs are similar, though not 100% equivalent to "webs of trust" or to the network that you wish to create - but yet users already have invested time in creating them. Can you give some insight on how you see them interacting, if at all?

We discuss this some in the trustnet paper. It is true that what we call the "follow network" defined by people's follow/friend relationships is a network just as the trust network is a network, but one cannot replace the other. I venture that nearly all of us follow people we do not trust. Examples might include crazy relatives or old classmates with whom we want to stay in touch, or politicians or other influential people who we distrust but whose opinions/statements are worth paying attention to because they impact our lives. On the other hand, there is a variety of sources that we might trust but not follow---we would like to benefit from their opinions about what information is true or false, but we are not generally interested in what they have to say. This might include experts on various topics (who post all sorts of random content), people who are trustworthy but boring, fact checking organizations, or news sites that are redundant with other news sites we follow.

The follow network might be useful as a bootstrapping mechanism for the trust network---it seems plausible that a large portion of a user's follow network is in fact trusted, so we could walk the user through the follow network to help them check of trusted sources for their trust network. But the two networks are independent of one another.

[lrosenthol]

@karger wrote:

The focus of Trustnet was to explore/demonstrate the user experience

So then should that be our (primary/only) focus when reviewing and/or discussing it? If so, that's fine, but knowing that up front will help me (at least) continue to focus where I invest my evaluation.

[lrosenthol]

Schema.org has a series of well defined grammars for "reviews", such as Claim Review that can be incorporated into a web page via Microdata (see https://schema.org/docs/gs.html)/

During your work on TrustNet, did you evaluate how one might leverage that work with yours?

[karger]

@karger wrote:

The focus of Trustnet was to explore/demonstrate the user experience

So then should that be our (primary/only) focus when reviewing and/or discussing it? If so, that's fine, but knowing that up front will help me (at least) continue to focus where I invest my evaluation.

Certainly not only. And in this group probably not even primary. The shortcut choices we made can provide a good jumping off point for discussing how to do things well. But I don't want to drill so far into the implementation details that we fail to discuss whether Trustnet's approach of explicitly surfacing trust and empowering individuals is a good one.

[karger]

Schema.org has a series of well defined grammars for "reviews", such as Claim Review that can be incorporated into a web page via Microdata (see https://schema.org/docs/gs.html)/

During your work on TrustNet, did you evaluate how one might leverage that work with yours?

We discussed the using claimreview but did not have the resources to do so. One could imagine various ways to incorporate claimreview's richer assessment of claims into trustnet, but you would also have to think about how to convince sources to author those richer assessments. What we found most potentially useful in claimreview hasn't been fully realized: the possibility of standardizing the claims being reviewed. Suppose that for a given news event, it were possible to publish certain standard "key claims" regarding that news event, and ithat news organizations were to use claimreview in their articles to indicate which of those key claims their own articles were actually supporting. Then trustnet could enable facilitators to asses those standard claims instead of assessing articles. And those assessments could be surfaced by trustnet in an article that was supporting those claims, even if nobody had yet assessed the article. This would make assessments more powerful/generalizable. It's not clear, though, how to make this robust: someone trying to disseminate misinformation could simply choose not to link to the standardized claims being made in the article. Even so, this approach suggests factoring trustnet into two layers, one concerned with determining, in a trustworthy way, which standardized claims are being made in articles, and a separate one concerned with capturing facilitators' assessments of those standardized claims.

[lrosenthol]

you would also have to think about how to convince sources to author those richer assessments.

No matter what assessment(s) are possible, getting humans/organizations to author them will always be the hardest part. (and let's not even dive into what happens when agents start leaving them!)

What we found most potentially useful in claimreview hasn't been fully realized: the possibility of standardizing the claims being reviewed.

Hmm...Curious why you think it is necessary (or even useful) to standardize the claims themselves? I can see things more where different types of uses for trustnet would use different aspects of the assessments - for example, professional news publishers vs. product reviewers. These are both important scenarios where trusting others is important - but the reviewers are clearly leaving different signals!

Even so, this approach suggests factoring trustnet into two layers, one concerned with determining, in a trustworthy way, which standardized claims are being made in articles, and a separate one concerned with capturing facilitators' assessments of those standardized claims.

Agreed, that depending on how deep you take this, you can indeed get into a tree/graph where you now need people who are evaluating the evaluations....but isn't that how things work in the physical world today, with multi-level reviews?

[agreiner]

Since an information source may be trustworthy in some domains but not in others, have you considered assigning trust by domain rather than a blanket trust/do-not-trust assessment?

[karger]

Since an information source may be trustworthy in some domains but not in others, have you considered assigning trust by domain rather than a blanket trust/do-not-trust assessment?

Yes, it seems clear that in certain circumstances you want trust by topic. However, since we were trying something novel, we wanted to begin with the simplest possible approach before trying something more complicated. In my more speculative moments, I wonder if we can establish a culture among a certain subset of society of people who only opine on things where they actually have knowledge-, so can be trusted whenever they opine---something like the "fair witness" of Stranger in a Strange land. But I know that's far-fetched.

[RByers]

On the Chrome team we've been thinking about the misinformation problem for years. There are lots of great ideas but we always get hung up on the problem of ensuring trust facilitation is durably aligned to user interests and able to reliably resist capture by special interests. Just managing trust in TLS certificates has really been a huge challenge for browsers (keeping CA behavior aligned to user security interests), and that's a tiny fraction of the difficulty and significance of the problem we're discussing here.

For last TPAC @jyasskin and I came up with a description of the properties we think we'd need to endorse and support a content trust system in Chrome. Then after the first W3C authentic web workshop we were thrilled to learn that David and Farnaz had built exactly that sort of system in TrustNet. While there are lots of details which could vary, we believe that having a framework with these core properties of trust network facilitation and user empowerment is a prerequisite for us to add any prominent UI to Chrome around content trust signals.

I wonder if perhaps existing provenance technologies like C2PA could be combined with a credibility facilitation network system like TrustNet to put users in control over how the C2PA signals are interpreted and leverage in their browsing UX?

[karger]

I wonder if perhaps existing provenance technologies like C2PA could be combined with a credibility facilitation network system like TrustNet to put users in control over how the C2PA signals are interpreted and leverage in their browsing UX?

Provenance is the essential other side of the coin to trust and the side that we shortcut in trustnet. I think of provenance as a way of embedding, in the content itself, the identity of a facilitator who is asserting the accuracy of the content. Provenance work thinks about all the things we skipped it trustnet, like how to carry that "endorsement" along as the content migrates and mutates. I am eager to explore ways to combine that with trust.