The Web Needs a Context Layer Built on a Shared Protocol

Two people can read the same headline and come away with opposite stories. One may see a public health measure, the other government overreach. Researchers have long known that communities don't just disagree on issues; they frame them in entirely different terms. The problem is not disagreement itself. A diverse society will always contain reasonable, competing interpretations of the same event. The issue is that the web typically gives users a single spotlight on a topic without making the surrounding perspectives easy to find. A claim can be accurate but still partial. What is missing is a way to see those competing frames side by side, mapping how the same conversation takes shape across the internet. This would give readers a broader view.

When a misleading post on X or Facebook appears with a note beneath it written by other platform users, that note is context: information, sources, and competing perspectives added alongside the content so readers can judge it more fully. Right now, that context layer is proprietary, created and owned by the platforms on which it appears. But context doesn't have to be built this way. Treating it as a protocol, a shared open standard any platform can adopt rather than a feature owned by one company, is an opportunity to build prosocial features into the infrastructure of the web.

In their paper "From local hacks to global standards: The hidden politics of internet protocols," Matthew Zook and Ate Poorthuis use three examples to illustrate that historically, infrastructure has started with a smaller use case and then scaled. The danger is that early informal decisions become global rules without enough consideration for human rights and other impacts.

One example they give is the country code top-level domain system, the familiar national suffixes like .FR for France or .UK for Britain, which were built according to ISO 3166, an existing list of two-letter country codes maintained by the International Organization for Standardization. But, from the 1980s until today, the ISO list itself has not been a neutral inventory of the world's nations. It is a list that elides the fraught question of what counts as a country. As a result, particular political histories and institutional relationships were adopted into the domain name system along with those embedded judgments. Territories with contested sovereignty, colonial dependencies, or without recognized statehood were included or excluded, baking political decisions about place into the architecture of the internet.

Context is a new, developing layer of the internet. The most promising tools for adding context to online content are community notes, used by both X and Meta, and which show real promise in reducing online harms like misinformation and disinformation. But these context layers are proprietary, owned and managed by these two platforms and, like the ISO list, they come with biases—in this case, those of these platforms’ unique user bases and their commercial interests. If we want context to scale and be scrutable, we need to facilitate context with protocols: that is the ‘how’ of building a context layer. An open, opt-in standard that any publisher, browser, or platform can adopt, rather than a feature each company builds and owns, can democratize context and appropriately place it within the realm of the political: that is the what.

Building such a protocol deliberately, in the open, lets us embed choice, user agency, and prosocial values into the context layer of the internet, a Broader View button (demo here) built into the web itself, rather than allowing closure to settle around whatever already exists before anyone has deliberately chosen, as happened with the domain name system.

Removals, labels and annotation

Most efforts to improve what people encounter online currently fall into three general categories: removal (take it down), labeling (flag it), and crowdsourced annotation (let users add context). The first two require a platform or trusted third party fact-checkers to decide what is true, which much of the public no longer trusts it to do.

Crowdsourced annotation like community notes was developed in part to address the issues of the first two categories, and the evidence suggests that it succeeds, at least in limiting the spread of misinformation and disinformation. The system's own designers found that algorithm-selected notes made users about 26 percent less likely to agree with a misleading claim, and that exposure to notes reduced likes and retweets by 25 to 34 percent in live deployment. A causal study, covering roughly 285,000 Community Notes on X (formerly Twitter), found that attaching a note cut subsequent retweets by about half and raised the chance the author deleted the post by around 80 percent. Issues appear, however, when trying to scale these efforts. The average note takes more than fifteen hours to appear, by which point roughly 80 percent of a post's reach has already happened, so the net effect on overall virality falls to between 16 and 21 percent.

Plus, many posts that perhaps should have notes never do. A note requires volunteers to notice the post, write a note, and reach cross-partisan agreement before it is published. This is a high barrier that only about 11 percent of proposed notes ever meet. Fewer than 10 percent of published notes reach "helpful" status, and 26 percent of those that do are later removed due to disagreement.

In addition, a great deal of online content is not false. It is accurate as far as it goes, but may show only one side of a contested issue. No current moderation system systematically surfaces the competing frames around a post that is true-but-partial, and a small, hyperactive minority of users produce most of what everyone sees, and that content skews more politically extreme than what the typical user posts, so the less partisan majority is rendered nearly invisible. On genuinely contested questions, the “true or false” binary is even less useful: the truth often isn’t settled for years, long after the moderation decision has been made and the post has done its work.

Rather than seeing this as an indictment of content moderation or Community Notes, which is the clearest proof we have that a context layer can work, we see it as evidence that the bottleneck is architectural: a layer run by volunteers inside one platform's user base will always face an upper ceiling that we believe only a shared standard can alleviate.

Why now? AI, obviously

A context layer that depends on volunteers noticing a post, writing a note, and reaching cross-partisan agreement will always be slower and more limited than the content it is intended to contextualize. What has changed is that large language models can now do part of this work by reducing the demand on the volunteer labor that made it scarce, and drawing from a wider range of relevant material.

A recent system, Supernotes, uses a language model to synthesize these fragments into a single candidate note, then scores that candidate by modeling how a politically diverse set of raters would respond to it. In testing, participants preferred the AI-synthesized notes to the best existing human-written ones roughly three times out of four. In this study, the model didn't decide what was true; it drafted and assembled content from existing human notes. Whether the result was helpful still came down to human cross-partisan agreement, and the AI was what let that agreement extend to far more content than volunteers could reach alone. This points to something a shared context layer could do beyond simply showing different perspectives side by side: surface where communities that usually disagree actually share ground, an approach sometimes called bridging. The algorithm behind Community Notes was also built on this principle, scoring a note highly only when people with otherwise opposed rating histories agree it is helpful, rather than relying on a simple majority. Supernotes extends that same bridging logic with AI.

There is a reason AI may be well-suited to this particular job. People often distrust context when it comes from a perceived opponent, and some research suggests they treat AI-generated summaries as comparatively impartial. We should be cautious, because AI carries its own biases that have to be managed openly. But for the narrow task of laying out how different communities frame an issue, that cites sources, AI may have an easier time being heard.

How? The protocol opportunity

A protocol-level approach asks: what if context were shared infrastructure, like a web standard, rather than a feature limited to one provider? A standard that lets any platform, publisher, or browser participate without each needing to build a feature from scratch, and lets context travel across services? And what if users had agency to choose their context provider?

Our model is the closed-captioning (CC) mark. It is instantly recognizable, works across virtually all video regardless of who made it, is owned by no single company, and turns on only when the viewer wants it. Part of the mark’s power is the mark itself: a single recognizable symbol compresses the whole idea into two letters anyone can spot on any screen. A universal mark is what makes an opt-in layer usable by ordinary people, not just legible to technologists.

We propose the same for context: a universal, opt-in icon which we call the “Broader View button,” that a reader clicks only if they want the fuller picture. On a contested political post, that might mean seeing how different communities understand the same event, what they agree on, where they disagree, and the perspectives and sources each community draws on, so a reader can understand the landscape and draw their own conclusions. On a video of a duck leading her ducklings across a highway, the button might open up a wider understanding about migration, habitat loss and how some cities are redesigning roads around wildlife. Context isn't only a corrective for our worst content; it's an invitation to be more curious about all of it.

No content is removed, no fact-checks are pushed into the feed. Because it adds speech rather than restricting it, the approach can hold support across a political spectrum that agrees on little else about online speech.

We propose that the standard should be provider-agnostic. Like choosing a default search engine, different providers could supply the context behind the same button, separating the standard (how context is displayed) from the curation (who, or which AI model, assembles it). Letting readers pick their own provider is a form of user agency, and experienced users judge content more favorably when they have actively chosen it rather than had it chosen for them.

We also propose that trust and safety belongs in the protocol itself, for instance, requiring that quoted text in a context window trace to a verifiable source and that off-topic pile-ons be filtered, so every implementation meets a minimum threshold.

A layer worth building

A key principle behind years of content moderation has been to remove false content and correct the record. But much of what hardens divides online is not false. Instead it is partial or one-sided, and no content moderation verdict can address this problem. What's missing is not a better judge or a jury. It's a layer that lets users who want it see a bigger, fuller picture.

Across established democracies, the spread of social media has tracked with falling trust and rising polarization, yet the research has gone overwhelmingly toward documenting that harm rather than testing ways out of it. One 2021 review of more than ninety studies notes how little work has explored how media might actually depolarize. In other words, we have mapped the problem in great detail, but we have barely begun to identify or fund the solutions.

Community Notes is the clearest proof we have that a context layer can work, and of its limits. They are not a reason to abandon the idea, but a reason to build it properly as shared infrastructure, rather than a feature owned by one company. A Broader View button will not be perfect, but a perfect solution does not exist, and there are costs for waiting.

Where should this work live?

Despite years of thinking from scholars like Francis Fukuyama and Renée DiResta, what are called “middleware” solutions to content moderation haven’t made it into the protocols standardization pipeline. We are still left with platforms, not protocols, implementing solutions, which Mike Masnick pointed out are not ideal.

Taking on a context layer has implications for any technical standards body's mandate already dealing with content, and those bodies are few. The W3C is the most natural home, since it already looks after the web and social standards, however its prior work on annotation would only be a partial help. ISO could also take it on as global trust frameworks like C2PA are increasingly within mandate and expertise. Whoever shepherds the work takes on more than writing the standard itself. They foster a community of trust and safety rules stewards and will likely bring together a huge cross section of web services and platform implementers.