Intro
Complaints about disinformation on social media have historically focused on the harms of false factual claims, such as false quotes or misleading statistics. This misses something at least as harmful: false information about what groups of people believe. Call these social mirages or fake views; they are implicit in the likes, comments, and shares of social platforms, forums, and comment pages.
For example, it’s 2025, and a progressive friend shared a Facebook post advocating for the abolishment of ICE. It has hundreds of likes, with exclusively approving comments. Two hundred people in an echo chamber have implicitly suggested that most people want to abolish ICE. Later, you open X and stumble into a different echo chamber where a video of someone getting pepper-sprayed in the face by ICE has dozens of people responding with a gleeful “this is what I voted for.” No moderate voices take the time to express a nuanced or mild approval, because they’re reading The Economist or otherwise not interested in an intra-coalition debate on X. So you only see glee or silence. The lack of mixed commentary gives both you (and them) false implicit social information.
These funhouse mirrors stoke outrage, drive polarization, and erode our mutual understanding.
Social mirages
Astroturfing - Consider the infamous Russian-sponsored @TEN_GOP account (screen caps and history here) posting anti-Clinton and pro-Trump content in 2016. While the posts themselves were often “fake news,” they also had sock puppet accounts voting them up and commenting, creating a mirage of support. Similarly, see China’s 50c party for an example of a state using similar tactics domestically. It is estimated that they make half a billion false comments a year.
Unchallenged fringes - In the wake of the United Health CEO’s murder, there were posts from progressives that were celebrating his assassin as justified, if not heroic. Many such threads had no moderate voices, either because they are conflicted or otherwise not passionate enough to say “maybe murdering bad guys is still bad” in that moment. If you’re a viewer from outside of the progressive coalition who can’t identify the subtle signs of the subgroup making these claims, who stumbles on such unchallenged comments, is likely to come away with the impression that leftists support violence generally.
Shallow winners - When Elon tweets 💯 in reply to a post saying “The reason that the mainstream media ignores the threat of the far-left is because the media are part of the far-left.” and you see that the post has 140k likes (and the post he quoted has 49k likes) you might walk away from that thinking quite a lot of people agree that the mainstream media is far left!
One solve
Community Notes has been effective at fighting factual disinformation; can a similar approach work with social disinformation?
Let’s say you come across a post you judge quite immoral, and are dismayed to find its top comments all agree and cheer it on. Now imagine this post also included a rundown of how your entire nation feels about it, with most of it sharing your disapproval. You could expand it to see the breakdown by key demographics, or you could open representative comments from each opinion cluster. When doing so, you find that at least some who approve it are operating in a different frame than you, and from that frame, you see some common ground.
Building this is not hard. We can combine the approaches of two existing organizations: YouGov and Talk to the City. What Talk to the City does is ask a population of people to weigh in with free-form opinions on a given subject, break down each thing they say, cluster similar thoughts, and surface who endorses each opinion cluster. They also surface key quotes from each.
Of course, we can’t ask the whole nation to offer their hot take on every social media post, so we need to add on this with YouGov’s approach to opinion polling, where they maintain a panel of a few thousand members who have opted into regularly answering opinion surveys for a small fee. Furthermore, YouGov asks panel members for their demographic information and verifies their identity. YouGov then utilizes a statistical technique called Multilevel Regression and Post-stratification (MRP), which allows them to weight and aggregate each panelist’s answer such that it has the same effect as running a demographically representative poll.
For example, YouGov asks, “Would you support or oppose it if National Guard troops from another state were currently being deployed to your state?” link
Bringing these two ideas together, then:
- Find a post you want a opinion survey for
- Show it to a a few thousand people for whom you trust you know the demographics of
- Ask for a few sentences of how they feel about it.
- (optional) have llm break disparate thoughts, otherwise treat as single thought
- Cluster thoughts based on vector embeddings. Try to ensure there’s at least a “approve” and “disapprove” but allow for others to emerge.
- Find the effect constant for each demographic variable and cluster
- Look at census data and find the cell size for each combination of demographic variables we’re tracking
- Big triple loop through each thought cluster, each cell, and each demographic attribute with simple math and you have an idea of how much of the nation would be in each cluster.
- (optional) Calculate a credibility interval for each demographic variable using a simulation.
- (optional) Repeat for more geographic areas, each US state, etc.
- For each cluster, find “representative thoughts” by ranking thoughts by how central they are to cluster, how readable they are.
- Create a report showing the clusters, membership size, and representative thoughts of each.
How it helps
For the astroturfing case, foreign actors trying to create false dissent might not be exposed as outright fakes, but at least their stated opinions would be shown to be fringe. You would need to take care not to allow the sock puppets into your survey panel. It’s out of the scope of this proposal, but it’s interesting to imagine how this could be deployed inside a state with domestic information control to combat its domestic propaganda.
For the unchallenged fringes case, consider our example of people celebrating the assassin’s actions. An opinion survey would have shown participants in this thread are a minority (I know because YouGov asked their panel about this specific event and calculated that merely 12% of Americans think violence is ever justified against corporations).
For shallow winners, think of our example Elon tweet about the mainstream media being far-left. An opinion survey would quickly show this to be at least out of line with the general population’s view. In this case, we have a traditional poll showing as much; however, there are many cases where we do not.
Building it
A small org could build this. Assemble a panel of people, present them with a “feed” of content for which their opinion is wanted. Run a bot account on Twitter/Threads/Bluesky that shares completed panels back to the original thread that generated them. Allow the bot to be tagged on posts, which acts as a vote that a post be considered for a panel. If a similar post has already been done, the bot can notice that and instantly bring in the relevant report. It is expensive to run, and grant writers may not be eager to fund it, so I suggest allowing sponsorship of reports. Someone might sponsor private reports for internal research or public reports when they expect their favored stance to be popular and thus reinforced as common knowledge.
An existing social platform could also create it as a feature, much like Community Notes was. Twitter already has a concept of verified accounts. They could go further by having people interested in being panelists verify a bit more (and provide demographic information) and then give them a tab in the app full of posts they can weigh in on. Such verified panelists can then comment on posts in this feed, creating a public comment that also serves as their submission to the report. The public nature of the comment might make it more trustworthy to outsiders, although preference falsification could undermine it. I’m not entirely sure why Twitter would make this a priority, but perhaps if they catch enough flak for their bot problems, they can “do something about it” with this feature. Or, their pride in being a town square may be motivation enough.