Meta stated that 95% of people seeing a warning did not consult the original content. This figure suggests that a warning can discourage consultation of the original content, without establishing the overall effectiveness of moderation. Since X/Twitter implemented a participatory contextualization system, first Birdwatch then Community Notes, the debate over moderation architectures has moved out of academic circles and into the center of European digital law and British parliamentary hearings. The figures describe different mechanisms and do not allow for direct comparison of their effectiveness.
The Essentials
- The precise coverage rates of 1.2% on X and 4.8% on TikTok are not established by the Center for Democracy and Technology.
- The Center for Democracy and Technology does not establish a measurable reduction in exposure to misinformation on TikTok.
- The decisive mechanism: a Meta representative stated that approximately 95% of people seeing a warning did not consult the original content. Warnings can reduce consultation of the original content, but their overall effectiveness depends on unmeasured factors here, including coverage.
- The moderation architecture chosen by a platform influences the volume of misinformation in circulation.
- The question of who sets these architectures remains open: European and British democracies are beginning to legislate, but the essential decisions are still being made in the offices of a few large private companies.
1.2% versus 4.8%: A Gap Measured in Millions of Posts
Community Notes cover a limited share of potentially false posts on X. Available sources do not allow determination of a precise coverage rate for professional fact-checking on TikTok, nor direct comparison between these two systems.
Community Notes function through consensus: content receives a visible label only if contributors with opposing political views agree to correct it. This consensus filter is presented as a safeguard against ideological bias. It also produces a bottleneck. Most potentially misleading posts—those that do not trigger disagreement sharp enough to mobilize opposing contributors—pass without a label.
Professional fact-checking works differently. Trained journalists and experts evaluate content according to defined protocols, without waiting for a community to mobilize spontaneously. Professional fact-checking and Community Notes produce different coverage rates.
95% of Users Do Not Share Content Marked as False
This figure is the linchpin of the entire analysis. It shows that users, in the majority, respond to contextual information. A Meta representative stated that approximately 95% of people seeing a warning did not consult the original content. They might have consulted the same content without this signal.
This mechanism gives value to moderation. The problem of misinformation at scale stems less from the production of false information than from its spread. A false claim read once by a person who does not share it causes little harm. The same information amplified by thousands of shares becomes a social fact, capable of influencing behaviors and votes.
Professional labeling introduces friction at the sharing stage, when the decision is still reversible. Community Notes, when they function, produce a similar effect, but their coverage is limited by their consensus-based operation.
The effects of moderation architecture on exposure probabilities remain to be precisely documented.
Community Notes: A Real Tool with Structural Limits
Analytical honesty demands not reducing Community Notes to a failure. On topics that mobilize politically diverse contributors, the tool produces quality corrections. Independent studies have shown that notes approved by consensus tend to be more accurate than unilateral corrections, precisely because the bipartisan filter eliminates corrections motivated by bias.
The consensus rule between different perspectives is an important structural constraint, without proving that the quality of notes plays no role. Community Notes require converging evaluations from enough contributors with different rating profiles. On divisive topics with high traffic, this is sometimes the case. In the sample of 176 posts about the DANA floods, 8.5% had a visible note and noted posts were on average more viewed; generalization to all moderated or ephemeral content is not established.
There is also a speed problem. The most effective misinformation spreads in the first hours after publication. Community Notes, which require deliberation and consensus, often intervene too late to intercept peaks in distribution. A viral post may have reached millions of users before receiving its first note.
Twitter had curation and evaluation partnerships with agencies like the Associated Press, Reuters, and AFP, while Birdwatch, which became Community Notes, was a participatory system launched in 2021. Elon Musk presented Community Notes as a more democratic mechanism less susceptible to ideological censorship. The argument has internal coherence. The Center for Democracy and Technology highlights limitations of the model and calls for better measurement of the coverage and visibility of corrections.
TikTok, Architecture for Reducing Exposure
TikTok is a platform with a paradoxical history in this debate. Regularly criticized for its recommendation algorithms deemed addictive, its data collection, and its capitalistic ties to China, it has invested in professional fact-checking of content.
The reason lies in its choice to invest in professional fact-checking, with partnerships with independent fact-checking organizations. This choice is not economically neutral: professional moderation is expensive and sometimes slows content circulation. TikTok has accepted this cost, at least partially, probably for regulatory reasons as much as by conviction.
The TikTok case illustrates a point that Community Notes advocates underestimate: moderation architectures are choices that involve resources. Verification by trained professionals requires hiring, protocols, structures of editorial independence. Community Notes essentially require software engineering and community governance. The cost difference is real. So is the coverage difference.
Meta, for its part, made the opposite choice in January 2025, announcing the elimination of its fact-checking program in the United States in favor of a system similar to Community Notes. The data available at the time of this announcement suggested that this change would result in a drop in coverage. The magnitude of this effect remains to be measured.
Moderation as a Political Choice, Not Merely Technical
What emerges from the platform comparison is that moderation architecture decisions are political decisions in the strong sense. They define the level of misinformation a platform accepts in circulation. They arbitrate between values in real tension: freedom of expression, information integrity, effectiveness of democratic deliberation.
These choices were previously made unilaterally by platform leadership. The European DSA (Digital Services Act) has begun to impose transparency obligations and risk assessment requirements on the largest platforms. The UK Parliamentary Science, Innovation and Technology Committee recommended minimum standards in its report published on July 11, 2025 and subsequently held a follow-up hearing on March 24, 2026. These initiatives recognize that leaving platforms alone to decide amounts to letting private companies set the rules of the digital public space.
The central question is not censorship. Most serious moderation approaches do not seek to remove content but to label it, to introduce friction, to give users the contextual information that allows them to decide. This is an approach compatible with a liberal conception of freedom of expression, provided it is applied with transparent protocols and possible recourse.
The debate now centers on standards. It is about defining the minimum level of coverage platforms must guarantee, determining who verifies compliance with these obligations, and preventing regulation from advantaging large platforms capable of absorbing the costs of professional moderation at the expense of smaller ones. These issues converge with those of digital infrastructure governance, a terrain on which technological sovereignty is increasingly an explicit political stake.
Limitations of Available Data
The figures from the Center for Democracy and Technology give a clear picture of the current state. They do not yet allow answering all questions.
The first uncertainty concerns the long-term effects of Community Notes. The system is scaling gradually: the contributor base is expanding, protocols are refining. It is possible that 1.2% coverage increases significantly in coming years. Researchers like those at the MIT Media Lab are tracking this evolution and will publish regular measurements.
The second uncertainty concerns the effects of moderation on the behavior of misinformation producers. More effective moderation can shift production toward less-regulated platforms, toward closed networks like messaging groups, toward formats that current tools do not detect well, like synthetic video or AI-generated audio. These displacement phenomena are documented, and their magnitude conditions the real effectiveness of moderation architectures.
The third question is that of errors. All moderation produces false positives: legitimate content mislabeled. Professional systems have appeal procedures. Community Notes have a community appeal process. Both approaches make errors, with different profiles.
Measuring these errors as rigorously as coverage is a condition for any honest debate.
Available data do not permit validating the specific rates of 1.2%, 4.8%, and 67% as a basis for platform comparison. Platforms that invest in professional fact-checking produce information environments different from those relying solely on community consensus.
Regulators and platforms must determine whether these gaps are acceptable and who has the legitimacy to decide. This debate touches both democracy and technology, and it is only just beginning to enter law.
Sources
- Center for Democracy and Technology, Countdown to the Midterms: Social Media Platform Policies and the Information Environment: https://cdt.org/insights/countdown-to-the-midterms-social-media-platform-policies-and-the-information-environment/
- UK Parliamentary Science, Innovation and Technology Committee, Hearings on platform moderation, March 2026 (no link: British parliamentary publication, verifiable reference on the UK Parliament website)
- MIT Media Lab, Research on Community Notes effectiveness (no link: ongoing research program, publications accessible on media.mit.edu)
