What counts as misinformation
Before anything can be measured, a researcher has to decide what is being measured. Terms are used in different ways, and the choice matters. A common distinction treats misinformation as false or misleading content shared without intent to deceive, and disinformation as false content spread deliberately. In practice, intent is almost impossible to observe from the outside, so many studies avoid the distinction and simply count false or misleading items.
Even that is less simple than it sounds. Some claims are clearly checkable and clearly wrong. Many more are partly true, missing context, outdated, satirical or matters of opinion. A study that treats only items rated false by fact-checkers as misinformation will produce a different picture from one that includes misleading framing or low-quality sources. Neither approach is wrong, but their results are not directly comparable.
How false claims get identified
Researchers typically use some mix of the following approaches.
- Fact-checker ratings. Items already assessed by professional fact-checking organisations are labelled and then followed. This is reliable for individual claims but covers only what fact-checkers chose to examine, which tends to be popular, political or newsworthy material.
- Source-based lists. Instead of judging each item, researchers classify publishers as more or less reliable and treat content from low-rated sources as a proxy. This scales well, but a generally unreliable site can publish accurate stories, and reliable outlets sometimes make mistakes.
- Automated classifiers. Machine-learning tools are trained on labelled examples to flag likely false content. They can process large volumes, but their accuracy depends on the training data, they can pick up on style rather than truth, and they may perform worse on new topics or languages.
- Human coding. Trained reviewers read a sample and assign labels. This can be careful, but it is slow, and reviewers can disagree, which is why studies often report how consistent their coders were.
Each method embeds judgments, so findings are best read as results under that particular definition.
Measuring spread: reach, exposure and engagement
“Spread” can mean several different things.
Shares and reposts count how often an item was passed along. They are visible on many services, and easy to tally, but they show only active sharing, not who saw the item.
Reach and exposure ask how many people actually encountered the content. Platforms often hold this information, but outside researchers frequently cannot access it, or only through limited programmes. Without it, studies may depend on engagement counts as a stand-in, even though engagement metrics capture only a portion of those who saw something.
Belief and attitude change are a further step. Seeing a false claim does not mean accepting it, and accepting it does not necessarily alter behaviour. Surveys and experiments can examine belief, but they depend on honest, accurate self-report and often measure it shortly after exposure rather than over time.
A study can therefore find that a false story was widely shared while showing little about how many people were persuaded. Conversely, content that few people share may still be seen by many.
Whose data, and what it misses
Most large-scale tracking relies on platform data, which has well-known limits. A study of one service says little about private messaging apps, where content is harder to observe, or about offline conversation. Access rules differ between services and change over time, and researchers can only study what the company chooses to make available.
Another issue is that samples drawn from public posts do not represent all users. People who post publicly often differ from those who read without posting. Studies that follow a set of accounts may also be skewed by automated activity, and the presence of fake accounts can inflate the apparent spread of a claim or make a fringe view look more common than it is.
The distribution of spread is usually very uneven. Findings often indicate that a small share of accounts or sources account for a large share of the false content shared, while most users encounter it rarely. That kind of concentration affects how any average figure should be understood.
Why recommendation systems complicate the picture
Content does not reach people only because friends share it. Recommendation algorithms decide what appears in feeds, and researchers disagree about how much they amplify false material compared with what people actively seek out. Separating the effect of the system from the preferences of users is hard, because the two interact: people click on things, the system learns, and the pattern reinforces itself. Studies that can test changes directly, for example when a platform adjusts a rule, offer clearer evidence than observational snapshots, but such opportunities are uncommon.
Counting corrections and claims that go viral
Some research looks at whether corrections work. Results typically depend on the topic, how the correction is delivered and how soon after the original claim. It is also difficult to measure whether a correction reaches the same people as the claim did.
Studies of viral memes show that a rumour can spread through a mix of attention, humour and identity, rather than because people believe it literally. People sometimes share content to signal group membership, to joke, or to warn others, which complicates any attempt to equate sharing with endorsement.
What to look for when reading a misinformation statistic
A figure about “how much” false content circulates is only as useful as its definitions. Questions worth asking include:
- How was misinformation defined, and who decided what was false?
- Which platforms, languages and time periods were covered?
- Does “spread” mean posts, shares, views or belief?
- Is the finding about the average user or about a small, highly active group?
- Was exposure measured directly, or estimated from engagement?
Misinformation research is a developing field, and its methods are being refined. Its most dependable lesson is a modest one: claims about the scale and effects of false content depend heavily on definitions and data access, so a single statistic rarely settles the matter.