Bot detection

No single test identifies a bot. Detection combines several signals: how often an account posts and at what hours, how quickly it replies, how new the account is, whether its profile is complete, how its followers and follows are distributed, how repetitive its text is, and whether it moves in step with other accounts. Machine-learning classifiers weigh these together and produce a likelihood, not a verdict.

Accuracy is limited in both directions. Sophisticated automation with human oversight can look human, and some people post so regularly, or so repetitively, that they resemble bots. Good practice treats scores as evidence to review and reports results with the uncertainty attached.

Automated amplification inflates the apparent size of a conversation. Before a team reports volume or reach to leadership, it helps to know how much of it came from accounts that are not people. Detection also helps separate organic criticism from coordinated activity such as brigading or astroturfing.

Not all bots are malicious. News feeds, alert accounts and customer-service assistants are automation too. And "bot" is often used loosely to dismiss real people who disagree, which is a reason to rely on evidence rather than impressions.

Bots are also only one part of manipulation. Many coordinated campaigns use real people or hybrid accounts, which is why analysts also look at coordinated inauthentic behavior.