About 20% of Active X Accounts Appear Bot-like

ByJames McCallef
bot percentage: A network of connected smartphone screens and profile icons suggesting social media bots

About one in five active, visible X accounts appears bot-like, fake, or spammy. A 2025 analysis of roughly 200 million users found bots averaged about 20% across major X event datasets. That closely matches a 2022 estimate that 19.42% of recently active public accounts were fake or spam.

So the best current public answer is about 20%—if you mean active public accounts or visible event conversation, not a platform-wide census. X’s own old disclosure measured something narrower: fewer than 5% of monetizable daily active users, or mDAU, in Q1 2022.

The best-supported public range is already laid out in our earlier bots on X estimate and X bot percentage analysis: the cluster near 20% comes from broader, more transparent public methods, while the much larger figure comes from a specific proprietary screen on a specific sample.

The best-supported public estimates range from about 20% to 64%

The strongest public platform-adjacent estimate for active public accounts is 19.42%, from a SparkToro and Followerwonk analysis of accounts that had tweeted in the prior 90 days. That study was explicit about scope: it looked at public, active accounts, not every account on the service, and counted fake or spam accounts rather than only automated bots.

A separate large-scale academic result lands in nearly the same place. A 2025 Scientific Reports paper analyzed roughly 5 billion tweets from about 200 million users across major events and used the BotHunter detection system. It reported a 20% bot baseline in online discussion, with some events spiking above that level. That is not a full census of all X accounts; it is a bot share within event datasets labeled by the algorithm.

Taken together, those two estimates are why “about one in five” is the best short answer for a cold reader. They are independent, recent, and surprisingly close despite using different methods.

The outlier is 64% from 5th Column AI, which said it analyzed 1.269 million X accounts and found that approximately that share was “potentially bots.” That is a real published estimate, so it belongs in the range. But it is a proprietary audit, not peer-reviewed, and not presented as a platform-wide measurement of all accounts. In practice, that makes it more like an aggressive fraud screen than a census.

There is also a much smaller number that answers a narrower question. A 2024 arXiv paper estimating fake profiles with GAN-generated faces found only 0.021% to 0.044% of active Twitter/X accounts fit that subtype, which the authors described as a lower bound. In daily terms, the paper estimated around 10,000 active accounts per day using AI-generated profile images. That does not mean fake accounts are rare overall; it means this one specific kind of fake account is only a slice of the total.

Here is the range in one view:

Estimate What it measured
<5% X/Twitter’s own estimate of false or spam accounts in mDAU, with known spam/fake accounts excluded from mDAU
19.42% Active public accounts that were fake or spam
20% Bot share in large event-discussion datasets on X/Twitter labeled with BotHunter
0.021%-0.044% Lower bound for active accounts using GAN-generated face profile photos
64% “Potentially bots” in a proprietary audit of 1.269 million sampled accounts

What X and Twitter have officially disclosed about spam and fake accounts

X’s own most cited number is still fewer than 5%, but that figure is narrower than most people think. In merger-litigation material preserved at the SEC, Twitter said that in Q1 2022 false or spam accounts represented fewer than 5% of monetizable daily active users.

That matters because mDAU was not “all users.” Twitter’s filing said it counted users who logged in and could be shown ads, while known spam and fake accounts were excluded from that figure. So the company’s famous number was never a claim that fewer than 5% of all accounts on the platform were bots.

The SEC’s June 15, 2022 comment letter underlined the point by asking Twitter to explain the methodology behind its disclosure that false or spam accounts during fiscal 2021 were fewer than 5% of mDAU. That is regulator language for a very ordinary problem: the denominator changes the story.

A rough way to think about it is this: asking “what share of mDAU is spam?” is like asking what fraction of passengers on a plane are ticketed travelers after airport security has already turned away some obvious gate crashers. Asking “what share of all accounts on the platform are fake, spam, or automated?” is a broader question.

Twitter’s disclosure said false or spam accounts were estimated at fewer than 5% of mDAU, and that known spam and fake accounts were excluded from mDAU.

X’s own most cited number is still fewer than 5%, but that figure is narrower than most people think.



Why estimates diverge across accounts, activity, and detection methods

The biggest reason estimates vary is that “bots,” “fake accounts,” “spam accounts,” and “bot activity” are not the same thing. A bot account is an account that posts or acts automatically. A fake account is an account pretending to be a real person or entity. A spam account is focused on promotion, scams, or low-value repetitive posting. And bot activity can come from mixed systems, including human-run accounts using automation tools.

That category problem alone can push estimates apart. The SparkToro/Followerwonk estimate of 19.42% grouped fake or spam active public accounts together. The Scientific Reports paper’s 20% figure focused on bot-like users in event datasets, labeled with BotHunter. The GAN-face paper measured only one visual tell of one subtype of fake profile. The 5th Column AI figure counted accounts judged “potentially bots” under its own proprietary system.

The second reason is sampling. A study of accounts active in the last 90 days is measuring a different population from a study of event chatter, and both differ from a study of all registered accounts, dormant accounts included. Dormant fake accounts can inflate one measure; highly active bot networks can dominate another.

The third reason is detection method. Researchers use combinations of:

  • posting frequency and timing
  • follower/following patterns
  • text repetition or coordination
  • profile metadata
  • image analysis for fake faces

The 2025 Scientific Reports paper relied on BotHunter, an algorithmic classifier applied to large event datasets. The 2024 GAN-face study used image-based detection to flag profile photos likely created by generative adversarial networks. The SparkToro/Followerwonk analysis used public-account signals. Each method catches some bad accounts and misses others.

That means the number can swing sharply if you change the threshold. Tighten the screen, and you count fewer suspicious accounts with higher confidence. Loosen it, and you catch more borderline cases along with more false positives. Public researchers also lack X’s private metadata, so outside estimates necessarily depend on samples, heuristics, or model settings that can shift the result.

The best public synthesis, then, is not that one estimate “wins” and all others are worthless. It is that the center of gravity sits around 19% to 20% for active, visible X activity, while lower official figures use a narrower denominator and higher figures come from narrower or more aggressive screens.

One useful derived number makes that concrete. If a dataset contains 200 million users and the bot baseline is 20%, that implies roughly 40 million bot-labeled users within those event datasets, a large enough block to shape what people see in fast-moving public conversations.

The next hard limit is visibility. X has internal signals that outsiders do not, including device patterns, IP clustering, enforcement history, and account-recovery data. Until the company publishes a modern, auditable methodology, about one in five active public accounts or conversations showing fake/spam/bot-like traits is the best-supported public answer.

Key Takeaways

Frequently Asked Questions

How many bots are on X in 2025 or 2026?

About one in five active public accounts or visible discussions on X is the best-supported public estimate. The closest recent public numbers are 19.42% for active public accounts judged fake or spam and a 20% bot baseline in a 2025 academic analysis of major X event datasets. Those are not full platform censuses, but they are the strongest transparent estimates in public view.

What percentage of X accounts are fake?

For active public accounts, the best public estimate is 19.42% fake or spam. X’s old official disclosure of fewer than 5% referred to monetizable daily active users, a narrower set that excluded known spam and fake accounts from the denominator.

Why is X’s under-5% number so different from the 20% estimate?

Because the numbers are measuring different populations. The company figure applies to mDAU, while outside researchers usually study active public accounts, public conversations, or sampled profiles. Different denominator, different filter, different result.

Are all fake accounts bots?

No. Some fake accounts are manually run sockpuppets, impersonators, or scam profiles. Some bots are openly automated service accounts and are not necessarily pretending to be human. On X, those categories overlap, but they are not interchangeable.

How do researchers detect bots on X?

Public researchers typically combine posting patterns, network structure, metadata, text repetition, coordination, and profile-image analysis. The Scientific Reports paper used BotHunter on event datasets, while the 2024 arXiv paper looked specifically for GAN-generated faces in profile photos. Each approach catches different failure modes.

References

Further Reading

Last reviewed: 2026-06