Botometer 101: Social bot practicum for computational social scientists
Abstract: Social bots have become an important component of online social media. Deceptive bots, in particular, can manipulate online discussions of important issues ranging from elections to public health, threatening the constructive exchange of information. Their ubiquity makes them an interesting research subject and requires researchers to properly handle them when conducting studies using social media data. Therefore, it is important for researchers to gain access to bot detection tools that are reliable and easy to use. This paper aims to provide an introductory tutorial of Botometer, a public tool for bot detection on Twitter, for readers who are new to this topic and may not be familiar with programming and machine learning. We introduce how Botometer works, the different ways users can access it, and present a case study as a demonstration. Readers can use the case study code as a template for their own research. We also discuss recommended practice for using Botometer.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Explain it Like I'm 14
Botometer 101 — A simple explanation
What is this paper about?
This paper is a friendly, step-by-step guide to Botometer, a tool that helps people figure out whether a Twitter account acts more like a human or like a bot (a software-controlled account). It explains how Botometer works, how to use it (through a website or code), shows a small example study, and gives tips on using the tool correctly and fairly.
What questions are the authors trying to answer?
The paper focuses on practical questions:
- What are social bots, and why do they matter for research?
- How does Botometer detect bot-like behavior?
- How can someone (even without much coding or machine learning experience) use Botometer?
- What do Botometer’s scores mean, and how should they be interpreted?
- What are smart, responsible ways to use bot detection in research?
How Botometer works (in simple terms)
To make the ideas easy to grasp, here are some plain-language explanations and analogies.
What is a social bot?
A social bot is a social media account partly or fully controlled by software. Some bots are harmless or helpful (like news updates). Others are deceptive: they may pretend to be people, spread misinformation, or push certain topics to make them look more popular.
The main idea: supervised machine learning
Think of teaching a friend to tell apples from oranges by showing lots of labeled examples (“this is an apple,” “this is an orange”). Over time, your friend learns patterns to tell them apart. Supervised machine learning is the same idea, but with a computer: it learns from many examples labeled “bot” or “human” to recognize patterns.
- Botometer looks at more than 1,000 “features” (characteristics) of an account. Examples:
- Profile clues: default picture or not, account age, screen name length.
- Behavior clues: when and how often it tweets, who it talks to.
- Content clues: what kinds of words it uses.
- Network clues: how it connects with others.
- It turns these clues into numbers (like a checklist), then runs them through a model.
The model: many small “judges” voting
Botometer’s main model is called a Random Forest. Imagine a panel of many small “judges” (decision trees). Each judge looks at some features and votes “more human-like” or “more bot-like.” The final bot score reflects the balance of votes.
Newer Botometer versions use an Ensemble of Specialized Classifiers (ESC). That’s like having different expert panels for different bot types and one for humans, then combining their opinions. This can create a “bimodal” pattern in scores (lots of very low and very high scores).
Scores and what they mean
- Raw bot score: a number from 0 to 1 (shown as 0 to 5 on the website). Closer to 1 (or 5) means “more bot-like.” This is not the probability of being a bot; it’s a ranking of how bot-like the behavior looks compared to training examples.
- CAP (Complete Automation Probability): a number that estimates the chance an account is automated, taking into account how common bots are overall. This helps choose a reasonable cutoff while balancing mistakes.
- Language matters: Many content features are based on English. For non-English accounts, Botometer also gives a language-independent score (ignoring language-based features). Use the score that fits the account’s language.
Speed vs. detail: two Botometer options
- Botometer V4 (detailed): Fetches an account’s latest 200 tweets and mentions to analyze. It’s thorough but slower (limited by Twitter’s data access rules).
- BotometerLite (fast): Uses only the user’s profile metadata (which is embedded in tweets and easier to collect). It’s much faster and good for huge datasets, but less detailed.
How people can use Botometer
- Website: Good for quickly checking a few accounts.
- API (with the official Python package): Good for checking many accounts in bulk; requires a Twitter developer account and an API subscription. Keep in mind “rate limits,” which are like speed limits for how much data you can fetch per time period.
How accurate is it?
On test datasets used in the paper, Botometer V4 did very well (AUC ≈ 0.99), meaning it’s very good at telling apart typical bots and typical humans in those data. But no tool is perfect:
- It can struggle with new kinds of bots not seen in training.
- It may do worse with non-English content if using English-based features.
- Very inactive accounts don’t provide enough data.
- Scores can change over time because they’re based on the most recent tweets.
What did the case study show?
The authors ran a small demonstration using tweets that mention three cashtags (like hashtags but for stocks/crypto): FLOKI (crypto), and SHIB andAAPL when looking at score distributions.
Why this matters: It shows how to analyze Botometer results carefully and how different analysis choices affect the conclusions.
Why is this important?
- Deceptive bots can shape conversations about elections, health, finance, and more. That can mislead people and distort research.
- Researchers need tools to detect and handle bot activity so their studies aren’t biased by automated accounts.
- Botometer offers a well-tested, widely used, and reasonably accessible way to estimate bot-like behavior.
Practical tips and good habits from the paper
Here are a few key recommendations the authors give to help you use Botometer wisely:
- Scores are snapshots: Because Botometer uses the latest tweets, an account’s score can change over time. If you’re doing research, collect tweets and run bot detection soon after.
- Compare groups, not just individuals: Looking at score distributions across topics or time periods lets you do statistical tests and reduces the risk of overinterpreting a single score.
- Validate your thresholds: If you must label accounts as “bot” or “human,” test different cutoffs or, ideally, hand-label a small sample to pick a cutoff that fits your goals (e.g., fewer false alarms).
- Use the right score for language: For non-English accounts, rely on the language-independent score.
- Be civil: Don’t use Botometer to attack people. Even if an account looks automated, it doesn’t mean it’s malicious. And no classifier is perfect.
Bottom line: What’s the impact?
This paper doesn’t just talk about bots—it gives you a toolkit and a playbook. It:
- Explains how Botometer works and how to access it.
- Shows a hands-on example you can copy.
- Clarifies how to interpret scores and avoid common mistakes.
- Encourages responsible, transparent, and statistically sound research practices.
As online discussions keep changing and bots evolve, tools like Botometer—and careful use of them—help researchers, journalists, and platforms understand what’s really happening in social media conversations.
Collections
Sign up for free to add this paper to one or more collections.