Christopher Bouzy

Christopher Bouzy @cbouzy

You said Bot Sentinel isn't perfect, but it's the most accurate publicly available tool. What did you mean? Anyone who understands machine learning and algorithms knows perfect doesn't exist... yet. But I believe Bot Sentinel is the most accurate publicly available tool... — PolitiTweet.org

Christopher Bouzy @cbouzy

Do you manually rate accounts? No. It would be impossible to rate millions of accounts manually. In a podcast, I said we manually flag fake news tweets for Tweet Alerts. That is something entirely different than rating individual accounts. Also, Tweet Alerts is now automatic... — PolitiTweet.org

Christopher Bouzy @cbouzy

Why does Twitter suspend accounts that are rated normal? An account rated normal doesn't mean it isn't violating Twitter's rules. The account could've been evading suspension or tweeted something bad enough to warrant a suspension. Twitter suspends accounts, not Bot Sentinel... — PolitiTweet.org

Christopher Bouzy @cbouzy

What about false positives? Yes, false positives are possible. However, Bot Sentinel classifies each tweet separately, and 300+ false positives is highly unlikely. An account that is rated problematic should be avoided, especially if the account has over 400 tweets... — PolitiTweet.org

Christopher Bouzy @cbouzy

The beauty of Bot Sentinel is that a few problematic tweets will not affect the overall rating. Everyone has a bad day, and even I have a few problematic tweets. Scores are eligible for update every 7 days, so someone can improve their score by not tweeting problematic content... — PolitiTweet.org

Christopher Bouzy @cbouzy

So what if someone has a 0% score? There are 3 reasons why an account receives a 0% score. 1. No problematic tweets were found. 2. No English tweets were found. 3. The account doesn't have at least 10 tweets. [I am taking a 30-minute break] ... — PolitiTweet.org

Christopher Bouzy @cbouzy

For example, an account with 100 tweets and a 15% Bot Sentinel score means Bot Sentinel identified 15 bad/problematic tweets. An account with 100 tweets and a 75% score means Bot Sentinel identified 75 bad/problematic tweets. Bot Sentinel analyzes a max of 400 recent tweets... — PolitiTweet.org

Christopher Bouzy @cbouzy

So how does Bot Sentinel rate an account? The classification process is the most confusing part for most people. Bot Sentinel classifies each recent tweet, and the rating is based on the number of "bad" tweets. The more bad tweets that are identified, the higher the score... — PolitiTweet.org

Christopher Bouzy @cbouzy

Conversations on social media platforms change rapidly, so a classification model that is 4 or 5 months old will no longer be reliable, which is why we update our models every 6-8 weeks. FYI: Our models are several gigabytes and loaded into memory... — PolitiTweet.org

Christopher Bouzy @cbouzy

Once we are satisfied with the model, it is uploaded to our production servers for public use. Sometimes people notice an increase in their scores when we update the classification model, and it's perfectly normal. It's a new model trained with completely different samples... — PolitiTweet.org

Christopher Bouzy @cbouzy

Once the model is trained, we test the accuracy using the accounts we monitor. Benchmarking and testing the model takes several days, and if we are not happy with the results, we start the training process over again. Using a poorly trained model would be disastrous... — PolitiTweet.org

Christopher Bouzy @cbouzy

Note: In a recent interview with Wired, I did share the proprietary step with the reporter who interviewed me. I actually shared a lot of what we do off the record... https://t.co/wga73c3OkO — PolitiTweet.org

Christopher Bouzy @cbouzy

Selecting the samples is the most crucial step. We added an extra proprietary step to the process to ensure the samples we use will improve the accuracy of our classifier. I can't share that step, but I can say monitoring accounts helps us improve the sample selection process... — PolitiTweet.org

Christopher Bouzy @cbouzy

We have built a tool to automatically label samples for the training process, but we also manually review samples because selecting samples isn't always clear-cut. Example: I want to mate my b*tch with a stud vs. I am a stud who can mate with any b*tch I want... — PolitiTweet.org

Christopher Bouzy @cbouzy

Here are a few examples: Good tweet: I love avocadoes. Bad tweet: I want to shove an avocado up your a.. Good tweet: Who is celebrating Kwanzaa? Bad tweet: Kwanzaa is for dirty n***ers. Good tweet: I want a new car. Bad tweet: I want to k*ll j*ws with my car... — PolitiTweet.org

Christopher Bouzy @cbouzy

So how exactly does Bot Sentinel work? It's pretty straightforward. We use millions of samples to train our models, labeling the samples "good" and "bad" tweets. We use Twitter's rules as a guide to determine which tweets are labeled bad. Sample selection is everything... — PolitiTweet.org

Christopher Bouzy @cbouzy

The same is true for human-controlled accounts. Human A tweets about baking and cats. Human B tweets election disinformation. Human A will receive a low score, while Human B will receive a high score. All Bot Sentinel cares about is what the account is tweeting... — PolitiTweet.org

Christopher Bouzy @cbouzy

Bot A tweets inspirational quotes. Bot B tweets COVID-19 disinformation that violates Twitter's policies. Bot A will receive a low Bot Sentinel score, and Bot B will receive a high Bot Sentinel score. Both are bots, but Bot A is harmless, while Bot B is problematic... — PolitiTweet.org

Christopher Bouzy @cbouzy

Just because an account is automated doesn't automatically signify a nefarious account. There are thousands of useful and benign automated accounts. Bot Sentinel doesn't care if an account is human-controlled or automated; it identifies policy violations in tweets... — PolitiTweet.org

Christopher Bouzy @cbouzy

The name Bot Sentinel means autonomous guard. In other words, we use a good "bot" to help identify and guard against inauthentic accounts and toxic trolls. That also includes automated accounts that violate Twitter's rules. Bot Sentinel classifies a lot more than just bots... — PolitiTweet.org

Christopher Bouzy @cbouzy

🧵There has been a lot of mis/disinformation about Bot Sentinel and how it works, so this thread will explain the training and classification process and how to use our analyzer properly. This thread will also answer some of the most common questions we receive... https://t.co/0uIASs53kK — PolitiTweet.org

Christopher Bouzy @cbouzy

A few of you have asked me if Twitter Blue is worth it, and the short is yes. I have used it for several months, and I would never recommend it to my followers if I didn't think it was worth it. The undo button has grown on me, and I use it several times a day. — PolitiTweet.org

Christopher Bouzy @cbouzy

@vmakinstall You need to be a Twitter Blue subscriber. — PolitiTweet.org

Christopher Bouzy @cbouzy

I was against the edit button, but Twitter did a good job. I still think 30 minutes is way too long, but overall I am happy with it. To be fair, Twitter is getting better with its feature updates. The mixed media update is a home run. — PolitiTweet.org

Christopher Bouzy @cbouzy

There is a lot of that going around lately. Did he fundraise for the lawsuit? — PolitiTweet.org

Christopher Bouzy @cbouzy

Sorry, I had to do it. https://t.co/ps5EqTSBnF — PolitiTweet.org

Christopher Bouzy @cbouzy

Ummm https://t.co/cqJ9tf5kty — PolitiTweet.org

Christopher Bouzy @cbouzy

RT @notcapnamerica: Do your due diligence when you're applying for these WFH jobs. There's so many scams going on right now and they're try… — PolitiTweet.org

Christopher Bouzy @cbouzy

And yet, here you are, tweeting at me.🤷🏾‍♂️ — PolitiTweet.org

Christopher Bouzy @cbouzy

RT @TristanSnell: BREAKING: 2 Russian refugees fleeing Putin's draft sailed to Alaska and have asked for asylum in the US. Next, will Ron… — PolitiTweet.org