Flagging hate speech and propaganda on Twitter using AI
Uncover the bleeps posted on social media about the U.S presidential election 2020
My Role in the startup
Product Designer (Remote)
Timeline
June 2020 - November 2020
*NDA is signed and only the data that can be displayed in public is mentioned here.
Overview
Bleepr.ai is a feed which flags hate speech and propaganda and its called as hateful feed in the beginning. Eventually, with the current pandemic and elections in the US, we decided to align the feed path towards those topics. I'm responsible for the whole design process and final designs.
Startup
in partnership with
Worked with a bunch of experienced and talented people
Dhruv Ghulati
CEO
Sameh Frihat
Data Scientist
Katia Stambolieva
Product Manager
Hasan Halacli
Web Developer
Ming Wang
Web Developer
Bleepr.ai is the #5 Product of the Day on ProductHunt
Context
Bleepr uses natural language processing (NLP) to automatically flag hate speech published on social media. In our first release, we scan Tweets of users with more than 10k followers, automatically label for 7 dimensions of unsafe language, and update daily.
Open the browser and visit bleepr.ai
You will find all the posts rated from different social media sites
One can give feedback on whether they agree with AI or not.
Problem
No one can access FB/Twitter closed platforms. When you flag something to them, not clear what happens next. They are slow at taking things down and have opaque policies. Lets put public pressure on them as the general public.
Goal
Be the one public newsfeed/destination everyone sees what platforms aren’t taking down.
Solution/Mitigation
We build a feed with the most hateful and actual topics that the community should see. This data builds an open-source hate speech algorithm.
Early Explorations
A stream of auto-scored content for misinformation, fake news, racism etc with the best possible models from keyword search and custom source lists. We called this "The Hateful Feed".
​
-
Display names as a key thing rather than Tweet e.g. this person said this.
-
We put the scores on each content item for each model
-
Hate speech, general misinformation, racism, sexism, threats, toxicity, political bias, obscenity, insults
-
-
Overall risk score (high, medium, low) on a scale
-
Show newest to oldest flagged items first, with the date on each content item.
-
Simple feedback like "Agree with AI", "Disagree with AI" on each flagged item. Just 1 feedback, not per model.
-
We use topic detection to auto-tag the content with topics
Early exploration called Hateful Feed
Bleepr Brand
The creative process of branding which is done in half day set off with logo design. It started with the brainstorming and analyzing vision of a new symbol. As we wanted the modernization, a new brand sign was expected to be more playful and fresh still simple with vision reflecting elements. In addition, I'm asked to keep the red colour palette as it corresponds to their vision of the product.
When all the tasks were considered and the direction was chosen, I started the creative process. The pencil sketching technique was applied to quickly visualize the first ideas on the paper. Such an approach allows for picturing various concepts without significant efforts and within a short time. Here are quick explorations for bleepr logo.
Sketches for logos in an hour
Wavy option based on the context that bleep is a sound with the entire word in it.
After some discussions, it was decided to step back to less abstract and simple shape with the minimum of details and reflecting the sense of harm.
A variant featuring a less abstract logo with the easily recognizable shape of an exclamation which signifies the harm present in text
New features and updated designs
We updated the hateful feed to bleepr feed under the light of new activities in the world like the US Elections. For more user interaction, we added some additional features to this as well.
Posts monitored from multiple social outlets
Showing data from FB and Reddit, and not only Twitter. We scan posts in bulk every day. We look at around 500 posts in more detail - and only those which were posted by profiles with more than 10K+ followers.
Highlighting the threat
We run our ML algorithms on those posts. Every post which has at least one of the automated labels ((sexism, racism, hate speech, obscenity, political bias, toxicity, insults and threat level)) with a score above a certain threshold is then listed on bleepr.ai.
We highlight these labels in red.
Bleep of the day , Bleep of the week, Bleep of the month
We added a new feature called bleep for the day, week and the month which is actually the post that contains more bleep in it than the others.
Bleepr.ai in action
Fortunately, we updated the hateful feed into the bleepr.ai but still, there are some things to be fixed and bugs in the product from the technical side.
Known issues
-
Sexism detection flags up higher than it should sometimes for words like ‘bi**h’ and ‘motherf****r’ even in contexts which are not hateful.
-
There are a few bleeps which contain swear words but are random and obscure and not truly hate speech Sometimes we flag news stories about controversial topics like homophobia, paedophile or sexual abuse, rather than original hate speech
-
We sometimes flag Tweets don’t exist anymore, which means Twitter has luckily already taken them down or authors have removed them.
-
Sometimes we flag Tweets that talk about political entities in a fairly coherent way, but use the words idiot a lot
What is coming next for bleepr.ai ?
Future releases include using human curation of Bleeps using a small curated QA community, adding more social media platforms to Bleepr, improving the models, and having live refresh of likes/shares/retweets.
Closing Notes
Help us make social media a safer space. You can do this by clicking on the AGREE/DISAGREE WITH AI buttons which are shown under every bleep on bleepr.ai. If you would like to contribute to our algorithms or join our community to take Bleepr to take it further, please contact us on info@factmata.com. We are just getting started.
​
This feed is one of its kind and I was able to build it from scratch and I'm looking to improve it better in the
future. I'm glad that I'm part of this.
Have a look at other recent works
