Koo, an Indian-made social media platform, is being considered a challenger to Twitter of late. Ever since Twitter defied the Indian government’s orders to block accounts and tweets related to the farmer’s protests, Koo’s currency has been on the rise.
Multiple ministers in the Narendra Modi government — including IT Minister Ravi Shankar Prasad and Railway Minister Piyush Goyal — and politicians from the ruling Bharatiya Janata Party have signed up on Koo, something the company described as a strategic counterweight to the American microblogging platform. Many government departments have followed suit as well. Currently, discourse on app, which won the government’s Atma Nirbhar Challenge in August 2020, largely tilts pro-government. Trending hashtags at the time of writing this piece include “#narendramodi”, “”#protecthindugirls”, “#atmanirbharbharat”, “#indiawithmodi” and “#khalistanplotout”. Hashtags such as “#congress” and “#rahulgandhi” were trending too, but most “koos” using them are critical of the opposition party.
We spoke to Koo co-founder Mayank Bidwatka, asking him about how the company will tackle the tricky subject of content moderation. Bidwakta said Koo would “empower” users to report hate speech, which will then be used to train the machine learning content moderation algorithm. We also spoke about the comparisons between Koo and American social media platform Parler (associated with the pro-Trump alt right in the US), and about the company’s business decision to focus on the mobile users versus the web. Following are excerpts of that conversation:
Note: Quotes have been edited slightly for the sake of clarity and brevity.
On mobile focus vs website
MediaNama: Koo seems to be a mobile-centric application and experience. How different has the response been on the web versus what you have gotten on mobile?
Mayank Bidwatka: You’re right, Koo has been largely designed for the mobile. The way we started out is to cater to vernacular. There are more mobiles than desktops. Hence our priority was to create a mobile experience. The usage on the desktop is largely by organisations, government agencies, professionals who will find it easier to use on the desktop or laptop because they are sitting, working and when they want to Koo, they end up using the desktop. But it’s been designed for the mobile.
There are certain countries where Koo is not yet available but they can access the website. The usage is a bit different. It is easier to do notifications on the app hence the usage is a lot more on the app in terms of engagement.
On content moderation
MediaNama: There is a lot of confusion about Koo’s content moderation. How many people do you have working on content moderation? How much of it is automated?
Mayank Bidwatka: You get a lot of content that is generated every day on the platform. Let’s take an example. If a person could go through a thousand Koos every day, and if you get a million Koos, you need a thousand people to go through them. You cannot do this manually, it’s not economically viable […] It is not possible to have so many people. So you have to depend on technology to be able to solve this moderation problem. You only looking for exceptions and outliers. We are building proprietary technologies to identify outlier content.
A majority of the content is peaceful and harmless. The semantics need to be in multiple languages, like the different kinds of topics. You need to be able to figure out what’s written in an image also, not just text. There are so many angles to content moderation […] See, content moderation has been a problem across the world right. All the global tech giants like Facebook, Twitter, Google have teams that work largely work through technology in the same way we will tackle it.
We want to harness the power of the community. There are more eyes in the community than any number of employees we can possibly have. Giving users reporting tools is an easier way of finding exceptions much faster than you or a team would have found out. We have reporting tools, blocking tools, if someone is reporting we have reporting codes. That makes our life much easier to find exceptions. You posted something and it was hate speech, and I reported it, you will be notified that there is someone who has reported your Koo. If there are thousands of these reports, you will understand that the community did not like what you just said. There is a good chance that you will take it down. This is like a self-moderating feature.
MediaNama: You support many languages. Detection of hate speech isn’t perfect even in English, and not easy in Indian ones. For a language like Odia, the number of users wouldn’t be high, and there isn’t enough annotated data to train the algorithms. Can you really build capacity when that’s not really technologically possible right now?
Mayank Bidwatka: So for a few of these things, you just have to work with dictionaries. Machine learning is basically there to identify patterns. I’ll give you an example. Let’s say you have semantics, you have some sort of dictionary that you have created, that helps you identify certain types of hate speech. You have got another set of data which was given to you by the community. You get multiple sources of data, and that is machine learned on. The machine will show up patterns saying usually these words are there, that’s why you make your dictionary even stronger?
MediaNama: So you’re saying the users will contribute to this dictionary?
Mayank Bidwatka: Absolutely. What happens is that machine learning cannot possibly happen unless there is some initial data. The machine is dumb and smart at the same time. It is dumb about what you are feeding into it. It is smart about recognising patterns. So if I give you all political Koos, and say this is equal to poems, the machine will take whatever you are saying and say “these are the patterns of the poems”. But it is actually political content.
MediaNama: If the users are made a part of the training process of the program, aren’t you including bias into the equation?
Mayank Bidwatka: No you are not understanding. Let’s say I have reported some sort of content and said this is hate speech. Hundred people have reported a hundred different Koos. From this hundred, a team will manually check “is this hate speech or not”. Let’s say it identifies 30 as hate speech. The machine will learn on the 30, not on the 100.
MediaNama: Will your team have the technical expertise to understand the context of that speech?
Mayank Bidwatka: We are a technology company and we have been running vernacular products even before Koo. That’s how Koo started actually. Koo is a part of a company called Bombinate, and the first product was Vokal. It’s all semantics.
MediaNama: Are you confident you’ll be able to get the context of say, anti-Dalit speech? Because a lot of people don’t have that understanding and context. How will you police that kind of speech?
Mayank Bidwatka: Yes, I am confident we will be able to do it. Because there is a certain methodology to identify stuff like that. It’s not very difficult to build. Once you’ve built it you know how to replicate that across different kinds of products.
On comparisons with Parler
MediaNama: Parler was recently taken down by both Google and Apple on the grounds that their content moderation was lax. How are you planning to avoid that?
Mayank Bidwatka: We will not be lax […] Parler had started for a particular reason. Why Koo has started has nothing to do with why Parler started. It’s not the same pattern. We started way before Aatmanirbhar challenge. We started last this year sometime because we understand the vernacular opportunity. Our reasons for existence is very different from why Parler existed in the first place. Had Mr Trump not been removed from Twitter, may be Parler wouldn’t have existed.
[Editor’s Note: Parler enjoyed popularity among the US alt-right long before Trump was suspended from Twitter. In fact, it was after the suspension that Parler was taken down from app marketplaces and had its contract with Amazon Web Services terminated.]
On being seen as pro-government social media
MediaNama: One of the first things we saw was that Koo has a partnership with Republic. That sort of comes with its own crowd right? As a business decision, is it wise to align with a political group since you might be excluding a lot of people that might be interested in a product like yours, but don’t want to be involved in that kind of discourse?
Mayank Bidwatka: I am happy with you also. I am happy to partner with your competitor also. I don’t care what political belief systems you have. Our objective is that we will partner with anyone. I only have friends. I have no enemies. I am blind to your ideology, your religion, your colour. I am blind to your objective. We are an open platform.
MediaNama: From looking at the content right now, it is clear which part of the political spectrum falls under, isn’t it?
Mayank Bidwatka: Let me make it clear. You have to be practical about this. Do you know how Twitter started? It started with the startup world. Startup founders and employees were using this. If I saw Twitter at that time, and I said “hey you guys are biased, you think entrepreneurs are god’s gift to mankind, you think startups are everything”. So there are early adopters, there are guys who adopt later. The larger you become, everyone has to adopt you.
If I look at early adopters, and I start extrapolating that to the lifetime of their companies, that doesn’t make sense. Same way, there are certain parties that understand the importance of something and say they understand the threat of not adopting something like that.
Certain parties will say “this guy is here, I hate him, I am not going to be there.” Guess what, if the whole country is there, you have to be there. You cannot avoid it. Now Republic has understood the value of having an Indian social media. They understand the issues that have happened with Twitter. They have taken a particular stance. They found a company, a very very small company like ours, and they said the whole country should know about this […] Tomorrow, NDTV will come and say you have great content that I would like to put up on my TV. Zee News will come, DD News will come. Everybody will come and partner with us. It starts off somewhere. If you look at the platform and see where it started, that becomes an incorrect thing.
On users and downloads
MediaNama: Can you tell us about the numbers — daily and monthly active users and so on?
Mayank Bidwatka: We are crossing around 4.5 million downloads with the app. The daily number keeps changing. The growth has been good in the last two weeks.
Support for Indian languages
Context: Currently, Koo requires users to choose a particular language, after which they are shown content only in that language. The app currently supports Hindi, English, Kannada, Tamil, Telugu and Marathi, with support for other languages planned soon.
MediaNama: That Koo is available in Indian languages seems to be very important to you. Are you planning to add any translation and transliteration features?
Mayank Bidwatka: The way our product is built is that it is immersive in that particular language. It is not a multi language experience. Let’s take Hindi and English. You will see the feed in the on language you have selected. Yes, you can go and switch the language but it’s one language at a time. So we are not allowing Hindi and English feed in one go. What that means is that you know English and Hindi, you are not going to go to Kannada feed and say “translate this”. That is not the user behaviour among the masses of India. The masses of India are comfortable with the language they know. We have designed this for a one language experience and hence have not done any translation.
As for transliteration, our keyboards are built for transliteration for the creators. If I select Hindi as my language, I go to create I could either type in Hindi, or type in English and it will come out in Hindi. That experience is already built in. We will work on linking the language ecosystems, but it is not an immediate priority for us. We will have it one day.
MediaNama: What are your main priorities currently?
Mayank Bidwatka: One is increasing number of languages. Other set of features will be ability to create in multiple languages in one go. We will look at stabilising the stuff that we have before we start adding new stuff. We have a lot of ideas for building new features.