Can AI help ‘preserve quality of the Internet’?
A rising startup aims to tackle disinformation using ‘quality metrics’ based on algorithms and expert knowledge
The number of fact-checking initiatives have mushroomed in recent years and yet many argue they are a losing battle against widespread disinformation. GEN spoke with Dhruv Ghulati, co-founder and CEO of Factmata, to understand what makes this rising AI startup unique and how it uses quality metrics to offer a kind of ‘nutritional label for anything you read online’. Ghulati says fact checkers are losing the battle because they are not on the offensive enough, and sees machine learning based on community and expert knowledge as a potential solution. GEN inquired about the inside-workings of Factmata and the limits of fact checking in the fight against disinformation.
GEN: Today fact checking initiatives abound around the world. What is different about Factmata (for example, in comparison with NewsGuard/Deepnews.ai)?
Dhruv Ghulati: Newsguard is humans rating publishers according to rules they develop as markers for good journalism, and how those publishers have historically followed these rules. Deepnews.ai is a black box deep learning system which rates news articles for “quality” and then correlates quality scores with publisher engagement metrics.
I personally believe Deepnews.ai is a super elegant solution, we are quite similar to them. But slightly different because we try not to produce black box predictions.
When we tell you an article is quality or not quality, you will see its scores given across each “quality metric” we build (e.g. how opinionated it is, how balanced it is etc is all a separate automatically calculated score) and even which words or sentences caused these judgements. This is more challenging, takes more time, but we think the way to go. I might be wrong, but I think having a group of like minded journalists rate a publisher for ethics is bound to be more biased than an open algorithm with open test data which rates what people write, not who they are or have been in the past. Finally, we want to publish our test data (i.e. what we deem to be opinionated, balanced, etc). This is a unique, transparent stance and I want more AI companies to do this.
No solution is perfect, I know every solution is trying very hard to get things right, so all credit to all initiatives.
Factmata uses ‘AI, communities and expert knowledge to identify and classify content’. Can you tell us more about how you select relevant communities and experts? What’s the role of AI?
Right now, mainly due to small scale, we try to do the smallest thing for the highest potential output. So we self-select who we think is the most relevant community for the problem accounting for factors like if they are happy to volunteer, if they have capacity, if they want to partner with us for the long term, etc.
An example project we’ve worked on is with Jugenschutz in Germany, who have worked with us to define fair definitions of hate speech. Their incredible community works to set these definitions along with a raft of other partners, we debate together as a group, and then Factmata adds the layer which builds the algorithms on top of these definitions. We are always open who our partners are and are open to having them reviewed regularly, but ultimately this is how we start.
I might be wrong, but I think having a group of like minded journalists rate a publisher for ethics is bound to be more biased than an open algorithm with open test data which rates what people write, not who they are or have been in the past.
Who are your major media partners? Do you work with legacy media as well as online outlets?
We are trialling with a number of major news organisations, mainly online outlets given we have to process text using machines. I cannot name the outlet, but we will be working with one of the largest media companies in the APAC region very soon.
Fact checkers are losing the battle because they are not on the offensive enough.
You compare Factmata’s scoring system to a ‘nutritional label for anything you read online’. Some would argue that a long-term solution to the disinformation problem is to invest in media literacy and critical thinking, so that people can decipher information by themselves instead of being provided with pre-validated/-approved facts. What are your thoughts?
I totally agree. A nutrition label never tells you what to do via a recommendation. It gives you more context to make the right decision yourself. The label might be yellow or unsure, but hopefully it gives you details of why its in the middle of the rating so you can decide either way (to eat the cookie or not, so to speak).
As platforms too are enhancing their fact-checking tools and adding new ones, could they take over smaller, independent fact checking initiatives? How do you envisage the future of Factmata in this regard?
We don’t compete with platforms. Their goal is to do everything in house and work with independent fact checkers to train their AIs (which I fear will mean ultimately the human fact checkers become irrelevant one day). Factmata’s goal is to provide algorithms to the rest of the internet that does not have the resources of a FB or Google, but still needs urgent help in content moderation, fact checking and more. We also provide products and services outside the media world, e.g. detecting rumours spreading about brands, as part of our business model.
Communities we speak to say they want to work with us much more than the larger platforms, for a variety of reasons. For example, we want to one day provide a revenue share to those we work with from the algorithms built from their data. This depends on Factmata reaching scale but this is our goal — we are on the side of the communities and journalists who ultimately form the backbone of the media ecosystem, yet get paid the least. My dream is fact checkers and journalists finally get rewarded for the invaluable work they do protecting the internet.
Despite the proliferating fact checking initiatives, disinformation remains on the rise and trust in news is declining. What do you think are the limits of fact-checking in combatting disinformation?
Fact checking is a valuable art, but only recently are people thinking of bringing the expert knowledge of a fact checker to a machine. Full Fact has in particular been doing great work here and should be applauded. I feel we need to tackle disinformation with equivalent tools on the fact checking side (i.e. strong distribution capabilities on facts, potentially even bot networks to spread facts or automated article correction mechanisms). This is a controversial view but in my opinion, fact checkers are losing the battle because they are not on the offensive enough — it is more a defensive layer on top of very powerful, well funded initiatives to spread lies and rumours that if unchecked I fear will get more powerful. Journalists who often spread biases also will have to be more careful about how they frame articles and be held to account.
Which recent/upcoming fact checking initiatives are you excited about?
Definitely Full Fact’s recent funding for an automated fact checking tool for journalists, due to be launched in 6 months I believe. I am also super excited about the upcoming Truth and Trust Online conference, which will bring the top researchers in this space in one place to discuss progress.
Dhruv Ghulati is CEO and co-founder of Factmata, an AI startup developing community-driven algorithms to solve the problem of online misinformation and create a quality media ecosystem. A Forbes 30 Under 30 leader in technology for Europe, Dhruv has built startups at Entrepreneur First and Techstars London. Having started his career in finance at Bank of America Merrill Lynch, he transitioned into being a product leader, engineer and scientist in the space of artificial intelligence and data science. Dhruv holds a 1st Class MSc in Computer Science from University College London, and a 1st Class BSc in Economics from the LSE.