Welcome to issue #2 of our recently launched innovation Q&A series, Startup Spotlight.
This series is all about diving into the world-class technological innovation conducted by exciting UK startups. Getting to grips with the technology, the mission, and even the stories behind them.
In each edition, we sit down with a different startup founder, leader, or operator in an easily digestible conversational style Q&A format.
Today, we interviewed Antony Cousins, CEO of Factmata about his startup's revolutionary machine learning technology that can immediately figure out the context behind online content. Things like bias, prejudice, and narratives.
- ⚖️ How Factmata measures to what degree content is racist, sexist, or uses hate speech
- 🤗 How an agency used Factmata to help a huge consumer brand with its brand reputation
- 🔮 How Factmata could eventually predict a brand's future sentiment
Do you trust fact-checkers? If the answer is no, it’s understandable why you don’t.
Sentiment analysis is still a human led thing, and if there are humans at the helm, it takes a ton of time and can be biased in itself.
So, what if everything was handled by an AI? No bias, no mistakes, just hard hitting insights that recognise the ins and outs of your content?
That’s what the team at Factmata has achieved.
With Factmata and its offshoot Content Score, brands and companies can instantly find out how content scores, from an insight perspective. It reads content for context and can detect racism, sexism, and hate speech.
After that, brands can use these results to guide their content and brand voice in a better direction. Or stick with it if that’s the goal.
But what exactly is Factmata, and what is Content Score?
How has it been used already?
And, what are the future possibilities with this type of technology?
We interviewed Antony to find out the answer to these questions. Enjoy! 👇
So, Antony, what is the mission of Factmata?
The overarching mission, is to make the internet a better and safer place for everyone.
We started out back in 2017, introducing automation to help with the problem of misinformation, disinformation, and fake news.
Solving that issue is still the same mission today. That’s the mission that we've worked on for a few years, and what we’ve based our underlying technology on.
A noble goal. But whilst noble, how exactly does Factmata monetise this?
Who is paying the most to distribute content on the internet and has their hard-earned reputation on the line? Brands and organisations.
We've got the underlying tech to mitigate reputational damage, but we want to avoid being the company that tries to solve this problem on behalf of everyone. It's not feasible.
To scale a business we realised that we should, to use a gold rush analogy, pursue a 'picks and shovels' strategy.
We want to avoid prospecting for gold nuggets, panning in the streams. Instead, we’ll sell the pans, picks, and shovels, to people that do. For example: selling our solution to agencies, who work with brands directly. It's self-serve and starts at $300 per month.
Some are still sceptical about fact-checkers in general. Why should they trust Factmata?
People don't trust governments anymore. People don't trust the media anymore. Weirdly, they do trust businesses, even though businesses are self-serving at core, they have to make profit to survive, not make the world a better place.
It's an interesting kind of juxtaposition. Because of that trust, we want to target brands to get them involved in this fight.
Effectively, they’ll have the tools they need to find misinformation, disinformation, fake news, and everything else which makes the internet a bad place. Toxicity, hate speech, racism, sexism, for example.
We noticed you deploy your technology into two differently branded offerings, one called 'Factmata' and the other 'Content Score'. Can you tell us about that, and why you did it?
We found out that our underlying technology has multiple use cases. The previous challenge was having one website, explaining all of those different uses, to entirely different audiences.
It was difficult from an SEO perspective.
That’s why we created Content Score, as a company, to focus on the use of the underlying tech itself.
So if you're a platform, or a company, and your user-generated content is being shared, how do you make sure they stay on top of this kind of content, that's going to get reported?
We built Content Score to provide that service for them, to prevent advertising on content that would not align with their values.
Things like racist content, sexist content, and COVID-19 misinformation. Whatever doesn’t align with their brand.
Currently, there's no alternative way to protect themselves, so Content Score reads URLs, and comes back with a safety score. It can say “don't post here, it's racist” or “don't post here, it’s sexist”, for example.
So, any brand can use Content Score to measure the safety of content, separate from Factmata? That’s pretty handy.
Let’s talk about R&D. What has been your technical journey? How has this developed over time?
Obviously, we started out in the automated fact checking space. What we were trying to answer was how do we analyse large quantities of data? How do we identify within the claims that are being made? What are people talking about? And then how do we automate the process of validating those claims against independent sources?
That's the kind of process we went through. What was concerning was this would make us responsible for what we are saying is truth, and what is also a lie.
It’s a space we do not want to be in, as the subjective nature of truth is perspective, and not commercially viable as a business.
So, that's why we created content classification models, which have the ability to read short form and long form text, and figure out what’s racist, what’s sexist, what’s hate speech, etcetera.
We’ve developed two further pieces of AI technology. One is topic clustering which, again, reads long form and short form content, and identifies similar opinions and groups them together. The other, is summarisation.
For summarisation, we generate a human-readable sentence, which explains the context of that grouped together content.
With these 2 technologies combined, they power our Factmata narrative monitoring platform, and it does so to great effect. It’s wonderful.
Topic clustering and summarisation sound like powerful tools for any brand concerned about its image.
We’ve heard you were involved with an agency that is working on behalf of huge consumer brand? Can you tell us about that?
Sure. The challenge was this particular brand, was that it is so big, with a huge number of mentions online across platforms and countries.
If you wanted to analyse everything that’s been said about them, you need a massive team of people to do it, every day, to process the mentions.
To understand the context, meaning and sentiment of every single mention is tough. But with Factmata, we can tell you in seconds. We can say to clients, “You don't need to do that. We'll tell you what it's about. Here's the sentence, which you can understand”. It works. Really, really well.
Once Factmata has worked out a narrative, it can do multiple things. It can work out if the narrative will be bigger tomorrow than yesterday, if the number of people involved are increasing, what the context of your response has been, how engaging the narrative is, and so on.
Once you have those narratives, the ability to analyse further, and measure more metrics, becomes fascinating. And that's where the value for a big company like this one is.
Let’s say a company finds Factmata, but has no clue on how to use it properly. What would you tell them?
If you're a marketer in any industry, you'll be familiar with sentiment analysis, which is measuring how people respond to using positive, negative, and neutral words.
If a brands' sentiment is all negative, it still doesn't help you identify whether people are specifically for a certain opinion or against a certain opinion. That's why a brand needs to make informed decisions.
With our topic clustering, combined with a specific model we've created called “Stance”, we can actually identify the intent of the people involved in these narratives towards the overall topic:
For example, if Factmata can work out if people are for or against a brand over a particular topic, they can actually make educated decisions based on that.
Here’s an example of the thought process an agency or brand would go through over something like that: If we want to weigh into this conversation with a certain point of view, how many people are for us? How many people are against us? Should we not get involved in this at all? Would we write off this segment of our audience by making a decision? How much is that worth to us?
Factmata’s still not going to be 100% accurate, as it’s near impossible to fully work out the buying power of a particular group. But we are heading in that direction. And it’s super exciting stuff.
This is amplified by the rise of 'cancel culture', which is 100% here to stay. I think the challenge for brands going forwards will be how quickly can you figure out which segment of their audience will they want to keep happy?
You can’t keep everyone happy. In some cases, not acting is seen as an action in and of itself. We've seen this lately with brands that have failed to take a position on Russia, particularly if they have business interests there.
How does Factmata and Content Score compare against competing offerings? What’s your unique selling point?
No one, that we know of, is using anything more complex than sentiment analysis, which has been around for 12 years.
It’s great at giving you basic facts, however. For example, using sentiment analysis by channel, it’ll give you a word cloud. It’ll break it down for you. But the actual insights are still the responsibility of humans, which our API automates.
That's the USP, what everyone wants. People don’t truly want the analysis, what they want is the insight. That's the bit we're building AI for. And that's where we differentiate.
What you have is exciting, and impressive by technological standards.
But what does the future hold? What’s the next level?
We're still refining the experience for getting those insights quickly to the brands that need them. But, the next cool step is prediction.
Prediction? Can you elaborate?
Currently, we're telling brands, here's what people are saying, here's the kind of people that are saying that, here's how many of them are saying this, here's the level of engagement right now, and back in time, and so on.
Basically, we can show you the trends until today. The exciting part is using different technology to work out the events that have contributed to these trends to make the predictions.
It's going to get less and less accurate, the further into the future you look, but what we can tell you is this.
This where you are today. If you don't do anything different, here's where you’ll be a week from now. Your stance will go through the floor. Your sentiment will go through the floor, the engagement with these particular narratives is going through the floor.
Essentially, we’re saying that a week from now, if you don't do something, you're going to end up on the front page of every news outlet in the world.
The eventual goal is future impact based on historical action. That’s the future for Factmata.
Amazing. You could use such insights and pair it with machine learning generated copywriting technology? Is that a possibility for you?
We’re also working on that. Can we automatically generate the copy that brands actually might need to use to counter a specific narrative? Would it be helpful?
Even if we do this, a human would still need to look at it. But can we give them a head start? Can we give them hints? Can we provide them with the specific words you have to include in their copy?
If we can automate the creation of the copy, you could even automate the identification of the right influencers to gauge on that topic? If it could, that’d be a gigantic time saver for many.
As a final question, it seems like your technology could also be utilised to help a brand target specific audiences with strong political or societal views. Is that a use case you have seen?
We haven't noticed anyone doing that just yet. And technically, yes, it is possible.
The issue is that people are still trying to wrap their head around narrative clustering. We're still educating the market on the value of that.
Everyone is so used to doing it the old way. So, for now, I think we'll have to pick it off a bite at a time.
That's it for today! To receive further content like this (and more!) drop your email below. 👇