Artificial intelligence can combat political bias. Yes, we all know that AI has had its own problems with bias in the past (and present). However, one news website is harnessing AI in the fight against overt political partisanship. And in age when the world is becoming more politically polarised, its efforts couldn’t be more timely.
This website is the Bipartisan Press. Founded in 2018, it has developed an AI model for determining the political bias of its own articles and any text you might find on the web. Based on a regression model for machine learning, it’s capable of natural language processing (NLP) and of text summarisation. And because it has been trained on a large database of articles (pre-categorised according to bias), it can classify texts according to their direction (“left” or “right”) and degree (“minimal” to “extreme”) of bias.
Essentially, the Bipartisan Press’ AI model can learn to reproduce classification systems devised by political scientists and journalists. And according to the website’s own research, it can classify the bias of articles to a 96% accuracy, with an average deviation of only 7%.
For now, the Bipartisan Press is using this model predominantly to classify the bias of its own articles, which receive scores from -42 (left-wing) to 42 (right-wing). This lends greater transparency to its own output, helping readers to judge just where it and its writers are coming from on the political spectrum. And seeing as how trust in journalism has sunk to all-time lows in recent years, such transparency could go some way to restoring public faith in the media.
“The site was started in 2018, around mid-way through Trump’s administration, to address the growing issue of media bias and political tribalism,” explains Winston Wang, the site’s founder.
“Initially, we published articles on our site where we included the viewpoints of various political ideologies on a certain political issue. However, we soon discovered this was logistically unfeasible at our current state, so we transitioned to publishing various opinions and pieces separately. To remain in the spirit of being ‘bipartisan’ we developed the rating system that we now show on all our articles, notifying readers of their bias. This ultimately led us to develop our AI bias system we now have.”
Not only does the Bipartisan Press want to be open about its own political viewpoints and inclinations, but it’s opening up its AI classification model to third-parties, so that other outlets have access to a quick and easy way of indicating political leanings.
To begin with, anyone can currently use the Bias Analyzer on its website to calculate the direction and degree of bias of any text they copy and paste. In fact, I copied and pasted the draft of this article into the Bias Analyzer, finding out that it apparently has a “minimal left” bias. The website also has a political bias bot on Reddit that can analyse the bias of any article posted to its sub-Reddit.
However, the website is also planning to launch a browser extension, which you can use to analyse the political bias of any website or page you happen to be browsing.
“We estimate our own extension can be launched in the next two or three months,” Wang tells me. “We actually currently are partnered with Daratos, which has an extension in the Chrome Web Store that uses our API to analyse bias. We do still plan on developing our own in-house extension that will be compatible with any website. The extension would, when activated, extract the text content from an article, analyse the bias, then return the bias and display that in the extension.”
An AI-powered browser extension may not be an especially ‘momentous’ product to launch, but it couldn’t come at a more opportune or urgent time. Given the increasing bias and subjectivity of many media outlets, it would be highly beneficial to provide the public with a means for discovering that the news they read is hardly ever ‘neutral.’
In terms of speed and scale, an AI-based tool can do a much better job of bias analysis than any human individual or team. Theoretically, every article on the web can be provided with a near-instant political bias score. Not only would this inform readers and voters, but it could exert pressure on media outlets to make their news coverage more balanced and factual. In turn, this would ultimately result in more well-informed readers and voters.
And the Bipartisan Press isn’t the only organisation working on exploiting the power of AI for the purposes of exposing bias. The likes of MIT, Purdue University and the University of Maryland have developed their own AI-based bias-detecting tools in recent years, paving the way for other companies or websites to produce their own solutions.
Of course, it would be naive to assume that AI alone will solve the political bias and polarisation issues currently plaguing print and digital media. For one, the Bipartisan Press admits that its model isn’t 100% accurate, in the sense of detecting bias exactly where a trained professional would detect bias.
On top of this, even if an AI could provide a 100% accurate replication of how human experts categorise bias, this would still potentially be problematic. That’s because one method or standard for categorising political bias may itself seem biased relative to some other method. In other words, what’s ‘centrist’ for one person may be ‘left-wing’ or ‘right-wing’ for others, and so on.
As such, all a bias-detecting AI model may achieve is the imposition of one political worldview on thousands or millions of people. The Bipartisan Press recognises this difficulty, but says that it can be avoided or at least mitigated with the use of databases that attract broad majority agreement.
“Fundamentally, when AI is biased, it’s almost always because of the dataset,” says Winston Wang. “In this case, we used data from Adfontesmedia, which uses a very extensive methodology to classify bias, including having three people of various political leanings rate each article.”
There are, of course, some caveats, Wang adds. “The definition of political bias has been notoriously hard to pin down, and political ideologies and wings greatly differ by country. To alleviate this, we simply used an authority that most people agreed with, audited it, and trained the AI on that.”
Yes, AI models may come with biases of their own. But generally, they will be no more biased than any alternative ‘manual’ system of classification. On top of this, they increasingly force us to be more conscious about how we arrive at classifications and build classificatory models in the first place. So with websites like the Bipartisan Press and with a growth in AI-based bias tools, we may take important steps towards improving how we inform the public.