Shit, Sherlock
A BBC ‘study’ into abuse sent to MPs is flawed, dangerous, and already being treated as gospel.
Previously: Lucan in all the wrong places
Yesterday, Peter Sherlock, editorial lead at the BBC Shared Data Unit, tweeted:
This has been an eight-month labour of love… using a machine learning tool to analyse three million tweets.
Just how does toxicity on social media shape our democracy?
It was a ‘bold’1 way to introduce the report and, at the time of writing, he hasn’t tweeted again. Why? Because that “eight-month labour of love” was rapidly revealed to be the product of terrible methodology and dubious assumptions.
The tool used by the BBC to analyse the tweets covered by the research — it almost physically hurts me to describe it like that — is Perspective API, a set of machine learning models designed by Jigsaw, a Google research unit, for content moderation. Its developers specifically say is “not meant to completely replace the work of human decision-makers.”
As Ash Sarkar writes in her dissection of the project’s many and manifest failings:
The AI appears incapable of distinguishing between swearing used in an abusive context, and swearing used in an innocuous one.
What’s more, it proves utterly useless when it comes to detecting racist content. I spent the afternoon playing slur bingo, to see what would and would not be picked up as a toxic tweet. I entered a selection of racist tweets I’d received in the past year into Perspective’s API, of varying lengths and sentence complexity. Just to make it easy for the machine, I deliberately chose one which included a common racial slur against South Asians.
None of these were registered as potentially toxic at all by the AI – but, “You’re a fucking G”, a compliment, popped up with a 90.29% likelihood of being toxic.
She correctly concludes that the “report is what academic peer reviewers would call ‘a load of horseshit’” (a comment rated 85.33% likely to be toxic by Perspective AI). The responses to Sherlock’s original tweet provide plenty of other useful critiques of the methodology. What I want to consider is what this report tells us about editorial processes at the BBC, assumptions in the British political media, and the corporation’s attitude and approach to criticism.
The BBC team accepted Perspective’s definition of a toxic comment — “rude, disrespectful or unreasonable” and “likely to make someone leave a conversation”. This entirely ignores the power imbalance between MPs and members of the public on Twitter as well as the context of political debate.
After 8 months working on this project, no one at the BBC seems to have questioned results that produced a list of “most abused” backbenchers topped by Ben Bradley — a Tory councillor and MP who has been forced to apologise for his own offensive tweets — and featuring only one ethnic minority MP.
They also don’t seem to have considered whether setting the bar for “offensive” tweets as low as using the word “stupid” — “This policy is stupid.” (71.76% likely to be offensive according to Perspective) — would produce a set of meaningless results. Nor that Perspective being predominantly trained on the comment sections of US news sites would lead it to miss the nuances of British discourse.
Compared to the frequently-cited 2017 study by Amnesty — which found that Diane Abbott received a third of all the abusive tweets it catalogued between January and June of that year — the BBC research did not cover a general election but was limited to a six-week during the normal run of parliament. That was almost certain to lead to an over-representation of Tories in the findings as they represent the governing party which naturally receives more coverage than the opposition outside of elections.
A good example of how little thought has been applied to the ‘results’ comes from the BBC’s longer document on the research which is hidden away on Google Docs, with its permissions to copy and paste disabled. The authors write:
Tory MPs were twice as likely to receive a toxic tweet as a Liberal Democrat.
That ‘fact’ is stated without the context that there are just 14 Lib Dem MPs in the current parliament and they receive substantially less press coverage than Tories. If the Lib Dems weren’t so lesser-spotted, they too might receive more criticism.
The BBC’s use of language and choice of quotes is crucial; it features MPs talking about the most extreme comments they receive while blithely talking about large numbers of tweets (“Former Prime Minister Boris Johnson received the largest number of tweets considered toxic at 19,000…”) that will have included justified criticism.
In its longer report, the BBC claims that “MPs who self-defined as being from an ethnic minority background were not more likely to receive a tweet rated above the 0.7 toxicity threshold”. But we know that “threshold” is hugely flawed and relies on a tool that does not consider a wide range of slurs and is unable to detect common forms of abuse. When the BBC makes that kind of spurious claim, it’s offering a new settled ‘fact’ that racists can quote.
But then the approach of the BBC researchers pushes a perspective in which a tweet calling someone racist is considered worse than many tweets that are racist. It is a study that seems designed to minimise abuse targeting ethnic minorities and LGBT+ people while treating messages that criticise misogyny, racism, transphobia, and homophobia as the real source of “toxicity”. It’s gaslighting with a spreadsheet.
Now, after a huge amount of detailed takedowns on Twitter, articles outlining the deep flaws in the ‘research’, and even a front page denouncing it, the BBC is keeping its head down. That’s a common response by the corporation to criticism of its output; it turtles up and hopes that the noise will quickly go away. But we absolutely should not shut up about this.
In the BBC News story, Ellen Judson, head of the Centre for the Analysis of Social Media at Demos, calls social media “democratically dangerous”. But what’s actually more “democratically dangerous” is the UK’s national broadcaster conflating sharp criticism — “liar” and “disgrace” are among the top terms it frames as toxic — with violent threats and abuse. This bad research provides further cover for government plans to restrict speech on social media under the guise of “safety”.
Paul Bradshaw, a data journalist who worked on the BBC story, says:
… the algorithms for detecting identity attacks (racism/homophobia) are not reliable enough yet, so we did not use them. What we were able to look at was toxicity, defined by the algorithm as "likely to make someone leave a conversation”.
That makes the headline Scale of abuse of politicians on Twitter revealed completely unjustified; it’s selling an answer to a question that the BBC team didn’t have the tools, approach, or perspective to answer. But then a headline that said Scale of tweets that might upset politicians who are overly sensitive to criticism based on an extremely subjective and flawed measure of ‘toxicity’ revealed is a bit long.
And that wouldn’t allow figures like Ben Bradley or Nadine Dorries — who are essentially trolls with expense accounts — to present themselves as the victims. Nor could Jess Phillips who proudly boasted of telling Diane Abbott to “fuck off” — a claim Abbott later said was a lie — be the face of the whole exercise.
I was struggling to find an ending for this edition, so I looked to the list of toxic terms in the BBC spreadsheet for inspiration. Far from being “outstanding” as some sycophantic tweets by other journalists suggested, this research is stupid ✅ pathetic ✅ bullshit ✅. (This conclusion was rated 96.07% likely to be toxic)
Thanks for reading. I hope you didn’t find it too toxic.
There are 5,873 subscribers to this newsletter (up 46 since last time). 463 people are currently paid subscribers (up 8 since the last edition). If you’re among them, thank you.
If you’re not, please consider upgrading to help support this newsletter. If I can reach 10% paid subscribers I can write more, commission guest posts, and offer more extra editions and bonus content.
‘Bold‘ used there to mean “very fucking stupid” (Perspective API rated this footnote 79.99% likely to be toxic.