MIT Technology Review
In her testimony, Haugen repeatedly emphasized that these phenomena are far worse in regions that do not speak English due to the uneven coverage of different languages by Facebook.
“In the case of Ethiopia, there are 100 million people and six languages. Facebook only supports two of these integrity systems languages, ”she said. “This strategy of focusing on language-specific, content-specific systems for AI to save us is doomed to failure.”
She continued, “So investing in non-content-based ways to slow the platform down not only protects our freedom of expression, but also protects people’s lives.”
I explain this in another article earlier this year on the limitations of large language models, or LLMs:
Although LLMs have these linguistic deficiencies, Facebook relies heavily on them to automate its content moderation worldwide. As the war in Tigray[, Ethiopia] first erupted in November, [AI ethics researcher Timnit] Gebru watched the platform float to get a grip on the flood of misinformation. This is emblematic of a persistent pattern that researchers have observed in the moderation of content. Communities that speak languages that are not prioritized by Silicon Valley suffer from the most hostile digital environments.
Gebru noted that the damage doesn’t end there either. If fake news, hate speech, and even death threats are not blocked, they will be used as training data to help build the next generation of LLMs. And these models, which parrot what they were trained on, in the end choke up these toxic language patterns on the Internet again.
How is Facebook’s content ranking related to teenage mental health?
One of the most shocking revelations from the journal’s Facebook files was Instagram’s internal investigation, which found the platform is worsening the mental health of teenage girls. “32 percent of teenage girls said they felt worse on Instagram when they felt bad,” researchers wrote in a March 2020 slide presentation.
Haugen also connects this phenomenon with engagement-based ranking systems, of which she said in front of the Senate today that “teenagers are more exposed to anorexic content”.
If Instagram is such a positive force, have we seen a golden age of teen mental health in the past 10 years? No, we’ve seen rising suicide and depression rates among teenagers, ”she continued. “There is a wide range of research to support the idea that social media use increases the risk of these mental health problems.”
In my own reporting, I heard from a former AI researcher who saw this effect carried over to Facebook.
The research team found that users who had a tendency to post or engage with melancholy content – a possible sign of depression – could easily lead to consuming increasingly negative material that risked further deterioration in their mental health.
But as with Haugen, the researcher found that leadership wasn’t interested in making fundamental algorithmic changes.
The team suggested tweaking the content ranking models for these users so that they no longer maximize engagement just to show them less depressing stuff. “The question for the leadership was, should we optimize engagement when you discover someone is in a vulnerable state of mind?” He recalls.
But anything that reduced engagement, even for reasons not to worsen someone’s depression, resulted in much hustle and bustle among the leadership. With their performance reviews and salaries tied to successful project completion, employees quickly learned to drop those who received a pushback and move on to those dictated from top to bottom.
The former employee no longer allows his daughter to use Facebook.
How do we fix this?
Haugen is against the dissolution of Facebook or the repeal of Section 230 of the US Communications Decency Act, which protects technology platforms from taking responsibility for the content they distribute.
Instead, she recommends making a more targeted algorithmic ranking exception in section 230, which she believes would “get rid of engagement-based ranking.” She also advocates a return to Facebook’s chronological news feed.