What causes your skin to be flawed

Instagram algorithm: if you want to be seen, you have to show your skin

Sarah is a grocery entrepreneur (name has been changed). Her company helps women feel good about their diet and promotes "intuitive eating". Like many small business owners, Sarah uses social networks to attract customers. Instagram, Europe's second largest social network after Facebook, is a marketing channel that she couldn't do without, says Sarah.

But on Instagram - which is heavily tailored to photos and videos - she has the feeling that her pictures do not reach the majority of her 53,000 followers, unless she poses in a swimsuit. In fact, she can be seen in a bikini on four of her seven most popular posts over the past few months. Ely Killeuse, a writer with 132,000 Instagram followers, agreed to have her name mentioned, saying that "almost all" of her most popular pictures show her in underwear or swimwear.

One explanation could be that Sarah and Ely's audience massively prefers pictures in bathing suits. Since 2016, however, Instagram has been arranging the images in a user's news feed in such a way that “those that are particularly important to the user appear at the top”. The fact that the other pictures of Sarah and Ely are less popular could also be due to the fact that they are shown less to their followers.

Which pictures are shown and which are not is not simply a matter of taste. Entrepreneurs who rely on Instagram to acquire customers have to adapt to the norms that the network favors in order to reach their followers. Even if these norms do not embody the values ​​on which they built their companies or which their clients and followers share.

2,400 photos analyzed

To understand which photos Instagram is prioritizing have that European Data Journalism Network and AlgorithmWatch Asked 26 volunteers to install a browser add-on and follow some professional content creators. We have selected 37 content creators from 12 countries (14 of them men) who use Instagram to promote brands or acquire new customers for their companies, primarily from the gastronomy, travel, fitness, fashion or cosmetics industries.

The add-on opens the Instagram homepage automatically at regular intervals and records which posts appear at the top of the test subjects' news feeds. This gives an overview of what the platform deems most relevant to each of them.

If Instagram doesn't interfere with the algorithm, the variety of posts in users' newsfeed should match the variety of posts made by the content creators they are following. And if Instagram personalizes the newsfeed of each user according to their preferences, the posts in the newsfeed would have to be distorted differently for each user. But this was not what we realized.

Between February and March, 1,737 posts with 2,400 photos uploaded by the content creators we observed were analyzed. Of these, 362 (or 21%) were identified as posts showing pictures of women in swimsuits or underwear or men with bare chests. In the newsfeeds of our test subjects, however, posts with such images made up 30% of all posts from the same accounts (some posts were shown more than once).

Posts with pictures of women in underwear or in bikinis had a 54% higher chance of appearing in our test subjects' news feed. Posts with pictures of men with naked torsos were 28% more likely. In contrast, posts that featured images of food or landscapes were 60% less likely to appear in the news feed.

These results, which can be viewed in detail here, meet the standards of statistical significance.

Waiting for a review

The bias in favor of nudity doesn't necessarily apply to all Instagram users. While it was clearly visible to most of the subjects, a small minority saw posts that were more likely to reflect the diversity of the content creators' content. It is likely that Instagram's algorithm generally favors nudity, but that personalization and other factors limit this effect for some users.

Our results do not provide a comprehensive review of Instagram's newsfeed algorithm. They only record what happened in the news feeds of our test subjects. (You can help us improve the results by installing the add-on; we will publish updates as soon as we have more data.) Without access to Facebook's internal data and production servers, it will never be possible to draw definitive conclusions pull.

Facebook did not answer our specific questions, but instead sent the following statements: “This investigation is flawed in many ways and shows a wrong understanding of how Instagram works. We'll organize the posts in your feed based on the content and accounts you've shown interest in, not arbitrary factors like the presence of swimwear. "

That said, we have reason to believe that our findings are representative of how Instagram operates.

Maximizing Loyalty

In a patent published in 2015, Facebook engineers - Instagram is operated by Facebook - explain how the news feed could control the selection and prioritization of images. According to the patent, when a user posts a picture, it is automatically analyzed immediately. Images are subject to a "retention metric" that is used to decide whether or not an image is shown in the user's news feed.

The retention metric is based in part on previous user behavior. When a user liked a particular brand and a photo shows a product from the same brand, the retention metric will increase. However, the metric can also be derived from the previous behavior all Users of the service will be charged. The patent explicitly states that gender, ethnicity, and the “degree of non-clothing” of people in a photo can be used to calculate an attachment metric.

While Instagram claims that the news feed is organized based on what is "most important" to a particular user, the patent says that the feed could very well be organized based on what Instagram thinks all Is important to users. Whether users see the images uploaded by the accounts they follow would then not only depend on their previous behavior, but also on what Instagram believes is important to other users of the platform.

Spurious correlations

Facebook automatically analyzes images with computer vision software before the algorithm decides which ones to display in the user's news feed. Such software draws automated conclusions from a training data set consisting of thousands of manually annotated images. Its limits could affect how Instagram prioritizes images in news feeds.

Computer scientists have known for years that such systems simulate and amplify the distortions of their training data set, which leads to spurious or deceptive correlations. For example, a program that has the task of identifying wolves and dogs on the basis of images from the Internet, which animals do not recognize animals like humans. Instead, it will identify any animal with snow in the background as a "wolf".

Computer vision training data is typically produced by underpaid workers who have an incentive to work fast and produce results that meet the expectations of their employers. This leads them to uncritically accepting categories that are offered to them and ignoring the subtleties a photo can contain, says Agathe Balayn, a PhD student at Delft University of Technology who researches distortions in automated systems.

The consequences can be far-reaching. In December, a Brazilian artist tried to promote one of his Instagram posts. The request was denied on the grounds that the post contained violence. It only showed a boy and Formula 1 driver Lewis Hamilton, both dark-skinned. In April, a yoga teacher was denied an advertisement for alleged profanity. The picture only showed the crane position. The teacher is American of Asian origin.

(In our experimental set-up we also used a computer vision system - Google Vision. Its results are debatable. For example, the attribute “beauty” was only used in women. It is very likely that the distortions are similar to those in Facebook's Computer Vision- System, if only because it was partly constructed by the same people.)

It's a fine line

Instagram's guidelines state that nudity is "not allowed" on the platform, but Instagram favors posts that show bare skin. The subtle distinction between what is encouraged and what is prohibited is determined by unverified and likely biased computer vision algorithms. Every time they upload an image, content creators have to walk a tightrope: showing enough to get their followers, but not so much that they get banned.

In 2019, a study by the US magazine showed Salty, which was carried out on 128 Instagram users, that improper removal of content is common. How often such incidents occur and whether women of color and women are disproportionately affected is impossible to say as long as Instagram's algorithms remain unchecked.

A study of 238 patents filed by Facebook containing the wording “Computer Vision” showed that out of 340 inventors named, only 27 were female. Environments dominated by men tend to produce outcomes that are detrimental to women. For example, seat belts in cars are only tested on male dummies, which leads to a higher injury rate in women. Our research shows that Facebook's algorithms could match this pattern.

Fear of the "shadowban"

Sarah and other entrepreneurs who rely on Instagram are afraid to speak to the press. Most professional Instagram content creators fear retaliation from Facebook, in the form of account deletion or "shadowbans" (in this case the posts of a user are not shown to any or only a few of their followers without them knowing about it). This would amount to a death sentence for their companies.

A young entrepreneur with around 70,000 followers, who said Instagram is “very important” for her business, asked AlgorithmWatch not to be named out of fear of a shadow ban. Ely Killeuse said that a second source of income was essential for her. Too much reliance on Instagram would mean losing her freedom and sanity.

The General Data Protection Regulation (GDPR) introduced in 2018 and the Platforms for Business (P2B) Regulation, which will come into force on July 12, 2020, already offer users and entrepreneurs many guarantees. In particular, the GDPR says that users have a “right to justify” automated decisions. The P2B guidelines are intended to force online services to disclose the “main parameters that determine the [algorithmic] ranking”.

These new measures are not intended to force the networks to reveal the internal mechanism of their algorithms, says Petra de Sutter, who chairs the Committee on Internal Market and Consumer Protection in the European Parliament. Preventing Instagram from sorting the users' newsfeeds would not be legally possible, she wrote in an email to AlgorithmWatch. Instead, the transparency that P2B will bring should be used to make informed policy decisions at a later date, she added. As for the fear of shadowbans, Ms. de Sutter thinks it is exaggerated. "A question has never resulted in retaliation," she wrote.

It may be different with P2B, but two years after the GDPR came into force, many experts complain that it was only implemented very poorly. One problem is that the Irish data protection authority, which is responsible for Facebook's Dublin-based European subsidiary, is dramatically understaffed and "doesn't seem to understand the GDPR," as an expert AlgorithmWatch explained. Another problem is the lack of supervision. No authority at EU or national level has the power or the resources it would take to scrutinize these huge platforms (Instagram included). Many GDPR regulations are therefore not implemented.

Possible discrimination

While our results show that male and female content creators are forced to show skin in order to reach their audiences, the effect on women could be greater and could be seen as discriminating against women entrepreneurs. But although the EU's Charter of Fundamental Rights prohibits discrimination on the basis of gender, there is no legal recourse for Instagram users to initiate proceedings. The specifics of companies in the field of social media are not taken into account in the legislation.

Miriam Kullmann, an assistant professor at the Vienna University of Economics and Business, wrote to AlgorithmWatch that European anti-discrimination legislation deals almost exclusively with employment relationships. Self-employed, like the content creators we've observed, aren't protected.

However, some groups do fight for the rights of independent creative people in social networks. IG Metall, Europe's largest trade union, supports a collective action by YouTubers who demand more fairness and transparency from Google, the owner of YouTube, when a video is taken off the market. However, they have no plans to expand their program to their colleagues on Instagram or other platforms.

Every third European

The number of European entrepreneurs creating content for Instagram is arguably several thousand. The impact of their posts is enormous, as they usually generate hundreds of thousands of followers. Facebook states that almost 140 million EU citizens (or one in three) used Instagram in April 2020.

Among the 18 to 24 year olds, Instagram penetration is 100% in every EU country. The quarantine regulations related to the Covid-19 pandemic have significantly increased the time spent on Instagram. Instagram traffic doubled within a week of the lockdown in Italy, Facebook told its investors.

A roll on its own

Almost a hundred years ago, the famous British writer Virginia Woolf wrote that women needed "a room to themselves" where their creativity could flourish. Submitting to the authority of other opinions, she wrote, was like inviting rottenness to spread to the heart of one's work.

On Instagram, there is no choice but to submit to the opinion of the authorities who constructed the newsfeed algorithm. Refusing to show skin automatically means limiting its range. Male and female entrepreneurs have to accept the rules made by Facebook engineers if they are to make a living.

Translation: Fyodor Shulgin (Voxeurop)