Hugh L. Stephens
Distinguished Fellow Asia Pacific Foundation of Canada Vice Chair Canadian Committee on Pacific Economic Cooperation (PECC)
This seems to be the prevailing view these days amongst the large digital social media and search platforms when the results of algorithmic selections they have programmed turn out to yield undesirable results. The poster child for this is Facebook as revelations from whistle blower Frances Haugen reveal how the company’s policies, and its algorithms, prioritize viewer engagement over safety and responsibility. Strident, opinionated, racist voices get people worked up. They generate traffic. Never mind that they are socially damaging, defamatory, incite violence, and are disconnected from factual reality. Never mind that some users of the service spread inflammatory misinformation about COVID vaccines or pursue other conspiracy theories. The algorithm boosts them and promotes more engagement. Never mind that many teenaged girls, addicted to Instagram, are worried or depressed about their body image. The algorithm that governs what they see makes the situation worse. As reported by the Guardian, “an algorithm tailors the photos and videos that a user sees, potentially creating a spiral of harmful content.” But it’s not the platform’s problem. It’s the algorithm, you see.
What exactly is an algorithm? We all have a general idea, I am sure. The Oxford Dictionary definition is “a process or set of rules to be followed in calculations or other problem-solving operations, especially by a computer”.
Algorithms are essential for the operation of many things, especially in the internet environment. For example, algorithms screen content for copyright violations and inappropriate sexual images. (They also get it wrong a certain proportion of the time, such as blocking Ruben’s nudes as well as contemporary explicit images). Algorithms have also been embraced by social media platforms, not only to block or filter but also to promote or recommend. When we talk about algorithms selecting what people will see, or be served, we are talking about a mechanism for sorting content based on relevancy to a user; how likely is it that they will want to see it? How does the algorithm do this? It works by “sniffing out” your preferences, based on likes, frequency of interaction, categorization of the type of content you have already seen, or seem to prefer, the content that you share with others, and so on. While this can be useful, it can also drive users into an echo-chamber where all they hear are voices that sound like their own. This is where the conspiracy theorists and others come in. Social media algorithms are driven by machine intelligence and machine learning, so the more you follow a predictable pattern, the more the algorithm will reinforce that behaviour.
Algorithms are normally proprietary, protected technology and intellectual property, with each platform jealously guarding its “secret sauce” that keeps consumers coming back for more. Given that they are created by “a real person”, even though they develop and grow under the impetus of machine learning, a chain of accountability can be established. It is the same principle as intellectual property, for example a copyrightable work being created by Artificial Intelligence (AI). AI can create musical, literary or artistic works, and these can be protected by copyright. But, as I wrote in a blog last year (AI, Ethics and Copyright),
“If only humans are capable of creating works that are subject to copyright protection (as the US Copyright Office has made clear in a recent interpretation), then how does one deal with works produced by a seemingly autonomous machine?”
This is not an easy question, but the answer increasingly seems to be that the copyright is conferred on the human being who created or directed the program, although it remains under study. Even in the case of machine and AI-generated works, there is always a human hand behind the creation. It may take some ingenuity at times to determine which human hand is predominant, but that is why we have courts. So, if AI can be traced back to a human creator (whoever designed the software program and applied it to the creation of the work), then so too the creators of algorithms must claim ownership, which also means bearing responsibility for what algorithms do.
Taking responsibility is the big issue of the day. Facebook maintains that it keeps its algorithms under constant review and refines them to eliminate undesired outcomes when necessary. But Haugen’s whistle blower testimony suggests that they are slow to act when problems surface and prefer to give priority to business outcomes (longer user engagement which drives more ad revenue) over social responsibility concerns. The criticism of Facebook covers how the company and its algorithms deal with everything from climate change deniers, political conspiracy theories, and sex trafficking to hate speech, young adults, teenage girls and health misinformation. An excellent summary can be found here, (thanks to The Markup.)
Inevitably this draws political scrutiny, in the US and elsewhere. Legislation can be tweaked to incentivize platforms to accept greater responsibility for their algorithms. For example, the US House of Representatives currently has two bills before it that deal with this issue. The first, (labelled the “Justice Against Malicious Algorithms” Act) would make internet platforms liable when they “knowingly or recklessly” use algorithms to recommend content that leads to physical or severe emotional harm by removing the “safe harbour” provisions against civil liability that platforms enjoy under Section 230 of the 1996 Communications Decency Act. Section 230 is a controversial provision that I have written about several times before (most recently, here) that allows digital platforms to evade civil legal responsibility for defamatory, misleading, obscene, racist or otherwise illegal or objectionable content posted by users on the questionable basis that the platforms are not “publishers” and therefore have no control over or responsibility for user-posted material distributed on their platforms. In other words, if this bill becomes law, there will be no platform immunity when the algorithm results in severe harm to the user.
A second bill, the “Protecting Americans Against Dangerous Algorithms” Act, would take this a step further by eliminating Section 230 protection in cases where;
“the interactive computer service used an algorithm, model, or other computational process to rank, order, promote, recommend, amplify, or similarly alter the delivery or display of information (including any text, image, audio, or video post, page, group, account, or affiliation) provided to a user of the service”,
“the information delivery or display is ranked, ordered, promoted, recommended, amplified, or similarly altered in a way that is obvious, understandable, and transparent to a reasonable user based only on the delivery or display of the information (without the need to reference the terms of service or any other agreement)…”
In plain language this means that where a user is manipulated by an algorithm or when an algorithm starts making content decisions unknown to the user, Section 230 liability protection no longer applies. This parallels the recommendation made by Facebook whistle-blower Frances Haugen with respect to Facebooks intentional algorithm-driven ranking decisions. However, who would make the decision as to whether a user is being manipulated is an unanswered question.
Another piece of legislation introduced earlier this summer in the Senate would remove liability protection from technology companies if their platforms spread health misinformation during a health crisis. This could occur when algorithms suggest and boost misinformation that seems to conform to what a user prefers, or that received “likes”, or was shared by a number of users. Small but organized sections of the population can boost particular sources of information, even if inaccurate or misleading, by creating lots of traffic, a tactic to which an algorithm will respond unless very carefully programmed.
Poorly designed algorithms can create their own distortions, as was shown by the case of “Carol’s Journey”. As documented by journalist Brandy Zarodzny, “Carol” was a fictitious user created by Facebook researchers back in the summer of 2019 to study the role of the platform’s algorithms in promoting misinformation to users. Carol’s profile indicated that she was “a politically conservative mother from Wilmington, North Carolina,…(with) an interest in politics, parenting and Christianity (who) followed a few of her favorite brands, including Fox News and then-President Donald Trump.”
How did Facebook’s recommendation algorithm treat this information? Within two days Facebook was recommending that she join groups associated with the conspiracy group QAnon. Within a week, her feed was replete with hate speech and disinformation, content that violated Facebook’s own rules. What did Facebook do with this information? Not much, according to Zarodzny’s report, although Facebook claimed that the research demonstrated how seriously it takes the issue of misinformation.
A problem with algorithms is that they can be manipulated by knowledgeable users to boost some content over others. It is common to see “search optimization” services offered to businesses to boost their rankings on search engines through various techniques that will appeal to and be picked up by the algorithm. Given that the software that sorts and ranks content can be tweaked by users, it follows that those who create and manage the algorithms must also be prepared to defend the algorithms they use. They need to take responsibility when their algorithms are either hijacked by users, or worse, misused by the platform itself to serve content to users that may be harmful to their well-being but which, like addicted gamblers, they continue to seek out.
Algorithms are here to stay. Like many things, they can be used for good or ill. In the hands of influential social media platforms that recommend content they are a powerful tool. Self-regulation does not appear to be working, so it is not surprising that various legal and legislative remedies are being proposed. Frankly, it is unlikely that any of the current bills in the US Congress will become law at the end of the day. They will likely meet the fate of much legislation that is introduced, debated, but not passed. Nonetheless, the fact that lawmakers are now seized of these issues is important and will hopefully bring more pressure to bear on the platforms to accept greater responsibility for their actions and the tools that they use to engage users. Hiding behind the algorithm won’t work. They need to own the problem and fix it. If stripping away some of their immunities brings about reform and greater accountability, I am all in favour.