Defining Meaningful Algorithmic Transparency Standards
Today's post is wonky, but algorithmic transparency laws have passed in the EU and are in progress in the US and UK, and we need to make laws meaningfully beneficial by being specific. And wonky.
The Digital Services act recently passed and will go into effect in 2024, with stakeholders now working to define what it specifically means to be in compliance. One section requires very large online platforms to be more transparent about their algorithms and the relevant Article (#29) reads as follows:
1.Very large online platforms that use recommender systems shall set out in their terms and conditions, in a clear, accessible and easily comprehensible manner, the main parameters used in their recommender systems, as well as any options for the recipients of the service to modify or influence those main parameters that they may have made available, including at least one option which is not based on profiling, within the meaning of Article 4 (4) of Regulation (EU) 2016/679.
2.Where several options are available pursuant to paragraph 1, very large online platforms shall provide an easily accessible functionality on their online interface allowing the recipient of the service to select and to modify at any time their preferred option for each of the recommender systems that determines the relative order of information presented to them.
Effectively, they are mandating that platforms deemed “very large” (45 million monthly active users in the EU), including TikTok and Facebook, would have to reveal the “main parameters” used in their recommender systems. How would we define this, such that compliance is a matter of following a standard vs. platforms deciding for themselves how to define “main parameters”?
Dall-E image generated of “a computer made of glass as digital art”
A goal of the DSA is to “create a safer digital space”, and so we need to provide enough information to understand why any given piece of “unsafe” content was recommended or amplified. In recent years, I’ve spent much of my time attempting to diagnose and improve recommendation algorithms, including cases where Facebook’s newsfeed has recommended content that reduces physical safety or incentivizes division. An engineer confronted with a vague statement like “Facebook is spreading hate”, will not know what to do. But if you give them a specific piece of content, the algorithmic formula used to recommend it, and the model predictions that fit into that algorithm, they can readily figure out what occurred and recommend steps to make changes. If we want to have that level of control over our recommendation systems, we need to have a similarly detailed view.
Some might ask why we wouldn’t just remove such content. I wrote a separate piece on the limits of content moderation, but I’ll provide a specific example here that might help. One of the worst episodes of violence linked to Facebook occurred in Ethiopia, where an activist expressed fear for his life at the hands of the government, leading to deadly violence between his supporters, who tended to be of one ethnic group, and members of other ethnic groups. Most human rights defenders would consider an activist's declaration of fear of government persecution to be especially protected speech, rather than hate speech, yet articles tied dozens of deaths to this post. There was certainly a need to remove hate speech that led up to this event, but in a country where ethnicity and politics are so tightly connected, there is no way to enforce against just the violence inducing speech while also respecting freedom of expression as the lines will never be clear. At some point, one has to address the algorithmic incentives toward divisive content and understand how those algorithms might be recommending and incentivizing dangerous authors and content, rather than the impossible task of attempting to enforce judgment on each piece of content.
If we agree that algorithms matter and so therefore we need to understand and fix them, how do we begin to build the transparency necessary to do that?
First, you need to know how the set of things that could be recommended is created. This could be based on all movies from genres you like or all content from people you follow. It may or may not remove things that are deemed potentially harmful or risky, based on some specific criteria (e.g. 75% chance of being “hate speech”). This allows us to understand who could possibly have been exposed to any piece of content, but we still won’t know how this piece of content fared against other content in this bucket, since users often see only a fraction of the things that are possible to show to them.
Next, this set of things needs to be ranked, so you need to understand specifically what “terms” any ranking algorithm is optimizing for. This cannot be something vague like “what people want” or find “meaningful”, but rather something you could actually put into code. You may use a predictive model score that is hard to understand, but that model score is predicting something specific. It may be a weighted sum of different user actions (e.g. the meaningful social interaction metric) or there may be many models that each predict individual events. But you need to know what all those terms in the algorithmic equation are and how specifically they are combined. The specificity should look something like an algebra equation and it may be long and complex, especially for large organizations that may have lots of variables to balance, but it is ultimately comprehensible to anyone who has taken high school algebra. Facebook gives some of the input terms for newsfeed in this post (e.g. “likely to comment”), but it doesn’t tell you how they are combined, and we don’t know if this is the full list of terms or a partial list. If we want to understand how bad content spreads, we need to know both the weights and the full list of terms. If comments are weighted 16 times more than likes, for example, that certainly could be an issue causing content with arguments in comments to get recommended beyond the desire of users.
Hypothetical example for newsfeed using some of the input terms mentioned: Ranking Score = (5*p(like))+(80*p(comment))+(2*p(generates_replies))+(1*p(survey_post_worth_your_time)))*IF(bad_for_integrity then .2 else 1))
The above may seem complex, but not so much that motivated observers cannot understand it and advocate for their communities. However, it won’t help you diagnose why something harmful is getting recommended without a couple more pieces of information.
We also need to know the distribution and top predictors for any given term. If a user commenting is 40 times more important than replies generated, but the range of a comment term is from zero to 1 and tends to be under .1, while the term for the number of replies generated could be infinite, then you can imagine a case where most of the predictive power for viral content is being driven primarily by an expectation of future replies. This is a hypothetical example, but the point is that you cannot know why something is recommended without knowing the distribution of input terms as extreme values can dominate.
At times, these inputs will be predictive models, as, for example, you can’t know in advance whether a post is or is not “worth your time”, since most users will not have answered a survey. One could imagine a system might use a variable like “poster is similar to user” to make this prediction, which may have diversity and fairness implications for posters who are not similar to that user. Without knowing the top features of predictive models used in terms within an algorithm, you can’t know whether there may be problematic input terms that are obscured by less problematic sounding models.
This will help you understand generally how the system works and why content is generally distributed in a certain way, but you still need one more input if you are trying to diagnose specific safety issues in a recommendation system. Specifically, you need to know the specific term values for any piece of content of interest. Consider the formula above again. Generally, it may seem reasonable if we look at the terms and the general distribution of inputs. But one particularly problematic piece of content could generate an incredibly high amount of comment replies. In fact, those scores could be so high that any demotion based on integrity signals may not make a real difference as 20% of a really high number can still be a really high number. Given the possible skew that outlier values can introduce, knowing the specific term values for a piece of content can help a person diagnose exactly why they are seeing any particular video or story, sometimes in contrast to the expectations of the algorithm designers.
The DSA is just one of many laws purporting to reduce risk and increase accountability by providing transparency of algorithms. As you can see in the article language above, the language is broad enough that implementation matters and without demanding specifics, companies could provide vague information that won’t actually allow the outside world to help find issues or enforce accountability. This also wouldn’t provide the true ability to “modify or influence” “main parameters” since one cannot meaningfully understand one’s influence without this level of information. Research shows that consumers value transparency, and so even if most users won’t access such information directly, they are likely to value it being visible and accessible.
Healthy algorithmic amplification is clearly an unsolved problem and companies need help with it as they can’t possibly find all the contextually dependent issues that users will have with respect to recommendations. Companies are under revenue pressure and are risk averse so are unlikely to take on the expense and potential embarrassment of doing this willingly. Any suggestion that this is impossible is likely a deflection. We should create a level playing field for everyone creating tomorrow’s online platforms by creating a robust and detailed standard for what meaningful algorithmic transparency means for increasingly influential recommendation systems. At a minimum, this should include:
how the set of things that could be recommended is created
specifically what “terms” any ranking algorithm is optimizing for
the distribution and top features for any given term
the specific term values for any piece of content of interest
Users and researchers value transparency and there is a psychological benefit to knowing more about the process. Yet, as we design more laws to mandate transparency, we may want to ask ourselves what our goal truly is. Do we want to feel better about the process or do we want to be able to specifically diagnose and change issues we see? If the goal is the later, then you need to know the specifics of how recommendation systems are ranked including the 4 specific answers above. To date, no legislation I have seen has specified the level of detail that would allow this, but there is room for regulators to define standards along these lines. Doing so would allow us to more effectively scrutinize the algorithms that impact our everyday experiences.
Do you have any research on user perceptions of algorithms that might inform efforts around algorithmic transparency? If so, please do send them my way (email@example.com) and we will consider them for inclusion in future newsletter posts.
Below are other articles we are reading compiled by Joo-Wha Hong, Human-AI Interaction Researcher at the USC Marshall School.
Burgoon et al. | International Journal of Human-Computer Studies | 2016
Life is full of surprises, including communicating with chatbots. Even though most chatbots are preprogrammed, it does not mean all their behaviors are predictable. In interpersonal communication between humans, the influence of unexpected behaviors has been explained by the expectancy-violation theory (EVT). Judee Burgoon, who first claimed EVT, tested whether her theory could also explain communication with machine agents.
Agudo et al. | Frontiers in Psychology | 2022
Some people are reluctant to get exposed to any information about art pieces before enjoying them. Those people should be now careful about the information about the artists because there are more artworks created by AI these days. The question is how much people are sensitive about AI being an art creator and whether it has an influence on the evaluation of AI-generated art, which this research aims to answer.
Jennifer Huh et al. | Journal of Global Fashion Marketing | 2022
Human-AI interaction has been applied to shopping, which facilitates consumer decision-making. However, if AI makes decisions, what is the joy of shopping? Huh and her colleagues examined how user autonomy, perceived human-likeness, trust, and purchase intention interact with each other when customers make shopping decisions with AI.
Caren Chesler | The Washington Post | November 12, 2022
We tend to think of death as what separates us from deceased ones and have wanted to cross the border to meet loved ones once again. While the current technologies cannot resurrect those who passed away, they may now at least give us a feeling of reunion. Check out how the recent developments of AI and the metaverse make it possible.
Cody Mello-Klein | News @ Northeastern | November 3, 2022
The metaverse has been a massive trend in both academia and in industry. However, the market value of Meta, one of the biggest companies leading the metaverse, has recently tanked. While Zuckerburg attempted to frame it as a minor issue that happens when a major investment is made, there are still conflicting views. Is it merely an instant problem a single company is facing or a red alert that the whole metaverse industry and researchers should be aware of?
Jaclyn Peiser | The Washington Post | November 13, 2022
The post-pandemic holiday season is now coming, so people are now busy looking for gifts for loved ones. However, stores would still be less crowded than expected because of social media. The article reports that Gen Z and millennial shoppers prefer using social media to buy gifts for many reasons. Check out the article and find out why buying gifts at a store is old-fashioned.