Promoting Fairness, Accountability, and Transparency Around Algorithmic Curation and Ranking Practices

As demonstrated in this report, the deployment of algorithmic curation and ranking practices by search engines and internet platforms have established their roles as gatekeepers of online information flows and online expression. Over the past few years, these practices have increasingly been dictated by the business models and services of these companies. Despite the growing prevalence of such algorithmic curation, internet platforms have demonstrated a fundamental lack of transparency and accountability around how such practices are implemented, and how these practices impact users, their worldviews, and publishers. Going forward, search engines and internet platforms, policymakers, and researchers should consider the following set of recommendations in order to promote greater fairness, accountability, and transparency around algorithmic decision-making in this space.

In particular, search engines and internet platforms that deploy algorithmic tools to curate and rank content in their search results or news feeds need to:

• Make a concerted and explicit effort to raise awareness around these practices, and provide adequate transparency around how they impact users’ experiences and free expression online. For the most part, there is little transparency around the algorithms that these platforms use, particularly when it comes to their operational mechanisms and how they translate inputs into outputs. Going forward, internet platforms should provide more effective and meaningful transparency around these algorithms’ decision-making practices.
  • Search engines and internet platforms should disclose and explain the various stages and procedures involved in collecting, curating, and ranking content in search engine results or in news feeds. This information should be updated whenever the platform decides to alter or refine its curation and ranking practices. This information should be available publicly and should be presented in a format that is easy to understand for members of the general public. This information should also be housed in a centralized and easy-to-access area, such as on a page dedicated to the topic on the platform’s website.
  • Search engines that employ a search quality rating process should publicly disclose what this procedure entails, including how any human raters are trained and evaluated, and how the search engine ensures that these individuals represent a diverse array of perspectives. In addition, if a search engine has developed guidelines for these raters, they should be publicly available, as this would enable users, researchers, and publishers to gain a better understanding of what values a search engine emphasizes in its curation and ranking process.
  • Where possible, platforms that engage in algorithmic content curation and ranking should provide additional resources that can help users, researchers, and publishers understand how a platform assesses and ranks content. A good example of this is the open-source code that Reddit provides for its news feed. However, because these resources often require a specific, technical expertise to understand and effectively use, they should not replace corporate public disclosures and explanations on algorithmic curation and ranking practices. Rather, they should supplement them.
  • Given that search engines are increasingly delisting search results due to legal requests, violations of their search engine guidelines, and requests under frameworks, such as the “right to be forgotten,” greater transparency and accountability needs to be provided around these procedures, and how they impact free expression. One method of doing this is by highlighting the scope and scale of these requests in a corporate transparency report. Search engines should also provide impacted website owners with notice of these removals and offer them the opportunity to appeal these decisions in a timely manner.
• Provide greater transparency to identify the different implicit and explicit preferences that are built into curation and ranking systems. Generally, this is difficult to do with algorithmic systems, as many of these biases are hidden and implicit. This is likely going to be the case for such curation and ranking systems as well. However, the fact that these systems are based on a hundreds of signals means that some of these preferences are also explicit.
  • Search engines and internet platforms should publicly disclose a comprehensive list of the signals their curation and ranking systems are based on. This information should also include how these signals interact and build off one another, and how they are weighted. If it is not possible to do this at a granular level, due to concerns over trade secrets and competition, for example, then these platforms should at least provide a comprehensive list of categories encompassing the different types of characteristics and signals that these systems consider, how these categories interact with one another, and how they are weighted. This will generate a greater understanding of which qualities are emphasized and valued more on a platform, and as a result which voices are amplified and silenced. This list of signals or categories of signals should be available in a single, public, central location that is easily accessible. This information should also be presented in a format that is easily comprehensible by a general audience.
  • When a search engine or internet platform updates the signals their ranking systems are based on, it can have a significant impact on the presentation and delivery of user and publisher content. As a result, platforms should strive to provide as many updates explaining these changes as possible. If the platform opts to publish categories of signals, rather than a comprehensive list of signals, these announcements should explain how these changes impact the overall category of signals. Currently, platforms make such announcements around changes such as those that aim to reduce the spread of misinformation or curb abuse. These announcements often outline the companies’ thinking and intention behind these algorithmic changes, and can be valuable for shining a light on the priorities of engineers who implemented these changes. This is because such announcements often underscore what values and assumptions the company made when approaching a content issue.
• Provide users with a robust set of controls which enable them to tailor their own search and news feed experiences. In particular, users should be able to:
  • Provide feedback on search results and posts in their news feeds. This should include the ability to hide, block, or filter out certain posts. These controls let users manage their own experiences and also provide the algorithms with valuable feedback on what content a user deems relevant and meaningful.
  • Determine whether and to what extent these systems are permitted to collect and use their personal data. Data such as a user’s location history, purchase history, or past browsing activity can be used to tailor a user’s online experience and inform targeted advertising. As a result, users should be able to control which of these data points are collected, particularly in instances where the data points are used for secondary purposes and are not integral for the company to provide the service. In addition, users should have strong controls related to data retention practices. This should include the ability for a user to clear their search history and delete the data a platform has collected on them.
  • Opt in to having their personal data used to refine and develop AI and machine-learning models. Currently, users are automatically opted in to having their data considered in datasets that train algorithmic models, with no method for opting out. This system raises significant privacy concerns, especially because these datasets can include highly personal information such as demographic characteristics, which can be used to target users in privacy-intrusive ways. In cases when highly personal data is being collected and used for potentially invasive purposes, such as facial recognition, users should always have to opt in to having their data used to train AI and machine-learning models. This is because as biometric information cannot be altered or changed like a credit card number can. Companies should also develop robust privacy policies around user data that is collected and used for training AI models.
  • Opt in to receiving algorithmically curated and ranked content. Ideally, users should only receive algorithmically curated and ranked content after affirmatively opting in—it should not be the default setting. However, if companies maintain that users should receive this content by default, users should be able to completely and easily opt out if they do not want to receive such content in their search results or news feeds. Additionally, this should be possible on search engines regardless of whether a user is browsing as a logged-in user, logged-out user, or private/anonymous browsing user.
• Enable publishers to understand and exercise some control over how their content is collected, curated, and ranked.
  • Both search engines and internet platforms should offer website publishers and content creators clear and detailed guidelines on how they can operate fairly and successfully on a given platform. For search engines, this means publicly sharing a detailed set of webmaster guidelines. For internet platforms with news feeds, this means publicly disclosing relevant rules and guidelines related to content creators.
  • Search engines should also let website publishers determine how their website is crawled. One way of doing this is using a sitemap. In addition, website publishers should be able to request recrawls if they feel their website was not adequately represented in a search engine’s index. Furthermore, website publishers should be able to opt out of crawling altogether if they so desire.
  • Website publishers who feel that their websites have been unfairly penalized in search engine results should be able to appeal these decisions and request a reconsideration. This appeals process should be timely and should provide adequate notice to the web page publisher of the outcome. Search engines that offer this appeals process should also educate website publishers so they know it exists and know how to use it.
  • Content creators who feel that their content has been unfairly penalized in news feeds should also be able to appeal these decisions. This appeals process should be timely. Internet platforms that offer this procedure should publicize information about this process and make it easy to understand.
• Internet platforms, policymakers, and researchers should collaborate on, promote, and fund further research on the impacts of algorithmic curation and content ranking. This is particularly important amid growing concerns around the impact of these algorithmic decision-making practices on democratic values, political polarization, and freedom of expression.
• Internet platforms, researchers, and civil society organizations should collaborate to develop a set of industry-wide best practices for transparency and accountability around algorithmic curation and ranking practices. These best practices should explicitly prioritize the public interest above corporate business models and concerns about trade secrets. This will help ensure that users are adequately educated and aware of these practices, have a range of meaningful controls at their disposal, and know how to use them. It will also promote greater accountability around these algorithmic decision-making practices and combat the need for knee-jerk legislation or regulation.