Cybersecurity & Tech

Exploring Tradeoffs in Ranking and Recommendation Algorithms

Liz Arcamona, Louisa Bartolo, Yvonne Lee, Sarah Shirazyan
Friday, September 29, 2023, 10:15 AM
How do social media companies rank and recommend content? What aims drive these systems?
"Man scrolling photos on smartphone while working on laptop." (Callum Hilton,; Public Domain,

Published by The Lawfare Institute
in Cooperation With

 Stakeholders ranging from civil society groups to scholars have raised a variety of concerns related to ranking and recommender systems on social media, including the amplification of potentially harmful content, limits on user agency due to digital “nudging,” the reductive nature of machine learning-based user profiling, and the erosion of user privacy through personalized ranking and recommendation algorithms.

At the heart of these concerns often lies an assumption that social media companies are motivated to optimize ranking algorithms to promote user engagement to improve their bottom line—even at the possible detriment of user well-being, platform integrity, or societal harmony.

However, our extensive collective experience, spanning two decades of working in and studying some of the world's largest platforms, challenges such a view of companies’ motivations and their resulting decisions about how to employ ranking and recommender systems. Focusing solely on engagement does not capture the complexity of these systems. Further, companies who want to ensure the longevity of their platforms must think more broadly about how to maintain platform health, including creating an environment that fosters healthy and enjoyable connections and access to information.

In fact, recommender systems rely on a complex array of signals—like metadata relevance, content recency and freshness, and user survey responses—in addition to engagement metrics. Instagram, where several of us work on developing the platform’s policies, is one of many platforms that has introduced recommendation-specific content policies making various types of content that could lead to poor user experiences (e.g. sensationalist or misleading content) ineligible for recommendation, regardless of how much engagement they attract.

Product and usage trends also affect how companies employ ranking and recommendation systems. For example, platforms like Instagram were built around the concept of connecting users based on their relationships, affiliations, and social connections—the “social graph.” However, as the volume of short-form videos in particular has exploded over the past few years, some platforms have optimized for content discovery and topical relevance over the strength of social ties. The resulting “content graph” places a premium on understanding and analyzing the content itself—its quality, context, and relevance—enabling platforms to deliver personalized content recommendations and fostering engagement based on user interests with respect to certain content, rather than solely on social connections. Content-based ranking models do not prioritize engagement, making the content itself—rather than users’ interactions with it—the driving factor of the ranking and recommendation process. As content ranking methods continue to evolve, traditional paradigms of content ranking need to be reexamined. 

Ultimately, content ranking is a complex interplay of various models, meticulously paired together to achieve a platform's goal. That goal can vary depending on the platform, service, or even feature within a platform. Our aim in writing this piece is twofold: to explain the key ranking models and their respective advantages and limitations, and to advocate for a more nuanced public discourse that recognizes the multifaceted nature of content ranking systems. 

Understanding Content Ranking

The issues surrounding social media platforms are not new. For years, digital platforms like news websites and shopping sites have grappled with devising algorithms to curate content in ways that resonate with individual users. We first provide a foundational view of what ranking and recommendation systems are, the key stakeholders who should be engaged in this debate, and the importance of merging the social and technical components of this discussion.

Defining Algorithmic Ranking and Recommender Systems

For the purpose of this piece, we draw a distinction between ranking and recommender systems, even though we recognize that such a dichotomy may not exist in people’s mental models.

Ranking refers to selecting, filtering, scoring, and then ordering items, which are then presented to users as a list. Ranking can be based on different factors, including recency, popularity, or relevance to the user. Ranking can follow from a specific user request, such as when a user searches for a specific item in a search bar, or without one.

This is where recommender systems come in: systems that suggest content to people they haven’t explicitly asked for. Recommender systems learn what users like, based on users’ explicit ratings or engagement, and couple that with metadata on the item itself to recommend similar items in the future. These systems also learn from patterns in user preferences and may optimize for the diversity or serendipity of content, so that users are not repeatedly presented with the same type of content. 

Sometimes, there is overlap between ranking and recommender systems: when a user searches for something and there’s no perfect match for their query, the results could also be thought of as recommendations. 

Both systems increasingly rely on sophisticated deep learning models that respond to countless regularly shifting signals. Thus, the exact outputs of algorithmic ranking and recommender systems are not always predictable, even to the developers building these systems.

Common Points of Concern with Algorithmic Ranking

In the past few years, scholars and civil society groups have drawn attention to a range of concerns related to ranking systems on digital platforms hosting user-generated content. Notably, we’ve heard these concerns generally across various types of ranking models, without a clear distinction on which risks are most salient for each ranking model or how to leverage model options to balance these risks. These include concerns about:

Key Stakeholders in Algorithmic Ranking

Algorithmic ranking and recommender systems are “multistakeholder” systems that involve different parties with a range of interests bound up in processes of algorithmic sorting and visibility. In the case of Instagram, for instance, a number of key stakeholders and their respective interests are at play, such as:

  • Creators: Creators have expressive and financial interests in being ranked highly and recommended. Creators range from influencers and entrepreneurs building brands to activists aiming to raise awareness or promote political causes.
  • Governments and Regulators: Emerging regulation, like the EU Digital Services Act and the UK Online Safety Bill, includes specific provisions around algorithmic recommender systems, which the largest platforms are required to comply with.
  • End-users: Users may have interests in seeing content they enjoy and that provides accurate information, and in having greater curatorial agency.
  • Advertisers: Advertisers often focus on brand safety and ensuring advertisements do not show up next to harmful content. They often work through initiatives like GARM, the Global Alliance for Responsible Media, to advocate for consistent guardrails and systems that keep brands and people safe online.
  • Instagram as a company: Instagram has an incentive to provide value to users, to enforce its Community Guidelines, and increasingly, to comply with external regulation around ranking and recommendation.

Ranking and Recommender Systems as “Socio-technical” Systems

In academic circles, ranking and recommender systems are referred to as ”socio-technical” systems—that is, they’re technical (complex machine learning systems) but they’re also inevitably social. The algorithmic code they rely on is produced by a team of people. Once that code is deployed, individual user behavior constantly feeds back into the system, which “learns” from that behavior. In fact, there is a growing body of scholarship forming under the banner of “critical algorithm studies,” that is concerned with investigating “algorithms as social concerns”: shaped by, and shaping, the societies in which they are embedded.

People can and do influence algorithmic recommendations simply by clicking on and engaging with content that they are drawn to. Research shows that as users become more aware of algorithmic systems and form various assumptions about how those systems work, they also consciously adapt their behavior to see more of what they want and less of what they don’t. For example, users can strategically click on certain types of content to “teach” personalized recommender systems what they wish to see—indeed, many platform features invite users to do so.

Content creators inform how ranking algorithms work and adopt practices to optimize the visibility of their content. This means that algorithmic ranking and recommendation are powerful not simply because they influence what users see, but because they can also influence what type of content gets created in the first place.

Content Ranking Models and Their Trade-offs

The largest social media platforms like Instagram, Facebook, TikTok, and YouTube offer a vast amount of content to their users. To view all the videos presently accessible on YouTube, one would need to dedicate over 17,000 years of consecutive video watching. Given the near limitless scale of content online, platforms play an important role in providing selection services via ranking and recommender systems to make it easier for people to find content that will be most valuable to them. The introduction of the ranked feed on Instagram was a result of this exact challenge—people were missing 70 percent of the content in their feeds, including almost half of the posts from their close connections.

While it is clear that ranking content provides value to users, there are several different methods. We outline the diverse strategies that platforms can employ, the benefits and drawbacks of each method, and how platforms can consider balancing such trade-offs.

1. Observe Engagement

As early as 1993, Brenda Laurel, in Computersas Theatre, explained how engagement is considered “a desirable—even essential—human response to computer-mediated activities.” Engagement has long-been recognized as an important metric for analyzing human-computer interaction, designing effective websites, and developing good recommender systems. When people choose to engage with content, that’s a strong indication they are getting genuine value out of it. News organizations, for example, customarily use this same principle when measuring the success of their publications via views or clicks as metrics. 

User engagement is among the most direct forms of input platforms can receive from people about content they want to see. It helps users find valuable, relevant information—drawing from people’s existing network of connections, interactions, and behaviors.

It is important to accurately determine why a piece of content may be relevant to a particular person. For example, one person may be interested in an athlete because they love the sport, while another may be interested in the athlete’s personal life; the goal is to show content about the athlete that’s relevant to why someone is interested in them.

Still, an over-reliance on engagement-based signals to guide ranking has aspects that many find concerning, including the following:

  • Entrenching popularity. Feedback loops in recommendations mean that popular creators tend to get more exposure, which leads them to get more engagement, generating the rich-get-richer effect referenced above. This can be mitigated with efforts to optimize for other factors, like diversity, novelty, and fairness, although of course those factors bring their own challenges.
  • Prioritizing users’ short-term interests as opposed to users' “ideal” preferences and long-term satisfaction. Efforts to solicit thoughtful feedback from users on their preferences to optimize for their long-term satisfaction may balance this out.
  • Leading people to consume and produce sensationalist or poor quality content. While trust and safety interventions can ensure that users don’t see harmful content, this also opens the door for broader normative debates about what “good” curation should look like.

2. Ask People via Surveys

In addition to examining user behaviors (e.g., liking posts), which isn’t always a perfect indicator of content that someone finds valuable or relevant, one now-common method for ranking content across social media platforms includes surveying people to better understand their perspectives and integrating such feedback into how content is ranked. For example, surveys that ask people “is this post worth your time?” help inform Facebook’s algorithms. YouTube measures “valued watchtime” through user surveys that ask people to rate content from one to five stars. Academics have also suggested listening to what users say in those surveys—not just their behavior—to enable key developments for recommender systems. 

Such surveys can be additive to other ranking methods. For example, social media platforms might predict that certain content is of interest to someone because they engaged with it, but users may respond in a survey that such content was not worth their time. Similarly, users may engage with a piece of content not because they found it valuable, but because they wanted to signify to someone that they had seen the post. Thus, user surveys can help elucidate the content that is most valuable to the user. 

Academics have pointed out that the effectiveness of surveys depends on asking the right questions. Many also argue that it is critical to ensure that the survey is up to date, as people’s preferences and interests change and develop over time, and platforms need to find creative ways to make surveys easy and appealing for users to engage with. For example, in addition to being asked for explicit information about their preferences, users could also be invited to adjust the parameters of their recommendations, an approach that blends surveying with user-level controls, something we come back to in the final section of this piece.

To scale survey results, platforms need to overcome challenges of extrapolation. The effective use of surveys depends on identifying people who are similar to the survey respondents. However, there are many reasons why people may be similar, such as sharing an interest or a geographic location. Accurately determining whether those similarities are relevant to defining the content people want to see online is the key to successfully scaling survey-based ranking.

3. Analyze Content (topic, sentiment, etc.)

Experts have unpacked the industry-wide trend of social media platforms shifting from a social graph (i.e., ranking based on who users are connected to) to a content-based graph (i.e., ranking based on specific topics). Content-based ranking can be tied to a number of factors, including the content’s format (e.g., whether it is a photo or video), topic (e.g., whether it’s about food, sports or fashion), or the sentiment associated with it (e.g., whether the content is happy, funny, or otherwise entertaining).

For example, researchers have encouraged platforms to create ranking models based on a predetermined set of human values. Others have suggested that platforms rank content based on the perceived “quality” of the post, pointing to platforms like YouTube that look for different indicators of content quality depending on the topic. Google's Search Quality Rater Guidelines also reveal this topic-specific approach. Facebook has continued to test and refine its approach to political content. Research has also suggested that the subject of the content—whether it is entertainment or politics, for example—being ranked matters, as people tend to have different expectations of good curation in different subject areas, and norms around curation differ based on the subject.

Because content-based ranking optimizes for interests and topics, rather than engagement, some point out that more creators (including up-and-coming creators) could have a chance for distribution to a broader audience. For example, you might be interested in exploring food ideas for your picky toddler, but your social graph may not include parents with similar interests. Proponents of content-based ranking models argue that topic-based signals can help expand the pool of content on a specific subject of interest to a user, irrespective of their social connections.

However, others suggest that excessive focus on content analysis may not provide the best ranking and recommendation outcomes in the long-term, because it can result in over reliance on a particular topic. For example, new parents could continually be served baby content, even if they are no longer interested in the topic. To counteract this effect, platforms would need an algorithm that is both able to adapt nearly instantaneously while not being overly sensitive to temporary changes in user behavior. With respect to sentiment-based ranking systems, people have different perceptions of what’s entertaining or funny, making it difficult to identify accurately what content corresponds to particular sentiments. Further, people may raise concerns that relying on sentiment-based signals allows platforms to manipulate users’ emotions, triggering questions about the influence of social media companies on users’ personal lives.

4. Empower Human Curators

Traditional media has long relied upon human creators. Journalists and editors, for example, select topics that they consider of interest to the reader and to society more broadly. These decisions are ultimately judgment calls made by trained professionals about which stories to publish and put in front of the public eye. 

Similarly, on social media, some argue that human curators may help improve people’s experience by identifying trends or topics with growing engagement, and recommending related content. Some journalists have suggested that reintegrating human curation could help create an online world that values quality over quantity, such as by disincentivizing “clickbait” content. Human curation is a promising path to show people content they may be interested in, although questions around scaling human curation continue to linger. 

Even if there was a path to scaling human curation, many tough questions surrounding what defines “good” ranking and recommendation remain. People may disagree on what counts as expertise and proxies for quality content.

On many social media platforms, traditional media intermingles with user-generated content. As a result, an important and ongoing debate centers on whether traditional media should receive special privileges, including higher ranking, based on the assumption that they provide more reliable information, and are already subject to regulation. 

Some platforms, like YouTube, raise content from official news media channels and health authorities in recommendations once users have run news or health related search queries, for example. While this is logical, the practice raises issues in cases where news channels spread misinformation. In the case of various contested sociocultural issues, people disagree about which voices should be raised and how platforms should manage “information contests.” Human curation would not solve these disagreements, because they are intensely human and political disagreements.

Further, a growing body of work on news recommender systems demonstrates that the process of agreeing on norms for human curators to use is highly complex (not to mention translating such norms into computer code).

User Agency in Ranking

Finally, platforms also seek to further empower users by giving them additional methods to control what they see. These tools can offer more granular controls, by offering users the option to indicate that they are “not interested” in a piece of content, or by including an option to view content reverse chronologically (from newest to oldest).

Providing such controls helps people make the most out of their time on social media. Ranked feeds prioritize posts that platforms predict that users want to see, but users know best what they want to see.

Experts also point out the limitations of user controls, as they exhibit an individualistic approach to ranking—which may not address the societal concerns associated with algorithmic ranking. For example, research has suggested that some customization tools, like allowing people to opt for reverse chronological feeds, can actually lead users to see more content from untrustworthy sources. In general, when it comes to  user customization options (like enabling users to curate their own feeds), there is a need to balance an interest in giving users more influence over what they see and mitigating societal concerns around growing fragmentation of the online information space. Thus, experts suggest that it may be worthwhile to explore other methods to empower user agency, outside of individualized customization tools. For example, recent work has experimented with encouraging users to fact-check content and downvote false content to influence a piece of content’s ranking.

Goals of Ranking on Instagram

In light of the benefits and drawbacks associated with different ranking models, our aim is to provide insight into how Instagram weighs the trade-offs in ranking.

Ultimately, Instagram’s goal when ranking content is not focused on keeping users on the platform for the longest time possible. Rather, the goal is to create user value, which manifests differently among users and even across distinct parts of the platform for the same user. To understand user value, Instagram considers a variety of factors—including findings from research and surveys, stakeholder engagement with the top experts in the field, and user activity on the platform.

Ranking differs across various parts of the platform to reflect how people use Instagram. For example, people tend to look for their closest friends in Stories, use Explore to discover new content and creators, and use Reels to be entertained.

Thus, in Stories, important signals for ranking include viewing history and engagement history, such as sending a like or a direct message—engagement that may be more common between close friends. These signals help us prioritize Stories from accounts Instagram thinks a user doesn't want to miss. Because users generally like to see, and tend to engage more with, content from their closest friends in Stories, content from brands or creators (who are not their friends) that they follow may appear lower in their Stories tray.

This is in contrast to Explore, where people are generally trying to discover new things. One of the signals that is more important in Explore than in Stories is how popular the post is on Instagram, including how many and how quickly other users are liking, commenting on, sharing, or saving a post. It's therefore less likely that a user will see content from a close friend in Explore as compared to Stories.

Finally, on Reels, a key goal is to show users content they find entertaining. Thus, Reels ranking relies in part on surveys, which help us get better at identifying Reels that entertain users. Survey data is valuable for ranking on Reels, because unlike some other parts of the app like Stories, the majority of what users see on Reels is from accounts they don't follow. Survey data, combined with other signals like content-based signals (e.g. whether there are audio tracks or visuals in a video) and engagement-based signals (e.g. what a person has liked or commented on) help show users the content most entertaining on Reels.

Balancing the Trade-offs in Ranking

Instagram often balances the trade-offs of engagement-based ranking by pairing it with other ranking models. For example, Instagram supplements engagement-based ranking with surveys to combat clickbait. Some users might engage with clickbait content, indicating an interest in seeing more of it. However, surveys often reveal that despite high engagement with clickbait, users don't find such content worthwhile. Hence, combining these two ranking models allows Instagram to better understand that clickbait content may not hold value for its users.

Additionally, Instagram reduces the visibility of content that our systems have assessed as likely in violation of our policies, where our content reviewers have not yet made an assessment. Instagram's success hinges on running a platform on which users have positive experiences and advertisers choose to run ads. It is therefore in the platform’s interest to reduce the visibility of content that might trigger bad experiences or discourage advertisers.

Instagram relies on several key principles to help assess the trade-offs between different ranking models. Such principles include, but are not limited to:

  • creating a safe environment online for a global user base;
  • maintaining a space where people can share their creativity, build small businesses, and grow their audience;
  • developing a scalable approach to enable various communities to find their home on Instagram; and
  • helping people understand the key mechanics of our ranking systems.

For example, Instagram demotes potentially problematic content to support a safe and positive user experience. However, to strike a balance between the principles of fostering a safe environment while preserving space for user expression, Instagram seeks to apply demotions only in a necessary and proportionate manner. 

Maintaining Integrity on Instagram

Instagram relies on a set of rules to ensure that the platform is a safe place online. All content that appears on the platform, whether it is ranked or unranked, is subject to the Community Guidelines. The Community Guidelines outline what is not allowed on Instagram and the goals that drive those prohibitions.

In addition to the Community Guidelines, distinct rules apply to ranked and recommended content, intended to address some of the potential risks of ranking and recommendation systems. Instagram's approach to content can be viewed through two distinct lenses: connected ranked content and recommended content.

Connected ranked content consists of posts from accounts that users have chosen to follow. The position of such content may be lowered in one’s Feed or Stories tray, such as content that Instagram predicts but have not yet confirmed may go against the Community Guidelines.

In contrast, recommended content, which users may not encounter without Instagram's proactive recommendation, is subject to even more stringent criteria. The Recommendation Guidelines outline the types of potentially low-quality, objectionable, or sensitive content Instagram aims to avoid recommending, even if engagement-based models would otherwise optimize showing that content (if users showed interest in it).

User Agency on Instagram

Instagram’s ranking system aims to show users content they are most likely to enjoy, but it is not perfect. Users may still find some content not valuable, despite a platform’s best efforts to balance the trade-offs of different ranking models. Therefore, providing controls becomes an important tool for giving users more agency, allowing them to see more of what they want online.

Instagram offers alternative ways for people to choose what posts they see in their feed, including a Favorites feed—which shows the latest posts from a list of specific accounts that a person has chosen—and Following, which shows posts from accounts that a user follows, arranged in reverse chronological order. 

However, though some users may prefer consuming content in reverse chronological order, it's worth noting that several experts have explained risks associated with making this the default experience. A default reverse chronological feed can make it more difficult for platforms to reduce the visibility of potentially harmful or borderline content, as all content is shown chronologically. Furthermore, it can inadvertently encourage excessive posting and spamming as the most recent content is what users will see, ultimately leading to a bad user experience.

Instagram also offers more granular controls to users, such as:

  • Sensitive Content Controls, which allows people over 18 to choose to see more or less potentially sensitive content in their recommendations.
  • Not Interested, which allows people to hide a suggested post that they are not interested in, and Instagram will also avoid suggesting this type of content to a user moving forward.
  • Hidden Words, which allows people to hide suggested posts with hashtags or captions that have specific words, phrases or emojis.
  • Mute, which allows people to hide another user’s posts, Stories or messages, without the other user knowing.
  • Snooze, which allows users to snooze all recommended content in their Feed for 30 days.

Such controls play an important role in allowing users to opt out of content that may not violate our policies or guidelines but could pose problems for specific people. For instance, a user who has recently experienced a called-off wedding may prefer to avoid encountering wedding-related content. Therefore, user controls, whether through options like "Not Interested" or the ability to hide posts containing wedding-related hashtags or captions, help support a valuable and personalized user experience. 

Finally, Instagram aims for users to have a purposeful and positive experience during their time on the platform. Therefore, Instagram offers daily time limits, as well as options for users to see how much time they are spending on Instagram and to receive reminders to take breaks while scrolling through Instagram.


Ranking and recommending content on social media entails navigating a complex web of trade-offs. The central challenge lies in curating content that resonates with users while upholding principles of fairness, safety, and transparency in online content distribution.

Instagram employs a diverse array of methods, including engagement-based, survey-based, and content-based models, to rank and recommend content. Instagram continues to experiment with innovative approaches, weighing the advantages and disadvantages of each ranking model, to present users with the most valuable experience. Over the next few years, academia, civil society, and regulators will also continue to play important roles in setting parameters for the development of ranking and recommendation algorithms.  

Content ranking ultimately represents a multifaceted landscape where various models are used to fulfill specific platform goals, which can differ significantly from one platform to another. Our primary aim in this discussion has been to both clarify the intricacies of key ranking models, revealing their distinct strengths and limitations, and emphasize the vital need for cultivating a more nuanced public discourse. Such nuanced conversations are essential in acknowledging and addressing the multifaceted nature of content ranking systems, ensuring that they continue to evolve in ways that benefit users, platforms, and the broader digital ecosystem.

Author’s Notes

This piece came out of a jointly hosted session by the authors at Rightscon 2023, the world’s leading summit on human rights in the digital age, on ranking and recommendation algorithms. The section “A Look at Ranking on Instagram” was written by the Instagram team; for all other sections, the authors contributed equally. There was no payment exchanged for this collaboration.

Meta provides financial support to Lawfare. This article was submitted and handled independently through Lawfare’s standard submissions process.

Liz Arcamona is the Director of Product Policy at Instagram where she leads a team that works on global content governance and building new technologies in a responsible manner. Liz graduated from the University of Florida and Emory School of Law, and has been at Meta since 2014.
Louisa is a PhD Candidate at the Queensland University of Technology Digital Media Research Centre (DMRC) and a student member of the ARC Centre of Excellence for Automated Decision-Making and Society (ADM+S). She researches the role of algorithmic recommender systems within platform governance.
Yvonne Lee is a Stakeholder Engagement Manager at Meta, where she works with academic and civil society organizations on issues surrounding misinformation, news, and algorithmic ranking. She is also a Visiting Fellow at Johns Hopkins University, where she taught an undergraduate course on the politics of algorithms.
Dr. Sarah Shirazyan is a Content Policy manager at Meta, the team responsible for writing and interpreting global policies governing what users can share on Meta’s platforms. She leads the company’s global efforts on stakeholder engagement for developing state media, misinformation, and algorithmic ranking policies. She is a Lecturer in Law at Stanford Law School, where she teaches a course on confronting misinformation online.

Subscribe to Lawfare