Before big tech engineers can improve the fairness of recommendation systems, such as social media feeds and online shopping results, they need to define what “fairness” even means.

Should an app show people only the content it predicts they will like most, or should it boost newer creators, small businesses or historically underrepresented groups? Should an online store rank products purely by past clicks and sales, or make sure independent sellers can compete with dominant brands?

“Recommendation systems are particularly prone to the ‘rich get richer’ effect,” said Allison Koenecke, assistant professor of information science at Cornell Tech. “Top-ranked items often get more clicks, which can lead to disproportionately inflated metrics for those items, cementing their place at the top of a search feed or social media page – and perhaps unfairly penalizing slightly-lower-ranked items that could be higher quality.”

Koenecke, along with co-lead authors Emma Harvey, a doctoral student in information science based at Cornell Tech, and Jing Nathan Yan, Ph.D. ’24, is an author of “Fairness-in-the-Workflow: How Machine Learning Practitioners at Big Tech Companies Approach Fairness in Recommender Systems,” which was presented at the 2026 Association for Computing Machinery CHI Conference on Human Factors in Computing Systems. The research was also authored by Junxiong Wang, Ph.D. ’24 and Jeffrey Rzeszotarski, assistant professor at Loyola University Maryland.

In addition to defining fairness, the study found, big tech companies need to incentivize fairness efforts, to establish a shared language around fairness issues so various specialists can communicate about it, and address fairness issues early in the development process, when they’re easier to change.

“There is so much academic work that suggests approaches for improving fairness – but to be impactful, that work has to actually be adopted in practice,” Harvey said. “It’s important to understand the barriers that practitioners face when trying to incorporate fairness into their workflows so that we can overcome those barriers.”

Unlike a problem with a simple technical fix, fairness often depends on human decisions: what a company chooses to reward, whose interests it prioritizes, and what tradeoffs it is willing to make. In interviews with 11 machine learning practitioners working on recommendation systems at large technology companies, the researchers found that those decisions are also shaped by how a company is organized – for example, whether past fairness efforts were documented, and whether legal or fairness teams are involved early in product development or only after a system has launched.

Another major barrier is incentives. The researchers found that most recommender system practitioners reported spending 10% or less of their time on fairness-related work. Their day-to-day responsibilities were more likely to focus on system performance, reliability and keeping products running smoothly.

“Even when technical practitioners genuinely care about creating fair systems, if they are only evaluated on system performance and uptime, those are the metrics that they will have to prioritize,” Harvey said.

In addition, fairness work often depends on lessons learned from past projects, but those lessons are not always well-documented or easy to reuse. “I was surprised by the extent to which documentation debt within big tech companies prevents practitioners from building institutional fairness knowledge,” Harvey said.

The study also found that fairness efforts can break down when teams speak different professional languages. Engineers, legal teams, policy specialists and fairness experts may all share the same goal, but approach problems through different frameworks.

“I was struck by the need for a ‘fairness lingua franca,’” Harvey said. “Even within the same company, technical practitioners struggle to connect with and understand fairness practitioners due to their different disciplinary backgrounds.”

The researchers found that timing matters, too. In many companies, fairness teams are brought in after a product or model has already launched – when changes can be slower, harder and costlier.

That matters because recommendation systems do not stay static after launch. They constantly adapt based on user behavior, creating feedback loops that can amplify bias or create new disparities over time. A system that favors already popular creators, for example, may make them even more dominant.

“It is imperative for companies to consider their organizational structure when prioritizing algorithmic fairness in technical processes,” Koenecke said. “This can be done by, for example, ensuring that internal fairness teams work directly with recommendation system product teams early on in the development process, such as during data collection and offline prototyping, rather than as a post hoc fairness check after a model has been released.”

This work was supported by a grant from the National Science Foundation.

Grace Stanley is a staff writer-editor for Cornell Tech.