🤯 LinkedIn’s Gender Secret: Engagement Explained 🚀

Tech & Science

One day in November, a product strategist – identified only as Michelle for the purpose of this experiment – logged into her LinkedIn account and altered her profile to reflect male pronouns and a new name, Michael, as she explained to TechCrunch. This action coincided with an ongoing experiment, #WearthePants, where numerous LinkedIn users – including prominent figures – were investigating the platform’s algorithm for potential bias against women. For months, many heavy LinkedIn users had voiced concerns about a decline in engagement and impressions following a statement from the company’s vice president of engineering, Tim Jurka, in August, who indicated that LinkedIn had recently implemented large language models (LLMs) to enhance content recommendations for users. Michelle, whose identity is known to TechCrunch, has over 10,000 followers and ghostwrites posts for her husband, who has approximately 2,000. Despite her significantly larger following, she and her husband consistently receive a similar number of post impressions. “The only significant variable,” she stated, “was gender.” Similarly, Marilynn Joyner, another founder, also adjusted her profile to reflect male pronouns. After changing her gender from female to male, she observed a dramatic increase of 238% in her post impressions within a single day, a trend echoed by Megan Cornish, Rosie Taylor, Jessica Doyle Mekkes, Abby Nydam, Felicity Menzies, and numerous other LinkedIn users.

The core findings of the #WearthePantsexperiment, initiated by entrepreneurs Cindy Gallop and Jane Evans, center on the complexities of social media algorithms. Experts agree that explicit sexism wasn’t the primary driver behind observed engagement declines, though implicit bias may be at play within the intricate systems that govern the Feed. Social algorithm experts explain that the Feed’s algorithms—a “complex symphony of mathematical and social levers” – continuously and simultaneously prioritize content. As data ethics consultant Brandeis Marshall told TechCrunch, “The changing of one’s profile photo and name is just one such lever.” The algorithm is also significantly influenced by a user’s existing interactions and engagement with content. Gallop and Evans, with a combined following of over 150,000 compared to the two men’s approximately 9,400 at the time, observed that Gallop’s post reached only 801 people. Importantly, the experiment highlights that side-by-side variations in feed updates, which aren’t perfectly representative or equal in reach, don’t automatically imply unfair treatment or bias—a key point repeatedly emphasized by social algorithm experts.

A man who posted the exact same content reached 10,408 people, exceeding the number of his followers by more than 100%. Subsequently, other women participated in similar experiments. Joyner, who utilizes LinkedIn to market her business, expressed concern, stating, “I’d really love to see LinkedIn take accountability for any bias that may exist within its algorithm.” However, LinkedIn, like other platforms reliant on large language models (LLMs), provides limited details regarding the training of its content-picking models. Marshall explained that many of these platforms inherently reflect a white, male, Western-centric viewpoint due to the individuals involved in their training. Researchers consistently identify human biases, such as sexism and racism, within popular LLMs, stemming from the models’ training on human-generated content and the ongoing involvement of humans in post-training or reinforcement learning. Despite these findings, the specific ways any individual company implements its AI systems remain largely obscured by the secrecy surrounding algorithmic black boxes. LinkedIn contends that the #WearthePants experiment could not have demonstrated gender bias against women. Jurka’s August statement, and subsequently reiterated by LinkedIn’s Head of Responsible AI and Governance, Sakshi Jain, affirmed that its systems do not utilize demographic information as a visibility signal. Instead, LinkedIn informed TechCrunch that it tests millions of posts to connect users to opportunities, and that demographic data is used solely for these testing purposes, specifically to assess whether posts “from different creators compete on equal footing.”

“The company emphasized that the scrolling experience—what users see within the feed—is consistent across all audiences,” LinkedIn stated to TechCrunch. LinkedIn has been actively researching and adjusting its algorithm to mitigate potential bias and deliver a more neutral experience for its users. According to Marshall, the specific reasons behind some observed increases in impressions for women, such as those who shifted their profile gender to male, remain largely due to unidentified variables. For instance, participating in a viral trend can significantly boost engagement, and accounts returning after extended periods of inactivity may have been rewarded by the algorithm. Michelle reported a dramatic shift—a 200% increase in impressions and a 27% rise in engagement—when she adjusted her tone to match a more direct, concise style, similar to that she uses for her husband. She concluded that the system wasn’t “explicitly sexist” but appeared to consider communication styles typically associated with women as a “proxy for lower value.” Researchers have long observed that stereotypical male writing styles tend to be more concise, while those associated with women are often perceived as softer and more emotional. If a Large Language Model (LLM) is trained to prioritize writing that conforms to male stereotypes, this represents a subtle, implicit bias, a phenomenon particularly prevalent in most LLMs as previously reported. LinkedIn’s AI systems assess hundreds of signals to determine content prioritization for users.

Insights derived from an individual’s profile, network, and activity play a key role in shaping the LinkedIn feed. The platform continually conducts tests to determine what content best serves users’ career needs, and member behavior—specifically, what they click, save, and engage with—influences the feed daily. This user interaction naturally impacts what content appears alongside updates from LinkedIn itself. Chad Johnson, a sales expert active on LinkedIn, noted that the new Large Language Model (LLM) system prioritizes writing quality over posting frequency or timing, stating, “It cares whether your writing shows understanding, clarity, and value.” Consequently, it’s increasingly difficult to pinpoint the cause of the observed #WearthePants results. Despite this complexity, several users have expressed dissatisfaction with LinkedIn’s algorithm. Shailvi Wakhulu, a data scientist, reports a significant decline in her post impressions, having averaged at least one post per day for five years and previously receiving thousands of views. Now, she and her husband average only a few hundred. “It’s demotivating for content creators with a large, loyal following,” she stated. Other users have experienced markedly different outcomes. One individual reported a 50% decrease in engagement over the past few months, while another noted a 100%+ increase in post impressions and reach during the same period. This disparity, according to this user, is driven by a targeted content strategy focused on specific topics and audiences, aligning with the algorithm’s intended function.

“We’re seeing rewarding results,” he told TechCrunch, adding that his clients are observing a comparable increase. However, Marshall, who is Black, believes that posts detailing her personal experiences receive fewer interactions than those focused on her professional expertise. “If Black women only receive engagement when they discuss their experiences as Black women, rather than their particular areas of knowledge, then that indicates a bias,” she explained. Dean, the researcher, suggests the algorithm may be amplifying existing signals, potentially rewarding certain posts not due to the demographics of the writer, but because of a history of prior responses across the platform. LinkedIn has provided some insights, noting a user base growth that has led to a 15% increase in posting activity year-over-year, alongside a 24% rise in comments. “I want transparency,” Michelle stated. Nevertheless, achieving transparency is a significant challenge, given the proprietary nature of content-picking algorithms and the potential for users to strategically ‘game’ them – a request unlikely to be fully fulfilled.