Gender-based discrepancies in the algorithmic delivery of political ads on social media

Researchers studied millions of political ads during the 2024 European elections. They found that social media algorithms are associated with showing populist and far-right ads to men more often than to women. This occurs even when the ads are not specifically targeting men. This suggests that algorithmic delivery may reflect inherent biases.

The study investigates a critical tension in modern digital democracy. It looks at the intersection of automated ad delivery and political equity. Social media platforms use sophisticated algorithms to decide which users see which advertisements. These systems optimize for engagement (the level of user interaction, such as clicks or likes). While advertisers set the initial parameters, the "black box" nature of these delivery engines means the final audience often differs from the intended one. Previous research has documented that algorithmic bias can distort audiences for jobs or housing. It remains unclear if these distortions extend to the political sphere and how they correlate with specific ideologies.

The invisible hand of political delivery

Current understandings of digital campaigning often assume that the audience a political party reaches is a direct result of their own targeting choices. If a party wants to reach young voters, they select age brackets in the platform's dashboard. However, this view may overlook the role of the delivery algorithm. This algorithm acts as a secondary, automated gatekeeper.

The authors propose a plausible mechanism for these discrepancies. They suggest that algorithms might optimize toward historical engagement patterns. If certain political rhetoric triggers higher engagement from male users, the algorithm may learn to prioritize that content for men. This could create a feedback loop. In such a loop, early engagement by specific groups influences subsequent delivery decisions. This might effectively skew the gender composition of a political movement's audience. Such a pattern could narrow the range of voters exposed to diverse viewpoints. As shown in, this is not a uniform phenomenon.

It varies significantly by country, with some nations seeing much sharper skews than others.

Isolating the algorithmic signal

To investigate if this gender skew is linked to the algorithm rather than intentional advertiser targeting, the authors used a regression analysis (a statistical method to estimate relationships between variables). Their goal was to isolate the "populist" variable from other factors. This isolation is crucial. It allows researchers to move beyond simple advertiser intent to see if the skew persists even when other causes are removed.

The researchers built their model around several layers of control:

1. Content Analysis: They used a RoBERTa-based transformer model (a type of deep learning architecture) to classify the sentiment of ad text. They also used a vision transformer called DINOv2 to extract features from images and videos. This helped account for the possibility that "masculine-coded" imagery was driving the skew.
2. Targeting Controls: They included the explicit demographic targeting chosen by the advertiser, such as age groups. This ensured they were not simply measuring the advertiser's intent.
3. Platform Dynamics: They accounted for competition levels (how many other ads were active in a country) and the specific platform used (Facebook vs. Instagram).
4. Geographic Heterogeneity: They utilized country-fixed effects (a technique to control for unobserved differences between countries). This accounted for national-level differences in regulation and baseline user demographics.

By stripping away these variables, the authors could look at the association between populism and gendered delivery.

Significant skews in the extremist spectrum

The results of this isolation are notable. The paper reports that, after controlling for all aforementioned factors, ads from populist parties are associated with a 6.2 percentage point higher excess share of male unique views compared to non-populist parties .

This means that even if a populist party runs a gender-neutral campaign, the algorithm is statistically likely to push that message toward men.

The authors further examined the ideological poles of the political spectrum. They found that the gender bias is not symmetric. According to the study, ads from far-right parties are associated with a 7.0 percentage point higher excess male audience share relative to centrist parties .

Conversely, ads from far-left parties show a tendency to reach more female users. They found a coefficient of -0.074 relative to centrist ads .

This suggests that algorithmic delivery is linked to political identity. The paper finds that far-right and populist messaging correlates with a significantly more male-skewed audience. This may inadvertently worsen political polarization by siloing different genders into different ideological echo chambers.

Limits of the digital audit

The authors are careful to note the boundaries of their findings. Because this is an observational study based on data from the Meta Ad Library, the researchers cannot make definitive causal claims. They can show a strong association between populism and male-skewed delivery. However, they cannot prove with absolute certainty that the algorithm is the sole cause.

Furthermore, the scope is limited to Meta's ecosystem (Facebook and Instagram). While these platforms represent roughly 90% of the European social media market, other platforms might operate under different optimization pressures. Finally, the study lacks visibility into the "auction" side of the business. The Meta Ad Library provides data on who saw an ad. However, it does not disclose the internal pricing dynamics or the specific engagement signals that drive the algorithm. This leaves a significant part of the mechanism obscured.

A verdict on algorithmic fairness

Is the algorithmic delivery of political ads currently fair? Based on this evidence, the answer is complex. The study demonstrates that the tools used for efficient communication may redistribute political discourse along gendered lines.

For practitioners and policymakers, the takeaway is clear. Auditing the intent of an advertiser is insufficient if the delivery is biased. The authors suggest that platforms need more transparent disclosure of delivery mechanisms. They also suggest that regulators might consider safeguards. These could include randomized delivery or quota-based exposure rules. Such steps might prevent these feedback loops from undermining democratic equality. Until then, the "automated" political campaign remains a process with significant demographic skews.

Figures from the paper