Most advertisers assume that if they are not explicitly targeting by race, gender, or age, their ads are being delivered neutrally. This assumption is wrong, and it has been wrong for years.
Advertising algorithms do not discriminate because someone programmed them to. They discriminate because they optimise. And optimisation, applied to real-world data that reflects historical patterns of behaviour and engagement, reliably reproduces and often amplifies the biases already present in that data.
This is not a theoretical concern. It has been documented, litigated, and settled. Understanding it is not just an ethics issue — it is a performance and strategic one.
How Algorithmic Bias Actually Works
When Meta’s delivery system decides who to show your ad to, it is working from an enormous model of who is likely to engage, click, and convert based on historical behaviour across billions of interactions. That model is extremely good at finding patterns. The problem is that patterns in data reflect the world as it has been, not the world as it is or should be.
If your historical converters are predominantly a particular demographic, the algorithm will look for more people like them. If engagement rates on certain content types are higher among certain groups, the algorithm will weight delivery towards those groups. If certain categories of product or service have historically attracted certain audiences, the delivery system will lean into those patterns — whether or not the advertiser has made any deliberate targeting choice.
This creates a loop. Biased historical data produces biased delivery. Biased delivery produces more data that looks biased. The algorithm gets more confident in the pattern. The skew compounds.
The Evidence Is Not New
In 2019, the US Department of Housing and Urban Development filed a complaint against Facebook, alleging that its ad delivery system enabled housing discrimination by allowing advertisers to exclude people based on race, gender, and national origin. Facebook settled. The outcome included commitments to overhaul how housing, employment, and credit ads were targeted.
But the more revealing finding came from subsequent academic research. Studies showed that even without any explicit demographic targeting, Facebook’s delivery algorithm distributed job ads with significant gender skews. Ads for roles in male-dominated industries were shown predominantly to men. Ads for roles in female-dominated industries were shown predominantly to women. The advertiser had not made a targeting choice. The algorithm had made it for them.
The mechanism was simple: the system optimised for engagement, and engagement patterns reflected existing labour market demographics. The algorithm was not being malicious. It was being efficient. Efficiency, applied to a biased baseline, produces biased outcomes.
What This Means for Advertisers
For most brands, the immediate implication is that your ad delivery is probably not as demographically even as you think it is, even if you are running broad targeting.
If your creative signals, your landing page, your offer, and your historical conversion data all skew towards a particular audience, the algorithm will find more of that audience. This might sound like good optimisation. Often it is. But it can also mean you are systematically underdelivering to audiences who would actually convert if they saw the ad — audiences the algorithm has deprioritised because they do not match the existing pattern.
This is particularly acute for brands trying to reach new or underrepresented audiences. If your existing customer base does not reflect the full potential market, your lookalike audiences and your delivery optimisation will reproduce that gap rather than help you close it.
The Women’s Health Dimension
For brands operating in women’s health, wellness, and femtech, algorithmic bias is not an abstract concern. It intersects directly with two real problems: reach limitations and engagement signal distortion.
Women’s health has historically been under-researched, under-discussed, and under-represented across media. The result is that algorithmic training data for categories like menstrual health, menopause, fertility, and pelvic health is thinner and more skewed than for mainstream health topics. Algorithms trained on engagement data have less signal to work with and more noise.
At the same time, the creative and messaging restrictions that platforms apply to health content — restrictions that are themselves applied unevenly — mean that compliant women’s health content often performs worse on pure engagement metrics than less restricted categories. The algorithm does not distinguish between low engagement caused by creative weakness and low engagement caused by platform suppression. It just sees a signal that says this content type does not perform, and it responds accordingly.
The practical effect is that women’s health brands can find themselves in a delivery loop where policy restrictions reduce engagement, reduced engagement signals low quality, low quality signals reduce distribution, and reduced distribution further limits the data available to optimise against.
Audit Your Own Delivery
The most useful thing most brands can do is look at their own delivery data with more scrutiny than the headline metrics typically receive.
Meta’s ad delivery reports break down impressions by age, gender, and placement. Most advertisers look at these reports to understand their audience. Few look at them to ask whether the delivery aligns with their intended reach, or whether there are groups the algorithm is systematically underserving.
If you are running broad targeting and your delivery is heavily skewed towards one demographic, that is worth questioning. It might reflect genuine affinity. It might reflect historical data creating a self-fulfilling pattern. It might reflect creative signals that are inadvertently excluding certain audiences. The report alone will not tell you which, but it will tell you whether to ask the question.
The Creative Signal Problem
Creative choices encode targeting signals whether advertisers intend them to or not.
The people in your ads, the language used, the visual aesthetic, the scenarios depicted — all of these communicate to the delivery algorithm who this ad is for. An ad featuring people who skew towards a particular demographic will tend to be delivered more towards that demographic. An ad using cultural references or terminology that resonates with one group more than another will reflect that in delivery patterns.
This is not an argument for sanitising creative into inoffensive neutrality, which would simply make it less effective. It is an argument for being intentional about the signals your creative sends, and testing whether those signals are serving your actual audience strategy or inadvertently narrowing your reach.
Optimisation Is Not Neutral
The broader point here is worth stating clearly: algorithms are not objective arbiters of relevance. They are systems trained on historical data, optimising for defined metrics, operating in a world that has never been neutral.
That does not make them useless. It makes them powerful tools that require intelligent oversight. The question for any advertiser is whether they are directing the algorithm with enough signal and scrutiny to get it pointing in the right direction, or whether they are letting historical patterns drive outcomes without examining whether those outcomes are actually what they want.
The brands that will use AI-driven platforms most effectively over the next few years will not be the ones who let the algorithm run. They will be the ones who understand what the algorithm is actually doing, and who bring enough strategic thinking to guide it towards the right result.
0 Comments