What a viral LinkedIn post, three layers of evidence, and one overlooked bias teach us about real AI literacy.
A cybersecurity leader recently shared an observation that resonated with me: her posts about pregnancy, IVF and life as a military spouse reach 25,000+ impressions. Her posts about cybersecurity, AI strategy and CISO leadership — her actual profession — reach about 500. And when the platform curated twenty “Top Perspectives” on a major security incident, three appeared to be women.
What the evidence actually shows — in three layers
Layer one: the viral experiments. In the #WearThePants experiments, women switched their LinkedIn profile gender to male and documented impression jumps of 200%+ within days. LinkedIn responded that its algorithms don’t use demographic signals for content visibility (at least not directly, but we all know demographic signals creep in usually thru proxies). In fairness, the experiments were confounded: several participants also changed their writing style when posting “as a man.” Suggestive, widely felt, not proof.
Layer two: the platform’s own published research. In 2019, LinkedIn’s researchers published a fairness-aware re-ranking framework for Talent Search. After deployment, the share of searches returning demographically representative results nearly tripled — with no loss on business metrics — and the system was rolled out to all Recruiter users. Read that backwards: it is a quantified, peer-reviewed admission of how skewed the rankings were before the fix. Credit where due — publishing and fixing is what responsible self-audit looks like. But it also tells us the skew was real, large, and invisible to users until the platform chose to measure it.
Layer three: the audit gap. A recent independent external audit of LinkedIn Recruiter’s rankings found demographic disparities in how stable candidates’ rank positions are over time — and, just as importantly, documented how close to impossible external auditing is: full platform access for independent researchers effectively doesn’t exist. Which leaves us with the honest answer to “is the feed suppressing women’s technical voices?”: we cannot independently verify it — and that is itself the finding. A system this consequential for careers, visibility and hiring is currently auditable only by itself.
Add one more layer that starts before the algorithm entirely: research on LinkedIn profiles shows men and women present themselves differently — down to how skills are listed and repeated — and ranking tools trained on those profiles inherit the difference. The bias doesn’t wait for the model. It walks in with the data.
The bias nobody prices in
Now follow the consequence downstream, because this is the part that I find missing from almost every AI-in-HR conversation.
For most organisations, LinkedIn is the first filter of the hiring funnel. It’s where sourcing starts, where recruiters search, where “the market” is operationally defined. So here is the uncomfortable arithmetic: if the pool is skewed before your process begins, it doesn’t matter how balanced and bias-free your AI-powered ATS is. You can audit the model to perfection, run every fairness metric, pass every vendor assessment — and still produce skewed outcomes, because your beautifully governed system is processing a pre-filtered sample. You audited the model. Nobody audited the funnel.
This is the class of bias that doesn’t occur to most people, because it lives upstream of everything we call “our hiring process.” It has no vendor to blame and no dashboard that shows it. And it quietly breaks the most common reassurance in the industry:
“But AI corrects human bias”
Partly true — and dangerously incomplete. Years ago I wrote about the cognitive biases that plague HR decisions: availability heuristics, proximity bias, anchoring, groupthink. Against those, data genuinely helps — a productivity analysis doesn’t care which employee sits closest to the manager, and a well-built survey beats hallway impressions every time. That’s the true half.
The false half: AI corrects the biases that show up as inconsistency between human judgments. It does nothing against the biases baked into its training data — those it doesn’t correct, it industrialises. A human recruiter’s skew touches one shortlist at a time. A model’s skew touches every shortlist, identically, at scale, with a confidence score attached. “AI fixes human bias” and “AI amplifies human bias” are both true — for different biases. Knowing which is which is precisely the skill our field is missing.
This is what AI literacy actually means
Which brings me to my real point. The market has decided that AI literacy means prompt engineering. I’d like to push back as firmly as I can.
Prompt engineering is tool fluency. AI literacy is the capability to ask, unprompted: where does risk enter this system, for which stakeholder, and what control do I set against it? Not just understanding the answer when someone explains it — having it occur to you. Seeing that a sourcing platform is part of your decision system even though you never procured it. Noticing that an input metric can disadvantage someone by design. Recognising that context, inputs, decisions and actions are four different places where harm can enter — and that each needs its own control.
And crucially: this literacy applies to the user, not just the platform. A recruiter who knows where their sourcing channel skews… sources differently. Adds channels. Checks slate composition before the process starts, not in the year-end diversity report. Treats “where did this pool come from?” as a standard control question, the way finance treats “where did this number come from?” That’s human agency in an AI-mediated process — not the ceremonial approval click at the end, but informed judgment at the point where the funnel is formed.
The fixes follow the layers. At platform level: external auditability — systems this consequential shouldn’t be auditable only by themselves, and regulation is slowly moving that way. At organisational level: funnel-stage controls — measure the composition of what enters your process, not only what exits it. At individual level: exactly what that cybersecurity leader did — elevate the expertise the algorithm doesn’t, and make your own sourcing choices consciously.
The platform decides what it amplifies. You decide where you look. AI governance, at its core, is the discipline of never confusing the two.
Fairness is tested across the whole funnel, with evidence.