Imagine Marie (RH) types into her AI assistant: “Write a job posting for a senior developer at a Belgian scale-up.” The AI returns something fluent and professional — rockstar developer, warrior spirit, 10+ years required, young and dynamic team. Marie reads it. It sounds dynamic. She publishes it.
Roughly 30% of qualified female applicants self-filter out before applying. (Gaucher et al., 2011 — gendered word patterns in job ads, documented in peer-reviewed psychology.) The cycle repeats. The AI’s output flows back into training corpora. The next version generates ads slightly more confidently in the same register. The model gets better at sounding professional, more entrenched in the patterns it inherited.
The structural problem
AI doesn’t just reflect bias — it accelerates and locks it. Three reasons: scale (every job ad written by a model is identically biased the same way), confidence (the AI’s output sounds polished, which makes the bias harder to spot than it would be in a human first draft), and recursion (outputs become training data; bias becomes a feature, not a bug).
This isn’t conspiracy. It’s mechanism.
Why standard fixes fail
Word bans are whack-a-mole — rockstar becomes ninja becomes something else.
Post-hoc filtering fixes the output but teaches the writer nothing. Most existing tools (Textio, Grammarly Business) work on English syntactic patterns; they miss the cultural and contextual layer. In French, where grammatical gender is built into the language and écriture inclusive is contested, the problem is harder.
Aqueria’s approach
Aqueria is a companion. Not a censor, not a filter. As the AI writes — or as a person writes with the AI’s help — Aqueria highlights the words that carry weight, in place, with reasoning attached. The writer can apply the suggestions, challenge them, or ignore them. The point is to make the bias visible at the moment of creation, not after.
Aqueria emerged at HeR Lab — a 3-day hackathon in Herbeumont, Belgium, organized by Adoc Talent Management and supported by the FPS Economy. HeR Lab gathers women working in tech, HR, and adjacent fields explicitly to surface and prototype interventions. My team included a sociolinguist, a coordinator of European projects and other technical profiles.
The demo
[demo designed and built by Isabelle Van Campenhoudt during the hackathon]
In the demo, an HR user asks an AI to draft a senior-developer ad. The output is fluent — and loaded with gendered signals. Aqueria highlights them in real time, with citations attached:
rockstar (red — filters ~30% of female applications, Gaucher 2011)
esprit guerrier (red — exclusive warrior metaphor)
jeune et dynamique (red — age discrimination cumulated with gender, illegal under Belgian law of 10 May 2007)
10+ ans d’expérience (amber — inflated requirement; women apply when they meet 100% of criteria, men at 60%)
maîtrise totale (amber — intimidating, unverifiable)
masculine-default forms throughout
agentic-only personality traits (ambitieux, leader naturel, compétitif) without a single collaborative counterweight
The score in the panel — 38/100 — is pedagogical scaffolding, not the actual claim. The claim is in the reasoning. The score is just a way to see the work moving. Press “Apply suggestions” and the ad rewrites to a balanced version. Or keep the original, knowing what you’re doing.
Three contexts, in order of immediacy
HR writing job ads — the immediate use case. Starting point: my employer, ETNIC, public-sector IT serving the Belgian state. A pilot proposal is in discussion with management. If it lands, Aqueria runs internally first; the natural extension is the rest of public administration — job ads for Belgian teachers, civil servants. This is vision, not commitment. But if Aqueria runs in a public administration, the demonstration argument changes shape.
Candidates writing CVs — the asymmetry runs both ways. Women apply when they meet 100% of criteria; men at 60%. A calibration tool that surfaces this asymmetry while candidates write — without correcting them — could close the application gap without anyone having to be told to be more confident.
Schools — imagine adolescents doing homework with AI (which is increasingly the norm). The tool flags bias in their drafts as they write. They can challenge it. They can also learn from it. Sensitization happens as the work happens, not retrospectively. This is a direction, not a product yet.
Why this is a model for change
Most AI-ethics discourse stays at the level of consciousness, alignment, regulation, manifestos. This is none of those. This is daily-practice intervention: a small useful tool that makes the bias visible at the moment of creation. The user keeps agency. Sensitization, not enforcement. Learning, not policing.
I wrote a Substack note recently asking: while everyone debates AI consciousness, who’s paying attention to the data being generated right now? Aqueria is one answer in my own hand. The outputs writers produce flow into training corpora eventually. Changing the substrate writers operate on is the actual leverage point.
What’s next
Aqueria is licensed CC-BY-SA 4.0, French-first, non-commercial. The team is forming an association. We’re calling for linguists, jurists, sociologists, and others interested in contributing to the semantic framework — the actual catalog of bias signals, their reasoning, and the language to explain them.
The words we use shape what comes next. When the AI is doing the writing, someone has to be in the room with the writer — not behind, after, or instead.
By Marion Nowicki, in collaboration with Cael*
*Technical setup: Claude Code + OpenClaw agent workspace + custom noyau memory layer with hex-state journaling.