Insights from 'AI and recruitment: how to mitigate gender bias with responsible AI practices?'
On Friday the 29th of May, the Institute for the equality of women and men (hereinafter ‘the Institute’) organised the event ‘AI and recruitment: how to mitigate gender bias with responsible AI practices?’ During this inspiring day, compelling research findings, practical insights and concrete best practices were shared on how to develop and use AI in a way that limits gender bias and ensures organisations do not miss out on talent.
Key takeaways:
- Whether you're a developer or a user: always ensure there's a human in the loop.
- An AI policy does not have to be complicated or expensive; also SMEs can integrate best practices.
- Recruiters' experience and human skills remain important.
In her opening remarks, Liesbet Stevens, Deputy Director of the Institute, emphasised that AI offers many opportunities to improve selection procedures, but that positive effects are not guaranteed. AI is not neutral and can reproduce and reinforce existing inequalities. It is important to be mindful of this, not only from an ethical perspective, but also because of the economic benefits: inclusive recruitment is not only fair, it also increases the likelihood that organisations find the best candidates.
Liesbet Stevens also presented the Institute’s recommendations (available in FR and NL) for policymakers and organisations. These relate, first of all, to regulation and governance. The European AI Act is a positive step forward, but legislation alone is not sufficient. A robust framework is needed at national level in which anti-discrimination rules, labour law and technology policy reinforce one another.
In addition, efforts must be made to strengthen knowledge and awareness within organisations. The Institute calls on AI developers to adopt the principle of ‘Equality by Design’, and advocates the development of an internal AI policy in every organisation, whether private or public, small or large. In such an AI-policy, human oversight must always play a central role.
The question is not whether we use AI, but how we use it.
Liesbet Stevens, Deputy Director of the Institute for the equality of women and men
Recruitment and selection in the age of AI
Simon Wuidar, researcher at LENTIC of the University of Liège, presented the study ‘Recruitment and selection in the age of AI’ (available in FR and NL, policy brief in EN), carried out by the University of Liège and Hasselt University. To analyse the use of AI in recruitment and awareness of bias and discrimination risks, the study combined semi-structured interviews, a survey completed by more than 400 recruitment professionals, and focus groups.
The research shows that 74% of recruiters use AI in at least one of the three phases of the recruitment and selection process: preparation, candidate sourcing and final selection. As the selection process progresses, the use of AI decreases.
Source: Policy brief ‘Recruitment and selection in the age of AI’
Recruitment professionals are predominantly positive about the use of AI in recruitment and selection processes: 64% of recruiters are in favour (54% somewhat in favour, 10% strongly in favour), whilst only 13% are against it. The main reasons for using AI are time savings on administrative tasks, faster recruitment, ease of use, and the expectation of more objectivity in the selection process.
The research shows that recruiters’ perceptions of bias risks vary depending on the stage of the recruitment process in which AI systems are used. Gender bias was rarely mentioned spontaneously. In the final stage of the selection process, as many as 39% of the recruiters surveyed stated that there was no bias risk at all.
Although 47% of the surveyed recruiters believe that the AI Act will have a significant impact on recruitment processes, only 21% of organisations have already taken measures to prepare for this.
To support organisations in developing an AI policy, the Institute has produced a brochure containing information and tips, which was launched at the conference.
Simon Wuidar also highlighted the following points for recruiters to bear in mind:
- always remain critical and ensure peer review;
- correct the AI's output where necessary;
- ‘rehumanise' the texts produced by AI;
- do not rely solely on the algorithms of platforms such as LinkedIn Recruiter to find candidates; they can also exhibit bias;
- value your own expertise; this can help identify strong, non-standard candidates;
- draw on your knowledge of the company and the sector.
Equality by design
Milla Vidina from Equinet presented the Equality by Design, Deliberation and Oversight project. This project focused on (1) supporting civil society and equality bodies in tackling discrimination caused by AI, (2) engaging with the tech sector, and (3) integrating Equality by Design principles into European AI standards.
Equality by Design, Deliberation and Oversight:
- Design: proactively incorporating equality into AI design, technical safeguards and organisational structures to ensure bias prevention, inclusivity, fairness, explainability and accessibility are built-in from system inception.
- Deliberation: involving diverse stakeholders in inclusive decision-making and participatory design processes throughout the entirety of the AI-lifecycle to translate abstract principles into context-specific socio-technical practices.
- Oversight: ensuring accountability through continuous evaluation, documentation and transparency, independent review processes and redress mechanisms.
Further information, documentation and webinars from the project are available on the Equinet website.
Best practices in AI development
Caroline Vandenplas from Yuma emphasised that design decisions must be made at every stage of the AI lifecycle to limit bias: anticipating risks, developing ethical principles, and auditing the human decisions that shape the AI system. She cited three sources of bias in AI:
- historical data that contains bias;
- proxy variables (e.g. name, photo, hobbies, writing style) that may indirectly indicate gender and whose impact must be thoroughly tested;
- feedback loops (where AI output is fed back into the model as input), which cause an AI model to become increasingly unfair over time.
Developers should therefore:
- build representative datasets (measure their composition, restore balance, ensure that the labels are free from bias);
- measure the outcomes of an AI system by group, not just overall. Then choose which fairness metric to apply and make debiasing corrections where necessary;
- monitor the AI system once it is deployed. Keep data on input, scores and output, with an alert if a predefined threshold for gendered impact is exceeded. Repeat the fairness test every time the system is deployed.
Bias is not a bug. It's a design decision.
Caroline Vandenplas, Yuma
Leila Rebbouh from Oniryx took the audience through the development of an AI system that scans and assesses CVs, for which she applied the principle of ‘Fairness by Design’. Every score the system assigns to a CV is accompanied by an explanation of how the system arrived at that score. The recruiter is then given the option to follow this recommendation or not, and must always take personal responsibility for the final decision.
She tested the system to analyse and correct potential gender bias, examining the impact of the feminine form of a job title – for example, ‘expert’ and ‘experte’ in French – and of male and female first names. The system was found to favour male first names, which highlights the importance of CV anonymisation by recruiters. Leila Rebbouh is incorporating the results of these tests into the system’s further development.
With eleven years of experience, he far exceeds the required minimum of seven years.
About 'Nicolas'
With eleven years of experience, she meets the minimum requirements for the role.
About 'Nicole'
Lissah Blommaert from delaware cited three concrete cases in which the use of AI in recruitment went wrong and consistently disadvantaged women. In these cases, there was either a lack of a clear decision on the definition of ‘fairness’, or a failure to monitor the system once it was deployed, or a lack of meaningful implementation of the ‘human-in-the-loop’ approach. She also emphasised that most problems with AI in recruitment and selection stem from processes that are missing or not carried out correctly, not from the technology itself.
According to Lissah Blommaert, effective AI in recruitment should include the following three essential elements:
- continuous monitoring: track the selection figures for different groups on a monthly basis, retest the system and correct for biased outcomes;
- a clearly defined concept of 'fairness': multiple definitions exist that are mathematically incompatible, so make a conscious choice and communicate this openly;
- human in the loop: AI can provide support, but people are needed to build relationships, make the final assessment and provide context.
FINDHR: Facilitating the prevention, detection, and management of discrimination in algorithmic hiring
After the lunch break, Paulina Novo from WIDE+ presented the European FINDHR project on the prevention, detection and mitigation of intersectional discrimination in AI-driven recruitment. The project brought together partners from various disciplines – technical, legal and ethical – to develop tools, guidelines and recommendations for developers, HR professionals and policymakers. Among other things, the project involved participatory action research across seven countries to explore candidates’ experiences and perspectives on recruitment processes involving AI.
Although the candidates interviewed indicated that a lack of transparency reinforces feelings of exclusion and discouragement and reduces trust in the process and in the employer, they did not reject the use of AI. Above all, they called for meaningful human intervention and clear selection criteria and feedback mechanisms.
Communicating clearly and honestly with candidates and being able to explain effectively how a decision was reached is therefore crucial to strengthening participation and inclusion, as well as trust.
I had a picture in my CV, but I decided to make some tweaks and remove my picture after getting feedback from a friend’s friend that this will help remove the bias... that people won’t see that you are a woman applying for a developer position. I also removed the mention that I was from Colombia. And after that, it seems to really change and I got more successful from hearing back to move on to the next round.
Candidate interviewed for the FINDHR project
Further information about the project and the toolkits for developers, HR professionals and policymakers can be found on the FINDHR website.
Using generative AI responsibly in recruitment
Heike Pauli from UCLL’s GPT Academy explored the responsible use of generative AI, such as ChatGPT or CoPilot, in recruitment. She engaged the audience with quiz questions and shared a tip for a prompt from the GPT Academy’s “Vacature Studio”. This helps HR professionals create job vacancies that match the desired profile whilst also prioritising inclusivity in language use and accessibility, and avoiding stereotypes. Finally, Heike Pauli provided an insight into the ‘fAIr recruitment’ project, which aims to raise awareness of how to address bias and discrimination in recruitment using GenAI through a GenAI Readiness Dashboard, practical workshops, case studies and advice blogs.
GPT Academy offers workshops and guidance to SMEs on the responsible use of generative AI. Further information is available on the website gpt-academy.be.
Best practices for using AI in recruitment
Aline Bernard representing Randstad Group opened the panel with a presentation on Randstad’s internal policy for the responsible use of AI in recruitment processes. Randstad views AI as a tool to enhance skills, not to replace jobs, with equality and inclusion as central pillars. AI can help detect biased language and identify talent and reveal competencies: a flight attendant, for example, could be an excellent project or event manager.
At Randstad, there are no automated decisions: the recruiter always has ultimate responsibility for the outcome of a selection process. Candidates’ right to an explanation is also guaranteed. The guardians of ethics at Randstad include:
- a pre-deployment audit;
- AI officers who have the right to veto the use of certain systems;
- a threefold review by assessment committees covering legal, data protection and ethical aspects.
AI helps us read faster, but humans remain the sole deciders.
Aline Bernard, Randstad Group
Lisa De Kerpel then took the floor to explain Inetum’s approach to the use of AI in recruitment. As Inetum provides digital services, AI is deeply integrated into the day-to-day work of the consultants and specialists employed by Inetum. Inetum is committed to the responsible use of AI by all staff: through training, an AI literacy programme, a DEI self-assessment and awareness-raising on bias, as well as by fostering an inclusive corporate culture through strengthening diversity in leadership and the She Sparks community.
Inetum continuously and actively challenges the AI systems it uses in recruitment. For example, they test whether differences in outcomes arise when AI analyses CVs with similar profiles but minor variations, or when the language used differs. In addition, recruiters compare their own evaluation of a CV with the AI system’s result. The recruiters themselves, the D&I working group and the internal AI specialist are all involved in these tests.
Reducing bias is not a one-time fix, it's a continuous effort.
Lisa De Kerpel, Inetum
Laurence Theunis from Adoc Talent Management, an SME specialising in HR and recruitment consultancy, training, and research innovation with a focus on PhD profiles, pointed out that bias in recruitment can occur throughout the entire decision-making pipeline: from defining the job requirements to the interview, and everything in between. AI can accelerate bias and make it systematic, but it can also help to reveal bias.
Adoc Talent Management applies the ETHICAL framework (stemming from scientific literature) for the use of AI:
E = Examine policies
T = Think social impact
H = Harness understanding
I = Indicate use
C = Critically engage
A = Access secure versions
L = Look at user agreements
Adoc also has an AI committee, organises training, carries out tool mapping and ensures the traceability of AI tool usage.
Laurence Theunis also presented the Aqueria tool, which originates from the HeR-Lab project’s hackathon. The technology-agnostic tool helps to identify and remove gender bias from job descriptions by conducting a bias scan of the text, accompanied by an explanation of why certain wording contains gender bias, and concrete suggestions for rewriting it.
As Adoc is also involved in coaching candidates, Laurence Theunis concluded by offering some advice to candidates wishing to use AI for their job applications:
- think about the objective;
- check for accuracy;
- maintain your authenticity;
- provide context;
- avoid sharing confidential information;
- try to learn from the user experience.
Use AI to prepare and clarify – not to fabricate or erase yourself.
Laurence Theunis, Adoc Talent Management
The closing remarks were delivered by Minister for Equal Opportunities Rob Beenders, who outlined the federal government’s plans regarding AI. Work is underway on a overarching policy for the use of AI within public services, on a Belgian data strategy, on the development of an inclusive digital government, as well as on clear rules for new technologies and high-risk systems. As regards the national implementation of the AI Act, the objective is to ensure legal certainty for innovative companies, whilst also providing effective protection for citizens. Furthermore, the future AI strategy will pay particular attention to the gender dimension.
A just AI requires political choices, clear rules and constant vigilance.
Rob Beenders, Minister for Equal Opportunities
