Active Sourcing Tools in Switzerland: Between Efficiency and Fairness

Large Swiss companies are increasingly relying on digital tools in recruiting – from LinkedIn Recruiter to ChatGPT. But how effective are these tools really? Caroline Straub and Mascha Kurpicz-Briki have investigated how widespread active sourcing tools are, where their strengths lie, and what role unconscious biases play in digital recruiting.

You’ve researched the human-centered aspects of HR sourcing software in Swiss companies. What exactly was investigated here? Why is this topic significant?

Caroline: Headhunters and HR departments frequently use active sourcing tools to identify potential candidates. Active sourcing is the direct approach of promising candidates to win them over for one’s own company. Digital matching tools are increasingly being used in active sourcing to track down suitable candidates online. Tools are meant to make recruiting processes more efficient. However, simple filtering and screening methods can lead to suitable candidates being overlooked and others being wrongly invited to more expensive selection rounds like interviews. This is a disadvantage for companies, and applicants risk discrimination. We researched how widespread the use of active sourcing tools is in Switzerland, how they are applied, and whether they lead to biases in selection processes.

How did you proceed?

Caroline: We first interviewed ten recruitment professionals (headhunters, HR recruiters in Swiss companies). Through the interviews, we obtained information about the prevalence, software type, and use cases of these tools. The interviews clarified how well professionals understand the tool, what problems occur, and how useful it is, for example, in searching for people and reducing the skilled worker shortage.

Mascha: In the second step, we conducted an analysis of the sourcing software products mentioned by the participants. The primary aim was to evaluate what technologies are behind them and what risks of discrimination they bring. For this, we conducted both an online research about the software (e.g., manufacturer website, experience reports) and a literature review (existing studies about this software). Based on this, we selected 2 software products and conducted further black box testing for potential bias. This involves testing the software with different inputs and analyzing the output. This is particularly relevant since we often don’t know how commercial software is implemented or what data was used for AI software. This is also a major problem for applications based on language models like ChatGPT.

Which tools are commonly used and in what situations?

Caroline: Most companies use active sourcing tools to search for passive candidates. These are people who aren’t looking for a new position and are contacted by companies to recruit them. All surveyed companies use LinkedIn Recruiter to search for passive candidates. Textkernel is also frequently used to search internal applicant pools. Some companies rely on Phenom, an AI-powered talent management tool. ChatGPT is also used, for instance, to generate outreach messages or job postings.

How should I imagine this? Does the tool do everything automatically?

Caroline: No, it’s always a combination of HR professional, line manager, and the tool. Typically, the line manager provides the job requirements and passes these to the active sourcer. The active sourcer creates a longlist of possible candidates using the tool. In the third step, this longlist is discussed with the line manager, and only then does (non-automated) contact initiation occur. Therefore, it remains human-centered.

How effective are the tools in searching for people?

Caroline: The tools are particularly suitable for areas like IT, marketing, and business development. They are less effective for blue-collar workers or in healthcare, as these target groups are less likely to be found on LinkedIn. In these cases, companies tend to rely on alternative methods like Facebook campaigns, Google Ads, or even offline advertising. The goal of these campaigns is for the company to showcase its employer benefits. When it comes to interaction, personal approach takes priority. Tool-based sourcing is more questionable in these cases.

What problems can occur during application?

Caroline: The interviewees cite LinkedIn’s limited search function as the main criticism. The search is based exclusively on keywords in profiles, meaning incomplete profiles or missing keywords can lead to overlooking suitable candidates. Soft skills and leadership experience can only be indirectly determined through job titles, which necessitates manual review of profiles.

What about biases? Mascha, do you identify bias in the tools’ applications?

Mascha: In our previous research, we have intensively dealt with the detection and reduction of such biases in word embeddings and language models. We were able to show that society’s stereotypes are present in these models. However, these methods require access to the models, which isn’t possible with commercial applications. With such applications, we make an input and look at the outputs. Based on such methods, we investigated two software products in this project.

Can bias in software be identified by testing different scenarios?

Mascha: In the project, we first looked at typical HR software used to extract information from documents, which can then be used for matchmaking, for example with job offerings. We examined whether this works equally well for application documents from men and women. Although we couldn’t find any significant differences here, this must be treated with caution, and further tests should be conducted with larger datasets, as well as with the matchmaking components themselves.

The situation was different with our second trial using ChatGPT. A bias was detected by our tests. The results will be published in a scientific article in the coming months. There’s still much to do, both technically and in raising user awareness.

Are active sourcers sensitized to the topic of biases? Do they question the tools’ suggestions, Caroline?

Caroline: We asked active sourcers if they notice anything in the generated suggestions. All sourcers are aware that they can only reach people who are registered on social networks, thereby excluding all others. Most notice that male profiles are found more often, which is why 60% of direct approaches are targeted at men and only 40% at women. Therefore, most sourcers wonder why male profiles are “more likely” to be discovered. Sourcers suspect different keyword usage in profiles as a possible cause.

Depending on the chosen search attributes, gender-specific biases can develop in the suggestions in tools like Visage. This underscores the importance of training HR personnel to recognize potential biases in the selection process – regardless of the tool being used.

What’s next?

Mascha: We will continue our tests and particularly investigate how the HR industry can responsibly benefit from new technologies. We support companies in defining suitable processes to ensure that humans remain the focus and issues like bias can be taken into account. The topic is also being investigated in the Horizon Europe project BIAS (biasproject.eu), in which the BFH is leading the technical work package.