The sensationalistic narratives surrounding Artificial Intelligence have a significant influence on how people make sense of and use the technology. These narratives often create unrealistic expectations about AI's capabilities that afect how people use it [ 1]. Adopting the concept of algorithmic imaginary, which refers to the way in which people perceive and experience algorithms in their interaction with them [2], I investigate how HR practitioners perceive and experience AI for hiring. With AI being increasingly used in recruitment processes, a better understanding of HR practitioners' interaction with the technology is important when considering matters of bias and fairness. This short paper is based on 8 semi-structured interviews with HR practitioners and recruiters in Italy with experience in using AI for their work and the observation of a selection process with an Italian head-hunting company using their own AI-powered system. From this data emerged how HR practitioners' perception of algorithms as neutral shapes an understanding of bias as something that can be overcome through training of the system. This view, however, reproduces a simplistic idea of bias which does not account for its intersectional and complex entanglement with power structures. Excessive trust in AI's neutrality should be replaced with more critical engagement with AI outcomes.
Artificial Intelligence has a strong link to science fiction and used as a narrative means to create
highly evocative and polarized narratives. These narratives have been shown to influence people’s
understanding of the technology [
In this paper, I will use the case of algorithms used for recruitment in human resources (HR) to
explore how practitioners’ algorithmic imaginary impacts AI fairness in hiring. In the last few years, the
human resources sector has known an increasing ofer of AI-powered tools that can rank or evaluate
CVs and show recommended candidates. On one side, these tools have been found to raise concerns
regarding bias, power imbalances, and fairness of the selection process [
This paper is based on 8 interviews with HR practitioners using algorithms for recruitment, and
CEUR
ceur-ws.org around 12 hours of observation with a head-hunting company using an AI-powered system for search and selection. Through analysis of interviews transcripts and observations notes, in this paper, I aim at presenting the main ideas behind the algorithmic imaginary shared by HR practitioners and how these might impact fairness in recruitment. First, considering the influence of science fiction on the perception of AI, I will present existing literature on the role of AI imaginaries and the algorithmic imaginary in shaping people’s use of AI. Secondly, I will present the specificity of the case of AI for recruitment focusing on the discussion about bias and fairness in AI when used for this task. Then, after explaining the method, I will present the algorithmic imaginary of recruiters that I identified in the data. Lastly, I will reflect on this imaginary which is strongly influenced by popular narratives and marketing narratives but also presents some interesting ambiguities that emerged when looking at people’s use of algorithms. Through this analysis, I aim at proposing some venues for future research and attention in the field of AI fairness for recruitment.
Thinking machines have been a classic element of science fiction since the beginning of the 20th century
[
How people make sense of and perceive the workings of AIs can be understood as an algorithmic
imaginary [
AI is being increasingly used in human resources (HR) for selection and recruitment. The platform LinkedIn, which makes use of an algorithmic recommendation system, has become an indispensable platform for recruiters in any sector, meaning that recruitment has employed some form of digital automation for a very long time [15]. Today, an increasing number of AI tools are being proposed on the market to perform several diferent tasks within the recruitment process. One of the most common uses pertains to the screening, ranking, or summarizing of candidates’ CVs or cover letters, providing an evaluation based on previous hiring decisions, the job requirements, or simply making a summary for ease of manual screening [16]. In diferent ways, these AI applications propose an analysis of candidates to recruiters which aims at supporting them in the initial screening phases and making recruitment faster.
The ethical implications of using these tools are a matter of great discussion. On one side, many
companies creating these systems adopt as one of their selling points the idea that AI decision-making
will reduce human bias ([
In short, perceptions of AI are translated into use can thus have an impact on fairness when AI is being used for decision-making. Mainly, the influence of fiction on people’s understanding of AI can prompt people to place excessive trust in the technology and discard their own agency in interpreting, modifying, or questioning AI results. Placing excessive trust in artificial intelligence and not questioning its outcome could especially eschew issues of bias and discrimination which require critical engagement with AI results. Facebook’s algorithm has given us many examples of the negative consequences of not being aware of algorithm workings on society, e.g. not being aware of the partial picture the Facebook news feed shows us can lead to the “bubble efect” and influence political views [ 19]. Similarly negative consequences have already emerged in the use of AI for hiring [20], which is why critical engagement is fundamental to ensure fairness. In this paper, I will investigate how HR practitioners experience and perceive algorithms in their everyday lives and reflect on the consequences this has on fairness in AI for recruitment.
This paper is based on semi-structured interviews and observations with recruiters and HR practitioners in Italy. This data collection was done as part of the activities for the European project BIAS, investigating bias in the labour market and empowering the HR and AI communities in mitigating it. Participants were first identified through BIAS events and workshops and then recruited through snowball sampling. The site for observations has been found through online search. In total, I have conducted 26 interviews mainly with HR practitioners and recruiters in Italy. From these 26 interviews, I selected a sub-set of 8 interviews which I deemed most relevant for this study as the interviewees had direct experience of AI systems for screening or ranking CVs. Interviews were conducted both online and in-person and lasted approximately one hour. The interview questionnaire (appendix A) was divided into six parts: work experience and educational background, standard selection process with or without AI, understanding of fairness of the selection process, use of AI for recruitment, perception of AI fairness, and expectations about the future of AI.
To complement this data, I conducted a series of observations during the span of two months for a total of around 12 hours with Tech Hire, a head-hunting company using AI for CV screening. Within Tech Hire, I followed one specific selection process for a sales manager in the cybersecurity sector observing primarily Alberto, the main recruiter, and holding regular check-ins with him where he updated me on the progress of the candidate search.
The interviews were conducted in Italian; thus, they have been transcribed and then translated. The
observations resulted in a series of notes which have also been translated. The interviews and notes
from the observations have been coded in Nvivo through thematic analysis. During the first round of
coding, I identified many quotes relating to perceptions of AI. For example, seeing AI as an assistant, a
tool, or a threat to their job. Within the code “perception” however, I noticed a diference between the
perceptions of interviewees who had never used AI and those who did. The perception of interviewees
who did use AI relied on their experience of use and were often presented as certainties. These quotes
seemed close to the experiences registered by Bucher [
A common theme across interviews is the general insecurity HR practitioners communicate about their knowledge of what AI is and how it works. On two occasions the interviewees asked the interviewer directly for confirmation that the tools they were describing were actually AI, which shows the dificulty in understanding this highly complex technology. In addition, the term AI is often used as a general term to indicate the dominant computational techniques [21] creating further confusion. During the observations as well, the recruiter mentioned that while the company’s co-founder (who programmed the platform) knows how everything works, the recruiter does not have the same overview and technical expertise and there are features he is not sure how they work. This insecurity regarding how the technology actually works is further exacerbated once people experience so-called AI hallucinations , which refer to AI output being erroneous or even fictional [ 22]. Experiencing these hallucinations further create a sense of mystery around the technology and reinforce the feeling of lack of transparency.
In one case this insecurity translated into a reflection about the system’s fairness and a resistance to AI use. Fiona, who tested an AI system for CV screening said: “And yet there are continuous challenges [to AI implementation]. Because how can we tell if this system has taken into account all the characteristics I want and has not discriminated against someone? Based on what is it showing only male and not female candidates for example? So there are still many doubts about the efectiveness and also the fairness of this system.”
Here Fiona reflects on how the lack of transparency of algorithms leads her to not trust the system, considering not only fairness but also efectiveness, comparing fair hiring decisions with good hiring decisions. On the other hand, interviewees who were currently using AI systems for recruitment fought this insecurity through continuous oversight of the system , which they felt they couldn’t trust. For example, Monica, a recruiter using AI for candidate, mentioned that recruiters in her company still check all of the candidates marked red by the AI as they know the system can make mistakes. She then reflects on how sometimes it is clear if the AI made a good ranking but other times it is more ambiguous and they need to do a manual screening of the CV to make sure it has been properly evaluated.
The most recurring theme across interviews referred to the perception of AI relying on data and probability. Maria, HR junior specialist, described it as being “based on science”, like a match-making platform proposing the best possible candidate. As a consequence, interviewees understand that the algorithm works best when given objective and numerical data, which it can easily understand. Weighing the pros and cons of using AI, HR consultant Valeria says: “For objective variables, it can probably be useful, in the sense that if I link some KPIs, or some indicators, to more objective indicators… i.e. how many projects you have managed, what level of satisfaction you have concerning customer service. [...] Or if you are in production, how many rubber pads have you produced, in what timeframe, in what ways.
When there are objective variables artificial intelligence can do well.”
Here Valeria reflects on the ways in which AI can be useful during CV screening. She draws a distinction between parts of the recruitment process that rely on objective data, usually measuring hard skills, and those who rely on subjective data that need to be evaluated by a human. Nevertheless, this distinction is not practically applied as recruiters evaluate hard and soft skills at the same time. Having to incorporate the use of AI with the full process of CV screening thus means that they have to turn their requirements into quantifiable data that can be understood by the algorithm. Nadia for example mentioned that: “the advantage of using artificial intelligence rather than a person is that AI works a lot based on data. To work a lot with data the person must also know how to classify them and give rules that generate clearer results.”
In this quote, Nadia emphasizes the crucial role of users inputting data into the system to actually get good results from it. While there are few examples in the interviews and observation of how recruiters translate requirements into objective data, the ones using AI mention having to learn how to use the system properly and adapt their use to the algorithm. Nadia, for example, complained about having to “be patient” and every day refine the way in which she used the system to screen candidates as the results never fully met her expectations.
Starting for the perception of AI relying on data, interviewees moved on to explaining how the algorithm needs to be trained and guided. The common narrative across the eight interviews is that, now, AI needs to be continuously checked by humans, because of the insecurity it elicits, but with continuous training and feedback loops the system is improving or will improve over time. For example, Matteo said: “Today the algorithm needs to be trained. […] The supplier is training the algorithm. Two years of training have now passed, and it is getting better and better.”
Similarly, Tech Hire employees mentioned training AI as an investment for the future. They take time now to train and give feedback to the system to reap the benefits of it later.
This training was sometimes interpreted also as a customization or personalization, trying to convey their own screening “style” the system. For example, Denise, reflecting on the possibility of using AI for CV screening, says: “For example, when I look at CVs I check if people have worked during university or if they have studied abroad and so on and so forth. I don’t know if I can ask the software to look for [this]. […] The day I decide to use it, I would like to be able to customize it as much as possible.”
Understanding training as a personalization further enhances the need for control over the AI system, which is expected to not just become a good recruiter but a good recruiter for their company, aware of their context and company identity and values.
Imagining the algorithm as a blank slate in which data is inputted, however, leads to perceiving AI mistakes as a consequence of human mistakes in training or in using the technology, either due to bias in training the AI or for lack of knowledge in using it. The problem of bias is seen as inherently human rather than technological. AI is then considered not biased but, if ever, trained on biased data generated by people. Thus, recruiters are weary of the results produced by AI but still hold the ambiguous belief that by training it and making sure it is fed good data it will become better, and thus neutral.
The HR algorithmic imaginary is highly ambiguous and even contradictory in some regards, perceiving AI as both objective but also not fully trusting it. Experiences of encounters with algorithms can help us understand how people made sense of this ambiguity in their use of AI and how diferent perceptions emerged according to how the algorithm made itself known. At times, the interaction between the input given by the user and the AI output based on it was misinterpreted, showing an unrealistic expectation for what AI should do.
Matteo and Monica, while talking about their use of their AI system for candidate screening and ranking, experienced the algorithm’s output as a way for themselves to check their own input and the job description.
“We are the first ones to check, not so much if the algorithm runs well, but if in setting the filters of the job requirements we have missed some important ones. Sometimes we realize that the candidate Mario Rossi ended up in the red CVs and we wonder why and we realize that we have not included that type of degree or those years of experience. Maybe the manager tells you ”Mario Rossi is a good profile for me”, so we go and check his CV because, for example, the manager could have had a recommendation from someone. But Mario Rossi is in the reds, why? Because you [manager] told me you wanted up to five years of experience and he has ten. Let’s include Mario Rossi but also let’s expand the research up to ten years of experience so that in addition to him other candidates are also included.”
Their experience shows a very grounded understanding of AI and its objectivity, which in turn highlights people’s bias in setting job requirements. In this case, AI is perceived as a neutral tool, and this perception leads them to question their own biases and rethink the recruitment process.
A diferent experience comes from the observations with Tech Hire. Alberto often described himself and his colleagues as “pedantic”, always checking the AI ranking and making sure there are no mistakes. He then mentions some examples of AI mistakes which I reported in my notes: “For example, we saw the CV of a suitable candidate, but he lives in Rome. The AI evaluated it 8/10 and set him at a high priority because the job description says that Milan/Turin/Rome are all acceptable locations, but in reality the priority is Milan. Alberto identifies this as an AI mistake. He evaluates the candidate 8 but sets him at a lower priority as he will first contact the ones living in Milan. A similar example had happened earlier in the day. A candidate had been flagged red (not meeting the must-have requirements) by the system because she indicated a level 7 of English and the minimum required was 8. Alberto, however, says that between 7 and 8 there is not much diference especially since it is a self-assessment, and the candidate matches all the other requirements. He again identifies this as an AI mistake and marks the candidate green.”
In this situation, the recruiter explained the diferences between the AI screening and the human screening as AI mistakes, conveying an unrealistic expectation of what AI could do since the system followed the job requirements . These types of information, such as Milan being a priority location and the language requirement not being so strict, are contextual information depending on the specific search. These could be communicated to the AI system but would require more work in the initial setting up phases. The perception, however, is that through continuous training, and what interviewees called “personalization”, this kind of contextual knowledge and HR expertise can be replicated by AI.
To summarize, the HR practitioners’ algorithmic imaginary relied mostly on the view of AI being based on data. Data, understood as objective and numeric data, as opposed to subjective requirements, are at the core of the algorithmic imaginary. The way in which algorithms elaborate and use these data, however, is still occasionally unclear, leading to a sense of insecurity and mistrust towards the algorithm and a need to continuously check its output. While checking the algorithm’s output is sometimes perceived as a burden, like in the case of Nadia mentioned before, it is also perceived as an investment, like in the case of Tech Hire. Recruiters check the algorithm with the intention of training or guiding it through feedback, improvements, or personalization to make it better. The algorithmic imaginary is that of a technology which can eventually become neutral, and thus fair but that for now needs to be monitored.
The presented algorithmic imaginary shows a clear influence from external narratives. First, it
borrows from public policies and popular media narratives about the inevitability of AI [23, 24]. HR
practitioners’ encounters with algorithms today suggest a need for continuous oversight, but do not
shake their belief that AI will be better in the future, even better than people. Secondly, their algorithmic
imaginary is strongly influenced by companies and marketing strategies presenting AI as neutral and
unbiased. This idea is upheld by interviewees as well, who adopt this view of AI as being objective, and
therefore, less biased than people. Since AI is based on data and probability, many assume it naturally
leads to better decisions, thus, mistakes or biases in AI are perceived as solvable through additional
training. However, previous research has already challenged this narrative and highlighted the issue it
could lead to in relation to fairness and bias [
Despite the importance of external imaginaries in influencing recruiters, I identified a tension between
perceptions of AI today and future expectations about AI which coexist within the same imaginary. On
one side, the policy and media narratives and the sensationalistic science fiction depictions of AI support
the hope for AI to become better and continue to improve. This expectation stems from depictions of
AIs as humans, too realistic to be true [
The ambiguity imbued in the algorithmic imaginary and the amount of trust that can be put into
algorithms finds certainty and stability in the adjacent data imaginary. Data, within the understanding
of AI, are not problematized. While interviewees recognize that data can be biased, bias in data is
perceived as a flaw that can be easily fixed by removing markers like gender, race, sexuality or other.
”In this sense, AI-powered hiring tools evoke a specific sociotechnical imaginary centred on the concept
of meritocracy” [
Artificial Intelligence has long been a staple of science fiction movies which propose sensationalistic narratives about intelligent machines. While these science fictions often have the scope of inviting reflections on other societal issues, one of their efects is that of creating false expectations about AI capabilities. These expectations form the algorithmic imaginary which inform use and perception. Through interviews and observations with recruiters and HR practitioners in Italy, I examine the algorithmic imaginary of AI in recruitment and how it could impact AI fairness. From the interviews and observation, I identified an algorithmic imaginary relying on the understanding of AI as something that works with data. There is high awareness of the fact that the algorithm needs to be trained, which interviewees think will make it better over time. While this algorithmic imaginary gives people agency to experiment with input, it also conveys this idea of algorithmic bias as a barrier to achieving otherwise neutral AI decisions about hiring. My findings show how the underlying assumption about AI is highly influenced by external sources: science fiction unrealistic depictions of AI and company’s marketing of AI as intrinsically neutral and unbiased. Despite these influences, however, recruiters’ algorithmic imaginary presents an internal ambiguity between the current experience of AI as flawed and opaque and the future expectation of human-like, neutral AIs. The algorithmic imaginary thus relates to both the present and future of AI and recruiters perceive their role as trainers of AI as a temporary step for AI development. Within this ambiguous algorithmic imaginary, I also recognize a strong influence from the adjacent data imaginary that presents data as trustworthy and neutral. However, the intrinsic situatedness and complexity of data and how it can also reflect unequal power structures is overlooked. The way in which the data imaginary is inherited and incorporated into the algorithmic imaginary and its impact on understandings of debiasing could be an interesting avenue for further research.
This work received financial support by the European Union’s Horizon Europe Research and Innovation program as part of the project “BIAS: Mitigating Diversity Biases of AI in the Labor Market” (grant agreement No. 101070468).
A. Semi-structured interview guide 1. What is your background? 2. What is your current position? 3. Could you explain to me step by step a typical recruitment process at your company? 4. Thinking about the work context and personnel selection activities in particular, what is your interpretation of the concept of fairness and equality? 5. Are measures implemented at your company to mitigate discrimination in the workplace and in the selection phases? Do you think they work? 6. Do you use or have you ever used an AI system for any phase of the recruitment process? 7. If yes, could you explain to me how it works? 8. Do you know how the data on which the AI system is based is collected? 9. Does the AI system play a role in the final hiring decision ? 10. What is your opinion on the introduction of AI in the workplace? 11. Thinking about the role of HR, what tasks do you think AI can replace and what tasks can only be done by humans? 12. What do you think is the impact of AI on fairness and equality in recruitment practices? 13. How do you think AI can be best exploited in personnel selection and management activities (especially considering fairness)? 14. What are your concerns related to the introduction of AI in the workplace? And your hopes? 15. What do you think the role of AI could be in 20 years? 16. Would you like to make some final comments or reflections?