IS A DOG LYING IN THE SNOW A WOLF?
By Corinne Fernandez, Partner & Industrial & Technology Practices Leader at Progress
Will «machine learning» replace Recruiting and Executive Search experts?
The current (and even recurrent) craze for artificial intelligence which «will soon replace man» and his occupations, affects all professions and therefore ours: the recruitment of Senior Executives.
(Initially) sceptical towards this threat, I was first alerted last year during a «courtesy interview» with a Human Resources Manager in search of a job and who was fascinated by the advances of AI in the recruitment process. He told me that I would «soon be replaced by a chatbot». I brushed off the remark and life continued … perhaps with a little less availability for courtesy interviews….
A few months later, one of our prospective clients did not confirm a search assignment because his company’s Head Office had purchased « machine learning[i] » software for recruitment and wanted to test it on the assignment that was to be entrusted to us. This second far more disconcerting alert pushed me to further pursue the subject with Progress’ research and documentation team. For the record, I am in charge of Progress’ research team, and so, to answer the legitimate questions of this team and those of our young recruits, I decided to evaluate the risk:
Where are we in terms of recruitment? What can machine learning do and what can’t it or won’t it be able to do in the near future?
After attending a few conferences and reading a lot of quality press articles, especially those published by our reference association, AESC, I interviewed some scientists and I would like to share with you an inventory of AI and my extrapolations to human resources. I have voluntarily chosen to start with scientists who don’t work in the field of HR to get an understanding of what is possible in general before focusing on the field concerned.
Is a dog lying in the snow a wolf?
Copyright Franz Marc Liegender Hund im Schnee (1910-1911)
To understand what machine learning can do, let’s take an illustrative example: in a «supervised learning» phase, a computer is shown thousands of images of a wolf, thousands of images of a dog and it ends up successfully distinguishing the wolf from the dog, and thus correctly making a «prediction» when it is shown a new image. But, this worked well until the day when the computer mistakenly identified a dog as a wolf. What happened? In fact, it is difficult to know how neural networks activated by machine learning functions work, similar to a «black box» but following an analysis, it was discovered that the majority of the images of wolves that had been presented were taken in the snow, which statistically or « probabilistically » deceived the machine.
On the other hand, bear in mind that it is sufficient for a child to see less than ten images in general to be able to distinguish an animal.
This example illustrates the power of human intelligence over that of machines and their algorithms, however powerful they may be.
: surveillance systems deployed in Chinese cities prove it. It also makes it possible to scan a great deal of information, process it and make associations.
First interview: Bruno Bouzy, a Professor in Mathematics and IT at the Paris Descartes University… When machines will play Bridge…
Bruno Bouzy works primarily on Artificial Intelligence applications in the gaming field and it is on the basis of his research, that computers are now beating humans at Go, something that was unthinkable a few years ago. Games present particular interest for scientists because with just a few simple rules, one can create a case and submit its solution to the computer. Real life is much more complex to model. Thus, the improvements in the capabilities of machines and « machine learning » have made it possible for the computer to beat man in a game of Go. Nevertheless, the computer is not yet capable of beating humans at bridge because some of the information (in the opponent’s or even the partner’s game) is hidden, unlike what happens on a chessboard or a goban.
Without knowing the field of human resources, which he qualifies as much more «unpredictable» than games; he launches some lines of thought:
- «Machine learning» succeeds in recognising images while text analysis is more challenging,
- In a job interview, the hidden information is more complex than in a bridge game …
Second interview: David Louapre, Scientific Director at Ubisoft and creator of the Youtube channel; «Science Etonnante», which endeavours to render science, more accessible: Human Resources Managers serving machines…
David Louapre stated at the outset that he is a physicist by training and did not work in the field of Human Resources. However, his general knowledge of AI and his ability to transpose his knowledge enables him to deduce that it is primarily in the analysis of the texts in job descriptions and candidates’ CVs and also reconciling the two that AI will be able to apply. In addition, designing templates will require supervised learning in order to save time for HR professionals.
Reminder: in «machine learning», there are two ways of «teaching»:
- «Supervised» (or assisted) learning and «unsupervised» learning. In the first case, the «machine» is fed with hundreds of thousands of images and by informing it what the image represents; neural networks gradually learn to create categories and become capable of making «predictions» on a statistical basis. This is the case with our example of the dog and the wolf.
- On the other hand, in «unsupervised» learning, the machine is fed with a very large amount of data and gradually, on its own, it «finds » «associations » and« links» that allow it to define categories.
In the case of recruitment, suppose that the «machine» is given tens of thousands of CVs of candidates presented for positions and thousands of CVs of candidates who have been recruited successfully on these same positions (it requires a few years to obtain this type of feedback), little by little, the «machine» will find the «correct» matching criteria for these recruitments, at the risk, as we saw in the case of Amazon, of reproducing biases that are harmful to diversity. Once the system is set, if we propose a new job description and thousands of CVs to the machine, it will be able to «predict»/«advise» on those to retain.
To create such a system in unsupervised learning, it would be necessary to have a large number of very patient in-house recruiters to assess the quality of hundreds or thousands of offers made and thus define the right sets of criteria.
To save time, supervised learning can be used, knowing that each company has an explicit or implicit skills reference of its own which should be integrated into the tool before launching the «matching» exercise between a job description and a profile. Yet, this skills’ reference is explained to the «recruiter» in a single conversation if the recruitment is outsourced or handled by the internal recruitment team who know it by heart.
Third interview: Joël Bentolila, CTO and co-founder of Talentsoft: Beware of unrealistically high expectations…or a 3rd AI winter is coming!
Joël Bentolila is the Chief Technology Officer and co-founder of Talentsoft. A real expert in his field, he has nearly 30 years of experience in software engineering management for major software companies. A graduate of the Ecole des Mines in Paris, Joel has been involved in a number of research projects on Artificial Intelligence at Carnegie Mellon University in Pennsylvania.
Joel began the interview by recalling that we have already experienced two «AI winters» and that to fulfil promises, the current spring must avoid the excessive hopes that led to the two previous winters. This is because even in the current craze he still sees «the same fantasies often conveyed by intelligent people» and which led to the previous winters.
In conclusion, he proposes a hierarchy of problems that can be solved by AI:
- Open information games (chess, go),
- Semi open information games (bridge),
- Diagnostics – possibly predictive «for breakdowns» (machines, cars),
- And lastly, human relations, «the ultimate of the «hazardous» domains as Bruno Bouzy would say ».
AI remains a very fertile area of research and development in the Human Resources field, as the startups which propose AI solutions endeavour to meet the HR Managers’ constant need to automate repetitive tasks. It is on this need and promise that they thrive. Talentsoft has conducted its own evaluation of the offers on the market and has found that so far, there are «many pilots and very few «full scope» deployments» and that generally companies that claim to have customers often only have pilots. Talentsoft invests in partnership with start-ups on matching and search software for their ATS as part of the HR Lab they run and which brings together a community of start-ups.
Nonetheless, Joël Bentolila believes that AI will improve the efficiency of the «search» (“sourcing”) for candidate profiles, of «matching» between profiles and positions or training programs and of chatbots for HR. Clearly, AI applies to:
ATS matching between job offers and candidate profiles
Matching in employee management tools:
– Between proposed jobs and employee profiles
– Between available training programs and employee training needs (adaptive
That said, before carrying out this matching, and because text paradoxically contains more «non-verbalised words» and ambiguity than an image, it is necessary to carry out a semantic analysis of the words and to constitute a dictionary which associates the words that could have the same meaning. For example: Sales Representative, Key Account Manager, Sales Engineer, Sales Executive, etc. This dictionary can be constituted by a human or done automatically by a machine, with risks of misinterpretation. Both options are to be considered in a context where trends are constantly creating new terms. We then speak of «semantic analysis» or preliminary «ontology». In Computer Science and Information science, an «ontology» is a structured set of terms and concepts representing the meaning of a field of information, whether by metadata of a namespace, or the elements of a domain of knowledge. Ontology is in itself a data model representative of a set of concepts in a domain as well as relationships between these concepts. It is used to deliberate on the objects of the specified domain. More simply, one can also say that «ontology is to data, what grammar is to language». Source- Wikipedia.
AI will also apply
1) to «searches» with a requisite close to Matching in order to search on networks, in profile or job databases (subject to authorisation) using keywords (as in traditional queries in a database) and then by analogy with a more «predictive» system,
2) through chatbots that clear the ground for candidates by guiding them to the right person or the right offer according to their criteria (geographical location, type of jobs sought …). By the way, Talentsoft has introduced «Bots» on recruitment websites.
A chatbot works in 2 ways:
- Via a decision tree
- Via an analysis of the request using Natural Language Processing (NLP).
In any case, Joel Bentolila concludes that AI can only offer assistance to a human operator but not replace him. It «handles» the simple issues and redirects towards a human as soon as the subject becomes too complex. AI remains a tool to help «increase» human capacities in the service of a professional.
And Executive Search in all this?
In Executive Search, by definition, one does not deal with large numbers of positions but just a few per year since the search concerns members of Executive Committees or Boards of Directors, or even executives suitable to fill these roles.
If, in order to widen the scope of possibilities, one is led to look for profiles on the social networks or official nominations via the press, the potential candidates are very few and can be counted in units or in tens at the most.
Beyond the necessary skills («hard skills») required for a given role (Financial Director, Human Resources Director, Director of Comp & Ben, CTO), one always looks for a /demeanour and interpersonal («soft») skills, a story, professional experiences in different companies (large groups and ETIs, family owned business and American «style» groups…). And lastly, the « (right) fit» during the face-to-face meeting, remains decisive.
How many behavioural tests must we administer to design this model which varies each time in accordance with the competency framework of each company? And above all, how many hired candidates correspond to the initial job specifications? This remark was made by our researchers and is based on an analysis of the profiles of the candidates that were hired compared to the initial criteria. It gives an idea of the «unpredictable» nature of our profession and can be explained by several factors:
- The requirements evolve during the course of the search according to the profiles met: the customers adapt to the market, and sometimes their own situation evolves,
- The face to face meeting is decisive,
- A common frame of criteria for both the decision-maker and the candidate can be key, we are in a context where trust is paramount,
- Potential, motivation and desire are often more decisive than the actual experience itself, and yet it is this experience that is highly accentuated in the job specifications.
At one of its conferences, Yatedo, a leader in the field of AI (together with Pipler, its search engine dedicated to recruitment), indicated that 11% of recruitment failures are due to technical skills and 89% to behavioural and relational issues or to problems in integrating the company. Thus, even if AI saves time by identifying qualified profiles, there will still be a need to work on the «human aspects» because Artificial Intelligence can never replace communication, intuition, persuasion and decision making.
Progress is therefore interested in AI as a tool to increase the scope of possibilities and we would like with this article, to open a dialogue with all those who are interested and who will bring us scientific and objective information….
And no, we will not be replaced by a chatbot 🙂 !
[i] Machine learning can be defined as automated learning (source David Louapre, Science Etonnante).
Machine learning or statistical learning is a field of study in artificial intelligence that is based on statistical approaches to give computers the ability to “learn” from data, that is to say, to improve their performance in solving specific tasks without being explicitly programmed for each one. More broadly, this concerns the design, analysis, development and implementation of such methods.Share