Recently, I wrote an article about the various ways to ensure the security of AI applications. I was curious to see how companies deal with this in practice. That’s why I went to visit Harver, a provider of AI-powered recruitment software, which was recently featured by WIRED magazine as one of Europe’s ‘hottest startups’.
Harver’s software maps out the skills and personality of a job applicant. An algorithm then filters out the most suitable candidates for the vacancy using machine learning technology and data analysis. At Harver HQ, overlooking the IJ river in Amsterdam, I asked CEO Barend Raaff how they deal with the issues around the security of systems, algorithms, and AI.
In a recent industry survey conducted by Tata Consultancy Services, the authors state that companies should take the following two measures in order to be able to use artificial intelligence (AI) in the future in a sustainable and consistent manner:
• Secure the system against hacking;
• Develop a system that makes good, reliable and safe decisions.
How do you deal with these two elements at Harver?
“The first point you mention, the protection of systems against hacking, in our opinion, does not only concern AI or machine learning but should have an impact on the entire organization. What’s important here, is that you strive for the right certification and that cybersecurity is a priority within the entire organization. This applies to every department and everything we do. We are talking about policy on storing data, sharing information among colleagues and even the layout of the office. The underlying idea, which is to make the company as safe as possible from a cybersecurity point of view, needs to be omnipresent. We don’t think of security as an incidental topic but as an ongoing process. We are constantly working to strengthen not only our software, but also our processes, communication, and security policy. This is a key point on our agenda that we are constantly working on. Developments in cybersecurity are fast, it’s thus important that the entire organization stays up to date and meets the latest industry standards.
The second point that you quote is very interesting. Developing a system that presents reliable data is of the utmost importance, especially within the recruitment sector. It starts with the way we collect candidate data. With the help of this data, a recruiter will be able to make a decision in the near future. Therefore it is crucial to make sure the data is valid. This is why we only use scientifically validated assessments to collect candidate data at Harver. It ensures that the input is reliable. In recruitment, you hear more and more about scraping data from social media, or automatically scanning resumes in search of certain keywords that are supposed to predict the potential of a candidate. We do not believe in that approach because you miss the context and lose control over the validity of the data.
Once we’ve collected the data, we find it important that the data is initially only used as ‘decision support’. What this means: the data only helps the recruiter to make choices, the system does not work autonomously. So at Harver, we do not take the final decision, we give data-driven advice. On this basis, a recruiter can decide for themselves whether they follow the advice or not.
A third element is what we call ‘performance feedback’. In other words: to make a system smarter, you have to know whether the decisions you have suggested are accurate or not. If an applicant goes through our systems, we give our customers the opportunity to follow these people during their employment at that company. This means that every three, six, and nine months we receive data on how an applicant is doing. Based on this data, we will see whether the decision advice we have given is accurate or not. With this information, we are able to make the system smarter and can continuously improve our advising.
The fourth element is aimed at protecting vulnerable groups and ensuring that ratios do not get out of balance. That’s why we monitor ratios such as man / woman, age, and other characteristics that can indicate whether or not an imbalance occurs and that can adjust the system if necessary.”
AI Now, a non-profit organization that advocates algorithmic freedom, works with a simple principle: if the designers / developers of an algorithm cannot explain a decision, you should not use the application. What stance does Harver have on this?
“The outcome of an algorithm becomes unpredictable if you lose sight of the basis on which you train your data. A good example coming from our world: you have parties that say – our algorithm copies the decision of a recruiter. But if this recruiter rejects five hundred out of a thousand people because of ‘unconscious bias, this bias will also be projected into the algorithm. The name says it all, unconscious bias comes from the subconscious and is impossible to identify. Explaining algorithm decisions then becomes a tricky story.
Hence we decide for ourselves which assessments will and will not be included in the final advice that we present to the recruiter. The system cannot take into consideration the data that we have not determined as relevant in advance. We do not look at names or addresses or even resumes. We only use validated assessment data in combination with the opinion of the recruiter.”
CEO of Harver Barend Raaff.
How do you design algorithms? Who is working on this?
“The team that works on the algorithms and assessment benchmarks at Harver consists of people who all have a background in work psychology and / or data science. A manual test is always added. We can manage that well ourselves because the dataset that we work with is still available.
When we start a new project with a customer, we combine our own expertise with the knowledge of the customer, that is twofold. Sometimes, the project is based on data from the past. In the call center industry, for example, we have so many customers that we can learn certain things from the past. Different industries and organizations are, of course, very diverse, so it is always good to involve the customer as much as possible in the process. This is how you get a system that is customized to what the customer needs.”
You work with several large clients. Do you see an effect with regard to the accumulation of data?
“What is exciting at Harver is the ‘performance feedback loop’. In low-entry jobs, for example, early turnover is a major problem. People are hired and leave again after twelve weeks. At Harver, after one year, we can see how many of these people are now successful and still employed. With reference to this feedback data, we can determine which of the data points we have collected at the beginning predict future success best. Based on this, the system can be adjusted and you can ultimately optimize the recruitment process. This feature is unique within the recruitment industry. That learning element, which is very common in online marketing, is something that’s very new within recruitment. It’s all about testing and experimenting: what works well and what doesn’t. And based on what works, you can improve.”
Where do you see Harver in the future? What is your vision?
“What we aim for is to achieve two things. On the one hand, if an applicant is linked to an employer, we want to be able to tell with the highest possible degree of certainty whether it will be a good match. On the other hand, if there is no match, we want to give the applicant an insight into a possible alternative. So let’s say, you apply for a job with a company that uses Harver and you are not accepted there. Then we want to show you jobs where the chance of employment for someone with your profile is a lot higher.
What you often see is that if you are not hired somewhere, the job application process is a dead end, resulting in a lot of frustration. We want to change this, but this happens one step at a time. Five years ago, the applicant received an email: “Unfortunately, you were not selected”. The first version of our system added a personal score report. What we already do now is indicate what a candidate’s good qualities are so that they can consider that for their next application. The next step is that we can really help someone get started and give more options. “You did not get the job here, but have a look at this alternative.” If we achieve both things, we’ve solved a big problem.”