AI Recruitment

What AI recruitment is

AI recruitment, also called AI in recruitment (التوظيف بالذكاء الاصطناعي), is the use of tools built on models that learn from past data at the stages of hiring: writing the job advertisement, searching for candidates, sorting and ranking applications, scheduling interviews, corresponding with applicants and assessing their answers.

The label gathers different things under a single name. Confusion in this area can start when an organisation buys one kind of tool in the belief that it has bought the other.

The question to ask of any AI recruitment tool: what it takes over

One distinction applies to every recruitment tool, and not only to automated interviews:

• A tool that takes over work or time. It drafts an advertisement, books an appointment, sends a reply, or extracts details from a CV into standard fields. The decision stays where it was.
• A tool that takes over judgement. It gives a candidate a score or a rank based on the content of their answer or their CV. It changes who decides, not how long the decision takes.

Both kinds are sold under the same name. So the vendor is asked which one is on offer before anything is bought, and asked separately about each stage, because a single product may combine the two.

Where AI recruitment enters the hiring process

• Before applications arrive: drafting the job description, suggesting where to advertise, and searching for matching profiles.
• When applications arrive: extracting fields from CVs and standardising them inside an applicant tracking system, and detecting duplicate records.
• At screening: ranking applicants by how closely they match written requirements.
• At assessment: scoring a recorded answer or a test.
• In correspondence and scheduling: replies, reminders and matching interview times.

How AI recruitment differs from the tools filed beside it

Several things are filed under this label that sit outside it. The axis that separates them is whether the tool learns from previous cases or applies something a person can read line by line. Telling them apart is not a matter of wording: each is bought at a different price and examined in a different way.

• Automation by rules. A condition written by the employer that excludes everyone without a driving licence is not learning from data. It is the execution of a rule that a person set, can read and can change. It carries less risk and is easier to review, and a tool marketed as AI may be doing exactly this.
• An applicant tracking system. It is the place where applications are stored and where they move from stage to stage. A ranking tool may be added to it, or may not. Having the system does not mean the organisation ranks candidates automatically, and not having it does not mean the organisation uses no such tool.
• An automated interview. It is one stage of the process, not the whole of it, and it can be run as a recording that people watch without any automatic score being given.
• Tests marked automatically before hiring. They are marked against a known answer key, which is marking rather than judgement. The difference is that an answer key can be reviewed item by item, while a model that learns from previous cases cannot be reviewed in that way.

Where the decision in AI recruitment stays human

The practice advised for automated screening in an applicant tracking system carries over to AI recruitment: the output is used to rank applicants, not to reject them outright.

A more precise point follows from the way automated interviews work. Automation moves the point of examination from the interviewer to the criterion; it does not remove it. A tool can apply a criterion more consistently than a person does, but consistency in applying a criterion says nothing about whether the criterion is right. A wrong criterion applied consistently produces a systematic error in place of a scattered one.

That is why what the tool learns from has to be examined. A model trained on the organisation’s past hiring decisions reproduces whatever those decisions contain, including implicit bias (التحيز الضمني), a leaning of which the person holding it is unaware. What a person is unaware of cannot be kept out of training data by good intentions.

A worked calculation of the AI recruitment threshold

An automatic ranking is not used on its own; it is used with a threshold, meaning the number of files from the top of the list that move on to the next stage. That number is the real decision, and it can be set without anyone calculating what it costs. The example below is entirely hypothetical. Its figures are chosen to show the relationship between the threshold and what is lost beneath it, and they describe the performance of no particular tool.

An organisation receives 1,200 applications for a single vacancy. Assume that among them are 24 candidates whom the hiring team would have invited to interview had it read every application itself:

• The team takes the top 10% of the tool’s ranking, which is 120 files, and 18 of the 24 fall within them. That captures 18 of the 24, which is 75%. The 6 left below the threshold will never be known to anyone, because excluded files are not reviewed.
• Among the files moved forward, the share of useful ones is 18 divided by 120, which is 15%. In other words, 102 of the files moved forward will be read and then set aside.
• If the threshold is raised to the top 20%, which is 240 files, the share captured rises to 22 of 24, about 92%, at the price of doubling the manual review load from 120 files to 240.

Two things follow from these figures. The first is that greater accuracy is bought with time: there is no option without a cost, only a point chosen along the line. The second is that the errors in the two directions are not equally visible. The 102 files moved forward without result are seen by the team, which complains about them; the 6 that fell away are never seen by anyone. So judgement in practice can lean towards narrowing the threshold, and narrowing it improves what is seen while worsening what is not. The threshold is examined alongside the selection ratio, which measures how much choice the organisation had when it decided.

The figures above are assumed for the purpose of the arithmetic. We found no published limit on which to base a share of automatic exclusion that could be called acceptable.

What AI recruitment cannot supply: the written criterion

An AI recruitment tool needs a written criterion to work on, and it does not create one. If the requirements of the role are vague, the tool produces a ranking that looks precise on a vague foundation.

The criterion comes from the structured interview: questions written before any candidate is seen, one order for everyone, and a rating scale that is described in advance. An organisation that did not have these before the tool will not have them after it. Skills based hiring points the same way, because a decision built on a described skill is easier to examine than one built on a job title or a place of graduation.

Situations in which an AI recruitment tool is weaker

General discussion of these tools assumes a vacancy that recurs and attracts a large number of applications, the setting of high volume hiring. Outside that description the calculation changes:

• A rare or specialist vacancy. A tool that ranks on the basis of previous cases needs previous cases, and an organisation that has filled the role three times in five years has nothing for it to learn from. The resulting ranking rests on matching words, not on learning from outcomes.
• Moving a model from one role to another. A model tuned on a sales role does not carry over to an accounting role by changing the job title, because what it learned is tied to the role it learned from. The risk arises when the tool is bought once and then used for every vacancy.
• The internal applicant. They are an employee applying for an internal vacancy, with a performance record held by the organisation itself. Placing that person in a ranking built from a written file sets aside the more precise evidence and puts weaker evidence in its place.
• Differences in language or in how CVs are written. A tool tuned on files written in one format can treat anything outside that format as a shortfall in matching rather than as a difference in writing. The candidate who loses out is not the weakest one but the one furthest from the template.
• A vacancy with one decisive requirement. If the deciding condition is single and known, a written rule is enough for it. Adding a ranking tool adds a layer without adding a decision.

AI recruitment and candidates’ personal data

Candidate files are personal data about people who are not employees of the organisation. Handling them is subject to the Personal Data Protection Law (نظام حماية البيانات الشخصية), which requires, among other things, that the purpose of collection be specified, that what is collected be limited to the minimum needed for that purpose, and that what is collected be protected. The same question arises in cybervetting, in which an organisation gathers what a candidate has published online.

The point to watch is that AI recruitment tools pull in the opposite direction: the more they are given, the better their output, so collecting more than is needed becomes a technical demand that runs into a legal limit. Entering a candidate’s data into a tool operated by another party is also processing of that data, not merely the transfer of a file.

The same tension arises over time. An organisation may want to keep files for a long period, so that the model can learn from them and so that a pool of candidates is available to return to, while the requirement to specify a purpose means that everything kept needs a purpose that still stands. These two aims pull against each other, and they are settled by a written decision on how long files are kept, not by leaving files wherever they happen to sit. In the sources we reviewed, we found no figure for that period on which to base a number, so none is given.

The definition of AI recruitment does not address any specific rule on decisions based on automated processing, or any right of a candidate to have such a decision explained or to object to it. Those questions have their own sources: the Personal Data Protection Law, its Implementing Regulations and the competent authority. That the definition leaves them aside does not mean that no such rule exists.

What undermines the use of AI recruitment

• An undeclared criterion inside the score. Fluency in front of a camera is one example, where the role does not require it. What is not written into the job description cannot be defended if the organisation is asked why a candidate was excluded.
• Accepting a score without its reason. A tool that gives a number without saying on which item the candidate fell short documents the decision but does not explain it.
• Believing the vendor’s claim without a trial. The practical test is to run through the tool a set of files whose outcome has already been decided, compare its ranking with the conclusions people reached, and examine the points of disagreement one by one. Each product is examined in its own right against the organisation’s written criterion; the descriptions above concern how such tools work, not a verdict on any product or category of products.
• Removing the alternative route. A weak connection, the lack of a device or a file that will not upload are not selection criteria, so an applicant held back by one of them needs another route to apply. That route is part of the candidate experience.
• Measuring success by speed alone. Speed can improve with other kinds of tool as well. What the organisation wants to know is whether the people actually hired are better. We found no measurement on which to base a view of the effect of these tools on the quality of the people hired, so it cannot be said that they improve it, and it cannot be said that they weaken it.

This is an explanation of the concept and of the statutory provisions cited, not legal advice.

Related terms

• Test Validity
• Predictive Validity
• Work Sample Test
• Cognitive Ability Test
• Boolean Search