Chapter 2
Emerging technologies and their limits
Distinguishing automation, prediction and generation to choose proportionate applications.
2.1 Machine learning
Machine learning fits models using examples. In Human Resources it can support document classification, skills searches or the identification of group patterns. Its usefulness depends on the quality of the objective, the representativeness of the data and evaluation independent of training.
A prediction does not itself identify the causes of an outcome. When the context changes, learned relationships may become unstable. The model must be compared with simple alternatives to determine whether the improvement justifies the additional cost, complexity and risks.
2.2 Language and generative assistants
Language models can help draft job descriptions, summarize policies or suggest training materials. Their responses must be checked before communicating employment conditions or interpreting rules. A fluent explanation does not demonstrate that a source exists or that the conclusion is correct.
An internal assistant must limit its sources, indicate uncertainty and refer sensitive matters to professionals. Retrieved documents may also contain malicious instructions. Testing should include ambiguous questions, conflicting policies and requests that exceed the user's permissions.
2.3 Process automation
Rule-based automation can handle repetitive tasks without artificial intelligence. Processing a complete application or checking a required field are examples of relatively bounded operations. Choosing a simple solution facilitates testing and reduces dependence on components that are difficult to explain.
Every automated workflow needs to handle exceptions and prevent errors from recurring at scale. There must be a review queue, a responsible person and a way to stop execution. Improvements are measured across the entire process, including the work needed to resolve incidents.
2.4 Emotions, biometrics and surveillance
It must not be assumed that voice, facial appearance or microexpressions support valid inferences about honesty, personality or job suitability. An apparent correlation may reflect context, disability, language or recording conditions. These uses require particularly rigorous scientific and legal evaluation.
Within the European framework, the Artificial Intelligence Act prohibits certain inferences about emotions in the workplace, with narrowly defined exceptions for medical or safety reasons. This is not a general authorization to monitor well-being. The current text and official guidance applicable to the case must be consulted [R1].
2.5 Proportionate technology selection
Selection begins with a verifiable need. Alternatives, required data, accessibility, maintenance costs and the possibility of withdrawal are then examined. A tool is not preferable simply because it is new. Some needs are better addressed through organizational change or training.
The analysis must include who benefits and who bears the risk. A solution that saves departmental time may create barriers for applicants. The decision must incorporate these perspectives and describe limits of use in understandable terms.
Questionnaires
Chapter knowledge & Human Resources dilemmas
Dilemmas receive no moral or employability score and are not used to assess job candidates.