Before using AI to screen job applicants, ask these questions
Before deploying AI to screen job applicants, employers must address accommodation, bias testing, vendor evidence, appeal processes, and human review. Responsibility cannot be outsourced to a vendor.
Why AI screening demands your attention
Using artificial intelligence to screen job applicants can save time and surface candidates a human might overlook. But it also introduces risks that, left unchecked, can lead to unlawful discrimination and costly legal exposure. In Hong Kong, where employers increasingly adopt automated hiring tools, understanding these risks is essential. Before you deploy any AI screening system, you need to ask hard questions about accommodation, bias, vendor evidence, appeal rights, human review, and where ultimate responsibility lies.
Does the tool accommodate disabilities?
Under the U.S. Americans with Disabilities Act (ADA), which sets a benchmark for many multinational employers, hiring technologies must not discriminate against people with disabilities. The U.S. Equal Employment Opportunity Commission (EEOC) and the Department of Justice have both issued guidance explaining that AI tools can violate the ADA if they screen out qualified individuals with disabilities or fail to provide reasonable accommodations.
For example, a video-interview tool that analyses facial expressions may disadvantage applicants with autism or certain mobility disabilities. A resume parser that penalises gaps in employment history may unfairly filter out people with chronic health conditions. Ask your vendor: What steps have you taken to ensure the tool does not discriminate against people with disabilities? If the vendor cannot provide a clear answer, that is a red flag.
You must also have a process for applicants to request accommodations. The ADA requires employers to provide reasonable accommodations unless doing so causes undue hardship. This might mean allowing an applicant to take a written test instead of a video interview, or providing extra time for a cognitive assessment. Your AI screening system must support such accommodations, not block them.
Have you tested for bias?
Bias testing is not optional. The EEOC’s Artificial Intelligence and Algorithmic Fairness Initiative specifically examines whether AI tools result in disparate impact based on race, sex, or disability. The EU AI Act, which applies to any organisation placing AI systems on the European market or using them to evaluate people, requires high-risk systems to undergo conformity assessments that include bias testing.
Ask your vendor for documented bias-testing results. What data was used to train the model? Does that data reflect the actual applicant pool for your roles? If the model was trained on historical data from a workforce that lacked diversity, it may perpetuate past exclusions. You should also run your own tests using synthetic or anonymised data that represents your local context, including Hong Kong’s multilingual and multicultural workforce. A transparent protocol: compare pass rates across demographic groups using a sample of at least 200 records per group. If any group’s pass rate is significantly lower, investigate the cause before using the tool.
Remember that bias testing must be ongoing. Models drift as new data comes in, so schedule regular audits at least annually or whenever the tool is updated.
What evidence does the vendor provide?
Vendors often claim their AI is fair, accurate, and compliant. You need evidence, not promises. Ask for third-party audit reports, validation studies, and documentation of the model’s limitations. The EU AI Act requires providers of high-risk AI systems to maintain technical documentation that includes a description of the system’s intended purpose, the data used for training, and the accuracy metrics. If your vendor cannot supply this, consider whether they are ready for regulated markets.
Also ask about the vendor’s own grievance and appeal process. If an applicant challenges a decision, can the vendor explain why the AI gave a particular score? Many AI systems are black boxes, meaning even the vendor cannot fully explain their outputs. This is a serious problem because you may need to justify your hiring decisions to regulators or in court.
Do not accept vague assurances. Request a written statement of compliance with relevant laws, including the ADA, the EU AI Act, and any local Hong Kong anti-discrimination ordinances. If the vendor hesitates, that is a warning sign.
Can applicants appeal AI decisions?
Applicants need a clear way to challenge decisions made or influenced by AI. The EEOC guidance emphasises that employers must provide reasonable accommodations, which includes allowing applicants to explain or correct information that the AI may have misinterpreted. The EU AI Act grants individuals the right to request an explanation of a high-risk AI system’s decision that significantly affects them.
Your process should include: a point of contact for appeals, a timeline for response (e.g., within 10 business days), and a mechanism for human review of the AI’s recommendation. Publish this process on your careers page so applicants know their rights. If an applicant says the AI misjudged their video interview because of a speech impediment, you must have a human ready to re-evaluate their application.
Is there meaningful human review?
Human review is not a rubber stamp. A human reviewer must have the authority to override the AI’s decision, and they must be trained to recognise potential bias and accommodation needs. The EU AI Act requires that high-risk systems be designed to allow effective human oversight. This means the human should understand the system’s capabilities and limitations, and be able to intervene when the system produces errors or biased outcomes.
In practice, this could mean: a recruiter reviews the top 10% of candidates flagged by the AI, or a hiring manager manually checks all applicants who request an accommodation. Do not let the AI become the sole gatekeeper. The human must have access to the full application, not just the AI’s score.
Document every instance where a human overrides the AI. This creates an audit trail that can help you identify patterns of bias or system failures.
Why you cannot outsource responsibility
Even if you buy an AI tool from a reputable vendor, you remain legally responsible for its outcomes. The EEOC and DOJ guidance is clear: an employer who uses a third-party hiring technology can still be held liable for discrimination. The EU AI Act places obligations on both the provider and the deployer of high-risk systems. In Hong Kong, the Equal Opportunities Commission may investigate complaints against employers who use discriminatory tools, regardless of who built them.
Outsourcing the technology does not outsource the risk. You must conduct your own due diligence, test for bias, provide accommodations, and maintain human oversight. If a vendor promises a fully automated, hands-off solution, treat that claim with scepticism.
Responsibility for fair hiring rests with your organisation. AI can be a useful assistant, but it cannot replace your duty to treat every applicant with dignity and respect.
Practical checklist before you deploy AI screening
- Confirm the vendor can document compliance with relevant anti-discrimination laws.
- Obtain bias-testing results and run your own tests using local data.
- Establish a process for applicants to request accommodations.
- Create a clear appeal mechanism with a defined timeline.
- Ensure human reviewers have authority to override AI decisions and are trained to spot bias.
- Schedule regular audits of the tool’s performance and impact.
- Document all override decisions and accommodation requests.
Taking these steps will not eliminate all risk, but it will significantly reduce the chance of unlawful discrimination and help you build a fairer, more transparent hiring process.
Sources and further reading
- EEOC: Artificial Intelligence and the ADA
- U.S. DOJ/ADA.gov: AI and disability discrimination in hiring
- EU AI Act text
Guidance and regulations can change; check the current primary sources before applying them to a consequential workplace decision.