Keep It Fair: The Honest Limits Rule for Recruiters
AI in recruiting comes with serious limits. Bias amplification, automated rejection, hallucinated facts, legal exposure, and privacy risks are all real. Every recruiter using AI needs to understand these before AI touches a screening decision.
Using AI to draft job descriptions is one thing. Using AI anywhere near a screening or selection decision is a different category of activity — one that carries real legal risk, real fairness risk, and real consequences for real people.
This lesson covers the limits that every recruiter needs to understand before AI touches the candidate evaluation side of the pipeline.
Bias Is Not Hypothetical
AI models are trained on data. In the context of hiring tools, that data often includes historical hiring decisions — which means historical patterns of who was hired and who wasn't.
If your organization, or the broader labor market from which training data was drawn, historically favored a certain type of candidate — a particular educational background, communication style, employment trajectory — then an AI model trained on that data will replicate those patterns. At scale. Without disclosing that it is doing so.
This is not a theoretical future problem. Documented examples exist of AI-generated job descriptions that used language statistically associated with male candidates. Of AI resume scanners that penalized employment gaps more common among women who took parental leave. Of tools that deprioritized candidates whose names indicated non-Western backgrounds.
The model does not know it is doing this. That is precisely the danger.
Your guardrail: AI assists the process. AI never screens people out. You read every summary, you make every advance or pass call, and you review every piece of AI output for language or framing that could proxy for a protected characteristic.
Automated Rejection Is Different From AI-Assisted Review
There is a meaningful legal distinction — recognized in a growing number of jurisdictions — between using AI as a drafting tool and using AI to make or substantially influence an employment decision.
If you ask AI to rank your candidate slate and then advance only the top tier without reviewing the others, you have effectively used AI to make an adverse employment decision against the candidates who were not advanced. That is an automated employment decision, regardless of whether a human approved the final offer.
New York City's Local Law 144, which went into effect in 2023, requires employers and employment agencies using "automated employment decision tools" in hiring or promotion to conduct annual bias audits by an independent auditor and to notify candidates that such tools are being used. This law defines the category broadly — it is worth understanding whether any tool you use for screening qualifies.
The EU AI Act classifies AI systems used in employment, worker management, and access to self-employment as high-risk. High-risk AI systems in the EU are subject to requirements including human oversight, transparency, and conformity assessments.
These are examples, not an exhaustive list. Regulations in this space are moving quickly. The correct posture is: loop in your legal and compliance team before deploying any AI tool that touches screening or selection.
AI Hallucinates — Including About Candidates
AI language models generate plausible text. Sometimes the plausible text is not accurate.
Ask an AI to summarize a resume and it may confidently state that the candidate has a certification they do not have. It may add a year of experience based on a plausible inference that turns out to be wrong. It may attribute accomplishments from one role to another.
If you use an AI-generated summary as input to a screening decision without reading the source document, you may advance or reject a candidate based on something that never happened.
The practice is straightforward: treat every AI-generated resume summary as a starting point, not a verdict. Read every one. Verify any specific fact — a certification, a company name, a tenure length — against the original resume before it influences your decision.
Candidate Privacy Is Not Optional
When a candidate sends you their resume, they are sharing personal information — often including contact details, employment history, educational background, and sometimes more. They are sharing that information with your company, in the context of an application to work there.
They are not consenting to having that information transmitted to a third-party AI service.
If you paste a candidate's resume into a public AI chat tool — the free tier of ChatGPT, the standard consumer interface of any major model — you are sending that person's personal information to a service whose data handling policies may not align with your organization's privacy commitments, or with applicable law. Under GDPR, this kind of data transfer requires a legal basis. Under CCPA and similar state laws, there are analogous obligations.
Your practice must be one of the following:
- Use an enterprise-tier AI tool that your organization has evaluated and signed a data-processing agreement with
- Anonymize before prompting — remove the candidate's name and contact information before pasting anything into AI, so the model is working with de-identified content
This is not a technicality. It is a basic obligation to the candidates who trusted you with their information.
The Standard to Hold
The test for any AI use in your recruiting workflow is simple: could you explain, to the candidate who was declined, exactly what criteria led to that outcome, and show that each criterion was job-related?
If the answer is no — because the criteria came from an AI ranking you accepted without review, or because the screening included language that proxied for a demographic, or because you advanced the top of an AI-sorted list without reading the rest — then the process does not meet that standard.
AI makes you faster. Your judgment makes the process fair. Both matter. Neither replaces the other.
The capstone at the end of this track asks you to apply this standard to a real role. Read "The Role Hiring Kit" — the next lesson — to see how fairness operates across the full kit of artifacts you will build.
This track is part of the AI Operator's Academy at academy.jeremyknox.ai. For broader context on AI in professional practice, visit jeremyknox.ai.