The world's first CV was Leonardo da Vinci's, written in 1482 to win work from the Regent of Milan. It worked. More than 500 years later, we are still hiring off the same document.
What has changed is what reads it. AI now sits across sourcing, screening and shortlisting, promising faster, cheaper, more objective hiring.
Here is the limit, up front. AI cannot actually take bias out of CV screening. The CV is the main place the bias sits, and the technology alone will not fix that.
That is not a reason to keep AI out of your hiring. It is a reason to be deliberate about where you put it, and what you ask it to do.
Where AI Earns Its Place
Start before the CV pile. A skills-first search matches people on the skills a job needs, rather than proxies like degrees, job titles or past employers. It widens your candidate pool. Why? Because it takes out the biases you are not even aware you are using to filter.
In screening itself, two uses hold up, and they have the same thing in common.
AI can sit in an interview with the interviewer, like a notetaker, and assess answers against evidence rather than impression. It does not warm to someone. It does not decide in the first ten seconds based on confidence. But that only works if every candidate gets the same questions, in the same order, planned in advance and tied to the job criteria. When you standardise the interview, the AI applies it consistently. Change the questions between candidates and you have automated an unfair comparison.
The same holds for benchmarking and summarising: same questions, same selection criteria in the prompt, same treatment for everyone.
In both cases the AI is applied to a process that you built, and one that has to be made fair. It is not being asked to work out what fair means.
Where Bias Actually Enters
In 2014 Amazon built a CV screening model trained on ten years of its own hiring history. It learned to penalise female candidates and CVs containing the word "women's", including "women's chess club captain", and marked down graduates of two all-women colleges. Why?
The model was not broken. It did what it was built to do: mimic past behaviour. There are few women in tech and fewer in tech leadership, so it reproduced that. A system trained on the past cannot produce a different future.
Yes, Amazon scrapped it.
In the last year I have worked with two clients who know these risks well. An international law firm reached the same conclusion before deploying, and decided the legal risk was not worth it. An Australian disability services not-for-profit hiring high volumes of support workers uses AI for candidate pooling, but only after a human first pass. Same technology, different position in the process, and the position is what makes it defensible.
Underneath all three: discrepancies go in, so discrepancies come out. A CV is a record of not only skills and experience but personal information that naturally has us forming a "who" in our minds. A profile we already unconsciously want to see, and screen out harshly and quickly in 6 seconds if they do not fit that image. What does that tell us about capability? Nothing. If you are measuring CVs, you are measuring a document riddled with bias triggers, built on that same 500-year-old format, and it does not give you the context your brain is looking for. Neither a machine nor a person can assess context they do not have, and a gap on a page does not say why it is there.
Bias also enters before the tool ever sees a candidate. Testing of commercial facial analysis systems found error rates under 1 per cent for lighter-skinned men and up to nearly 35 per cent for darker-skinned women. Nobody designed that. It was built and tested on data that under-represented those users, and the gap shipped. Recruitment tools are built the same way.
Four Questions to Ask Before You Buy or Build Your Recruitment AI
- What data are you capturing, and why? If nobody can answer the why for a field, that field is a liability.
- Who designed it, and against what? What did the training data look like, and what are the error rates across different groups? A vendor who has measured this will tell you. A vendor who has not is telling you something too.
- Where does it sit in the process? Screening people out at the front is a different risk from organising a pool a person has already reviewed.
- What are you asking it to analyse? It should feed back to core job assessment. Anything it evaluates that is not the job is something you will have to justify later.
What This Is Really About
You do not need to decide whether AI belongs in recruitment. It is already in your stack.
The decision is narrower: which part of your process are you handing over, and have you made that part fair before you automate it? Get that order right and the technology compounds good practice. Get it backwards and it industrialises whatever you were already doing.
What would you put your own hiring process through first?
Not sure? That is OK. There are many ways to bake a cake. Get in touch with me and I can help you build the fair processes that match your audience.
Samantha Philpot is the Director of Find Your People. She has facilitated since 2010 and spent ten years in recruitment and HR, shaping a difference in employment systems.