Skip to main content

Our lab was testing constrained LLM architectures when we realized something surprising: by redefining how the AI “thinks,” we could structurally prevent bias — not just detect it.

So we googled and realized it could be useful to HR, and we had Claude AI make some fake JD’s and resumes that it said was loaded with biases and….it worked.

Instead of filtering or training the AI, we have managed to erase demographic and proxy bias (names, schools, ZIP codes) at the cognitive level. It’s pretty cool. Testing shows near-100% blocking of known bias vectors, with full audit trails.

We tried to make a demo video comparing the base model with one with this...firewall...we are calling it. It’s here, if you’re interested.

But...now what? We’re not an HR company. We weren’t even really focusing on this. But it strips away all prestige/race/class/gender biases and evaluates strictly on merit, and that seems pretty important

Is this valuable to practitioners? Should we be talking to compliance teams, product leads, DEI groups? Supposedly it can help an EU AI Safety Act?

We’d love to hear from anyone at Personio (or anyone using Personio) who can point us in the right direction.

Thanks!

Very interesting ​@Chris.Fii ! 

As you can imagine, AI is something that comes up quite a bit in this community. It’s no surprise, considering how often People pros have to hear about it.

I’d love to hear some thoughts from folks in our community like ​@damayantichowdhury09 (whose written an excellent contribution about AI here), ​@Naturally Mindful , ​@JHBEM , ​@nina.johansson, ​@HRJoy , ​@SabbuSchreiber , ​@HRHappiness, ​@Nathan Jolly 

wow, ​@Chris.Fii this is really impressive if this works as described. Erasing bias at the cognitive level sounds like the kind of disruptive innovation HR desperately needs!

A few questions that come to mind:

  • How do you define and implement bias removal at the cognitive level? Are there risks of losing useful context or nuance?

  • How adaptable is this to different hiring workflows and ATS platforms? Integration is often a dealbreaker.

  • Have you tested this with real-world hiring data or just synthetic samples?

  • What does the audit trail look like in practice? Transparency and explainability are critical for compliance.

  • Can this help with subtle or systemic biases beyond explicit proxies like names and zip codes?

I’m not a recruiter myself, but as someone who cares about fair processes, I’m really interested to see how this evolves and if it can scale beyond the lab. 😃 If you decide to take this further or run pilots, I’d definitely be interested in hearing updates. Please keep the community posted!

Your reply


Terms & Conditions of the Personio Voyager Community