Hidden Bias in LLMs: A Practical Deep Dive
Details
Most AI talks stay on the surface. This session looks underneath. Informal and friendly tech talk with networking afterwards at the pub.
### Who am I?
Hi, I’m Gabriele — a data scientist and AI researcher.
I run small, independent AI events in London as a volunteer.
You can read more about my work here:
👉 https://www.linkedin.com/in/gabrielemonti/
Since I’m not sponsored by any company, I ask for a small contribution to cover the cost of the venue.
please purchase tickets on https://luma.com/eulq7drh
You can pay directly on Luma, or message me on LinkedIn if you prefer another option.
---
LLMs are not neutral, and the bias doesn’t just come from training data — it also comes from the embeddings that power every RAG and agent system.
Alongside the embedding experiments, I’ll also walk through a few core ideas from my slide deck: how majority-class effects work, how model architecture shapes bias, where guardrails fail, and why even simple data distributions can push models toward predictable errors. It’s a small dose of data science behind the scenes, enough to understand why these biases appear before we even get to the embeddings.
Then we’ll dive into a set of experiments comparing several popular embedding models (MPNet v1/v2, MiniLM, DistilBERT) and show how each one produces different gender associations — and how those differences silently affect retrieval, ranking, and agent behaviour.
This is a practical session based on real slides and real cosine-similarity results — not theory and not a startup pitch.
***
### What we’ll cover
- Where bias in LLMs actually starts
- How embeddings from different models contradict each other
- Why some modern models “overcorrect” and flip stereotypes
- How majority-class behaviour shows up in neural nets
- How embedding drift leads to retrieval drift (and unpredictable agents)
- Simple ways to test and monitor this in your own systems
***
### Who it’s for
Anyone building with LLMs — data scientists, ML engineers, founders, or anyone curious about what’s happening under the hood.
The event contains elements of maths and code but it is SUITABLE FOR ANYONE CURIOUS!
***
### Format
Short talk + live examples + Q&A.
Small room near Liverpool Street.
PUB afterwards for networking
Event is not free
£12 to cover the venue.
please purchase tickets on https://luma.com/eulq7drh