Atlas of Weak Signals
Poison Piling
Group Members: Beste, Ayal, Agnese, Max, Heba, Aiman
Description: Poison pilling (or data poisoning) is the deliberate insertion of malicious, misleading, or corrupted data into datasets used to train machine learning or AI models. The goal is to “poison” the model’s learning process so it behaves incorrectly, produces biased or nonsensical outputs, or even leaks sensitive data.
The definition was copied from ChatGPT when asked: “Do you know what poison pilling is as it relates to putting in data that messed up ai data training?”
We later asked ChatGPT how it felt about this practice. Ask us if you want to read its response.
https://chatgpt.com/share/6937e4f3-3b4c-8009-a287-3997a9f7259c
We also had visual depiction of the same:
As a group we spent time discussing relevant topics to this such as Repair workshops, Sudan protests and creative ways to resist, parasitic architecture, anarchist cookbooks. Its fascinating how one trigger topic can open such a variety of discussions and topics.