FLCC presenters recommend auditing sources, careful prompting and tools like Nightshade, Latimer and NotebookLM
Summary
To reduce bias, presenters suggested auditing training data, using tools that preserve source citations (NotebookLM), choosing culturally fluent vendors (example: Latimer), and selective local LLM deployment; they also described Nightshade as an artist defense.
Gadeau reviewed mitigation techniques available to practitioners and creators. He described Nightshade, a tool artists use to perturb images so models misclassify them, explaining that it "will take that picture of Seneca Lake and change just a few pixels just enough" to confuse AI while remaining imperceptible to humans. He also recommended use of models or services that publish source lists or provide verifiable citations and noted NotebookLM as an example of a tool that lets users control and audit sources.
For organizations handling sensitive data, Gadeau suggested running LLMs locally where feasible, and for general use he advised explicit prompting (for example, request diverse casts in images or ask the model to "ignore race, gender and age"). He emphasized that prompting and user feedback (thumbs up/down) can help services learn, but warned that feedback loops are imperfect and that vendor and policy actions are also necessary.
AI generated
The text on this page is AI generated. Summaries, highlights, analysis, and video transcripts are all produced from the original source material.
AI can make mistakes, so if you spot one, and we will fix it for everyone.
Note: the source content is unaltered by us. Any content source we link to, be it a video, an audio recording, or a document, is presented exactly as its publisher released it. That publisher is usually a government body, sometimes an individual official or another organisation.
