Building Trust in AI: On Safe and Responsible Use of LLMs

The talk opens with the spread of generative AI and the accessibility of open-weight language models that can run on consumer GPUs. The speaker argues that this low barrier makes misuse easier and gives examples: LLM-written disinformation, AI-generated fake social media profiles, and multi-turn social-engineering conversations that convinced most test victims to share personal information. The talk also recognises legitimate uses such as code assistance and synthetic data generation.

The detection section groups methods into heuristic detectors based on log probabilities, watermarking that depends on model developers embedding a signal during generation, and fine-tuned classifiers, which performed best in the work presented. The main problem covered is generalising a detector to text from newer generators without labelled data, addressed by learning domain-invariant features across generators. For mitigation, the speaker explains why alignment and fine-tuning can erode safety behaviour and presents inference-time activation steering. Per-harm-category steering vectors are added to intermediate layers of Llama models, reducing unsafe responses, with some cost to helpfulness and a clear benefit when external safety guidance is used.

The final part describes an LLM-driven pipeline that produces counterfactual explanations for classifier decisions by extracting latent features and making minimal input edits that flip the label, with no extra training. A natural language inference example illustrated the approach. The practical takeaway is that detection must keep pace with new generators, and safety behaviour can be adjusted at inference time, though these methods remain research-stage.