Chris Russell, Brent Mittelstadt, Sandra Wachter, Kaivalya Rawal, Daria Onitiu
We have not lifted any functions out of this paper's repositories yet, so there is nothing we have run. If it links a repository, it is listed below.
Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.
Explainable Artificial Intelligence (XAI) seeks to explain how an Artificial Intelligence (AI) system arrived at a particular decision. We propose "Rule of Thumb" (ROT) explanations, a new approach to XAI based upon a novel formulation that identifies the most relevant features for predicting the behaviour of an AI system, for a particular datapoint. We show how ROT is well-suited to enable XAI in: (a) zero-shot classification using large language models (LLMs), (b) auditing of opaque AI systems without model access, and (c) the use of AI in scientific discovery. Additionally, ROT meets specific requirements from leading AI regulations, provides a familiar interface and visualisations for XAI practitioners, is model-agnostic, and is substantially faster than alternatives. Code available at: https://github.com/KaiRawal/Rule-of-Thumb-Explaining -Artificial-Intelligence-Systems-using-Partial-Information Artificial intelligence (AI) in the form of large language models has become ubiquitous. One of the key components that has made it disruptive and widespread is its ability to perform zero-shot tasks. Given an applicant's resume, it can say if they are a good fit for a job, or identify an animal in a photo. But does it work? As it stands, the only way to find out if a large language model (LLM) can perform a task, is to label a significant subset of data, and measure how consistent the LLM is with these labels. This data labeling process substantially decreases the disruptive nature of AI. While the AI itself may be zero-shot, the process of deciding if we should deploy it is not. Explainable AI (XAI) has long been promised as a tool to help with this kind of deployment decision. Better understanding of how, and in some cases, why AIs make decisions can increase confidence in deployment and help users identify and correct errors [1,2]. However, precisely when you would expect XAI to be making a resurgence, it is nowhere to be seen. In no small part, this is because of fundamental limitations of the methods. Existing Explainable AI methods try to either peer inside the algorithmic black-box and use access to the internals of models (grad-cam [4], integrated gradients [5], counterfactuals [6], or mechanistic interpretability [7]) which is inapplicable to many widely used closed source LLMs that only allow access via an application programming interface (API), or 1
The same record, over MCP at https://syntology.ai/mcp:
get_harvested_code_for_paper("2608.10766")
get_code_for_paper("2608.10766")
have("2608.10766")
Connect an agent — have() is free.