Towards Transparent AI: A Survey on Explainable Language Models

Language models lack transparency in high-stakes applications: The survey emphasizes that the "black-box nature" of language models creates critical accountability problems, particularly in domains like healthcare, law, and finance where stakeholders need to understand why a model made a specific decision.

Existing explainability methods have significant limitations for language models: While explainable AI techniques exist for other types of AI systems, they struggle when applied to language models due to their complex architectures, massive training datasets, and broad generalization abilities—meaning we can't simply adapt old solutions to new problems.

Different language model architectures require different explanation approaches: The survey identifies that encoder-only, decoder-only, and encoder-decoder models need tailored explainability methods, suggesting that a one-size-fits-all approach to AI transparency won't work as these systems continue to evolve.

Two key standards for evaluating explanations: The survey assesses explainability techniques using "plausibility" (whether explanations sound reasonable) and "faithfulness" (whether explanations accurately reflect what the model actually does)—highlighting that some explanations may seem convincing but not represent the model's true reasoning.

Significant research gaps remain in making language models interpretable: The survey identifies open challenges in this field, indicating that achieving genuine transparency in language models is an ongoing problem requiring continued research and development before these systems can be safely deployed in critical decision-making contexts.