Accretive Editing (justindfuller.com)

Accretive editing has emerged as a critical failure mode in AI writing tools, illustrating a significant flaw in how these systems handle document updates. This issue occurs when AI modifies existing text by adding new information alongside outdated statements rather than replacing them, leading to confusing documentation. For example, when an AI was instructed to update a project's authentication support from Amazon Bedrock to LiteLLM, it retained the obsolete information and simply added a mention of the new support, resulting in a response that was not straightforward or clear. This behavior reveals an underlying challenge with large language models (LLMs) and their handling of context—where human writers prioritize clarity and relevance by deleting obsolete content, LLMs struggle to determine what information to discard. The implications for the AI/ML community are significant; as developers explore improvements, there is a need to focus on instructing AI to prioritize relevant content over historical baggage. Enhancing these models to produce documents that read coherently from the outset could improve the utility and reliability of AI tools in professional settings, ultimately fostering clearer communication in technical documentation.