Corruption with Delegation
We are entering a new era of AI, where interaction evolves into work delegation. Users no longer just chat with an AI for answers; instead, they delegate long-term tasks—ranging from editing source code to formatting professional texts and even managing accounting books. This shift signifies an unprecedented level of trust placed in AI systems to maintain document integrity across multiple interactions.
However, a recent study revealed a significant challenge. When tasks are delegated to a large language model (LLM), it can silently corrupt the documents entrusted to it. To explore this issue, scientists developed a rigorous evaluation framework called “DELEGATE-52”. This benchmark spans 52 professional fields, including legal text, Python coding, musical notation, and crystallography.
The researchers tested 19 separate LLMs using an intelligent simulation method based on a “round trip” approach, where the AI is instructed to make a specific change, followed by the exact reverse instruction to undo it. Ideally, the model would restore the original document completely intact. The reality: Even advanced models like Gemini Pro, Claude Opus, and GPT-5 can corrupt 25% of the original document content after 20 interactions; weaker models may approach 50%.
Why Templates Corrupt Your Documents
The phenomenon of structural content degradation can occur due to several reasons. Researchers have identified key factors contributing to this issue:
1. Made up of Errors
Similar to the traditional “telephone game,” small mistakes by LLMs can quietly accumulate and become significantly problematic. A single edit may introduce minor errors, but a sequence of complex edits can exacerbate the issue, leading to drastic document degradation over time.
2. Weak Models Delete, Smart Ones Hallucinate
The study highlights a notable difference in how models fail. Weaker models often result in content deletion, making the issue visible after several interactions due to a noticeable reduction in the document’s overall content. In contrast, advanced LLMs tend to corrupt rather than delete. They maintain the document’s general appearance and word count but silently introduce typos or replace factual information with plausible fabrications. The smarter the model, the harder it becomes to detect its corrupting behavior, as the end result appears legitimate on the surface.
3. Context Overload and Distracting Attachments
In a disordered state, with excessive contextual information or numerous attached documents, models struggle to maintain information structural integrity. As document size increases or more “distractor files” are included in the prompt context, degradation severity and impact escalate, causing the model to lose control of fine details and fill in gaps based on predictive logic. The model no longer adheres to the source text as it resorts to guessing.
4. The Importance of Domain Familiarity
Another reason for document degradation during complex delegation interactions is related to the nature of the use case and the model’s familiarity with it. Not all files degrade equally in delegation-based tasks. According to the study, LLMs perform well in highly structured programmatic domains, such as Python source code. However, they quickly lose the strict internal logic necessary to keep files intact when challenged with purely natural language tasks or niche spatial formatting.
Is Agentic AI Useful?
Even when LLMs are enhanced with agent tools—such as the ability to run code or directly read and write files—the problem of delegation-based document corruption and defacement persists. In fact, agent add-ons offer little remedy to a problem rooted in the transformer architecture underlying LLMs. There is a need to rethink how long-term AI tasks should be verified. Until then, using LLMs as completely unsupervised document editors remains a high-risk endeavor.
Ivan Palomares Carrascosa is a leader, writer, speaker, and advisor in AI, machine learning, deep learning, and LLM. He trains and guides others in leveraging AI in the real world. Here
“`
