LLMs and Documents: The Delegation Leading to Silent Corruption

Delegating Tasks to LLMs: A Risk to Document Integrity

We are living in an era where artificial intelligence is no longer limited to answering questions but is taking on complex and prolonged tasks. Users now trust large language models (LLMs) to handle a variety of tasks, ranging from code editing to formatting professional documents, and even managing accounting books. This increased trust raises questions about these systems' ability to preserve document integrity through multiple interactions.

A recent study highlights a concerning issue: when tasks are delegated to an LLM, it can silently corrupt documents. To explore this phenomenon, researchers established a rigorous evaluation framework called "DELEGATE-52." This benchmark covers 52 professional domains, including legal texts, Python programming, musical notation, and even crystallography.

The researchers tested a total of 19 distinct LLMs using an intelligent simulation method based on a "round-trip" approach. This method involves asking the AI to make a specific modification and then performing the reverse instruction to undo the changes. In an ideal scenario, the model should return the original document intact. However, even the most performant models, such as Gemini Pro, Claude Opus, and GPT-5, can corrupt 25% of the original content after 20 interactions. Less performant models can reach a corruption rate of 50%.

Causes of Document Corruption by LLMs

Accumulation of Errors

One of the main reasons for this corruption is the accumulation of errors. Like in the game of telephone, small mistakes made by LLMs can pile up and become significant. A single modification can introduce localized errors, but a sequence of complex changes can exacerbate the problem over time, leading to drastic degradation of the document.

Deletion and Hallucination

The study also reveals a striking contrast in how different types of models fail. Weaker models tend to delete content, making the problem visible after several interactions due to an obvious decrease in the overall content of the document. In contrast, leading LLMs, like GPT-5, do not delete content but corrupt it. They maintain the overall appearance of the document, even keeping the word count almost intact, but silently modify, distort, or replace factual information with fabrications that still seem plausible. The more intelligent the model, the harder it becomes to detect its corrupted behavior, as the final output appears legitimate at first glance.

Context Overload

Models also struggle to maintain structurally intact information in messy conditions, meaning when there is a lot of contextual information or excessive attached documents present. As the size of the document increases or more "distractor files" are included in the prompt context, the severity and impact of degradation explode. The model no longer adheres to the source text, as it finds it easier to simply guess.

Domain Familiarity

Finally, the model's familiarity with the application domain plays a crucial role. Not all files degrade in the same way in delegated tasks. According to the study, LLMs perform well in highly structured and programmatic domains, such as Python source code. It is when they are pushed to pure natural language tasks or niche spatial formatting that they quickly lose the strict sense of internal logic necessary to keep files completely intact.

Limitations of Agentic Tools

Even when LLMs are enhanced with agentic tools, such as the ability to execute code or read and write files directly, the problem of corruption and degradation of delegated documents does not disappear. In fact, agentic additions do little or nothing to prevent a problem that occurs at the core of the underlying transformer architecture of LLMs. It is necessary to rethink how long-term AI tasks should be verified. Until then, using LLMs as completely unsupervised document editors remains a high-risk gamble.

Iván Palomares Carrascosa, a recognized expert in the field of AI, emphasizes the importance of training and guiding users to leverage these technologies while being aware of their limitations.