Creating concise, accurate summaries of text is something humans take for granted, but it's been a challenge for AI. How do you teach a machine to distill the essence of a document? And how do you even *test* if an AI summary is good? A new UNT study tackled this problem, investigating how different automatic and manual methods stack up against each other when judging the quality of AI-generated summaries, specifically of technical patent documents. Traditional automatic methods, which often rely on comparing an AI summary to a human-written one, weren't always reliable. Surprisingly, these traditional metrics sometimes disagreed with human judgments of quality. This led the researchers to explore a fascinating alternative: using Large Language Models (LLMs), like those powering ChatGPT, as judges themselves. They found that LLMs showed a remarkably close agreement with human evaluations of summaries. In fact, even open-source LLMs delivered results comparable to the state-of-the-art GPT models. This suggests LLMs themselves may be a powerful tool for scoring how well AI summarizes information—without the expense and time investment of human reviews. The study also explored how to use LLM feedback to *improve* summarization. By feeding AI summaries back into the LLM along with quality judgments, the researchers found they could iteratively refine the summaries to become clearer and more comprehensive. However, a small trade-off in accuracy was observed, hinting at areas for future refinement. While the current study focused specifically on patent documents, the researchers hope to broaden their findings to other types of text. This exciting research highlights the potential of LLMs not only to generate text but also to assess and even enhance the quality of AI-generated summaries, opening doors to more reliable and efficient text processing tools.