Large language models (LLMs) are revolutionizing coding, offering impressive code generation capabilities. However, they sometimes produce incorrect code, which can be frustrating and time-consuming for developers to debug. Imagine an AI coding assistant that only showed you its work when it was confident it was correct. That's the idea behind HonestCoder, a new approach to LLM-based code generation. HonestCoder estimates the LLM's confidence in the code it generates and selectively shows results only when the LLM is highly certain. This approach is based on a fascinating insight: highly confident LLMs tend to produce similar code outputs even with slightly different inputs, almost like a human consistently giving the same answer to a question they know well. When an LLM is uncertain, its outputs vary significantly. HonestCoder leverages this observation by generating multiple code variations for a given prompt and then analyzing the similarity between these variations using multiple metrics, capturing not just textual similarity but also structural and semantic aspects of the code. If the similarity is high, HonestCoder deems the LLM confident and shows the code. Otherwise, it politely declines, saying, "Sorry, I cannot solve this requirement." To test HonestCoder, researchers created TruthCodeBench, a new benchmark covering Python and Java, and applied it to several leading LLMs like DeepSeek-Coder and Code Llama. The results are promising: HonestCoder not only significantly reduces the number of erroneous programs shown to developers, but also accurately predicts whether an LLM can successfully solve a coding problem. This innovative approach could fundamentally shift how developers interact with AI assistants, making them true collaborative partners and boosting productivity. The future of HonestCoder could involve even more sophisticated confidence estimation methods and integrations with retrieval-augmented generation techniques, potentially leading to AI assistants that can explain their uncertainty, learn from developer feedback, and contribute even more effectively to the software development lifecycle.