We are changing the rules regarding the use of generative AI in short AtCoder Heuristic Contests (AHCs). ## Previous Rules Until now, AHCs have broadly allowed the use of generative AI, while prohibiting automated uses such as repeatedly having generative AI improve a solution based on execution results or scores. The intent was to allow generative AI to assist participants with ideation and implementation, while prohibiting the automation of large-scale trial-and-error processes, which we considered likely to significantly undermine the competitive nature of AHCs. ## Background of the Change Recent generative AI systems have improved rapidly and are now capable of performing a wide range of tasks required in AHCs at a high level, including developing solution ideas, generating code, and making improvements based on results, simply from being given the problem statement. One of the main appeals of AHCs is the process of analyzing the problem, forming hypotheses, and repeatedly implementing and experimenting to improve a solution. As the ideation and trial-and-error capabilities of generative AI have advanced dramatically, it has become increasingly difficult under the previous rules to preserve the extent to which participants’ own reasoning and experimentation affect their results. We are also concerned that more expensive AI plans may provide access to more capable models, longer reasoning time, and a greater number of trials, creating a significant competitive advantage based on how much participants are able to spend on generative AI. For these reasons, we have decided to prohibit the use of generative AI in principle during short AHCs. ## Summary of the New Rules The use of generative AI during an ongoing short AHC is prohibited in principle. The main exceptions are: * Translating the problem statement using the specified prompt * Translating contest-provided tools into another programming language using the specified prompt * Using the following approved code-completion features: * GitHub Copilot Code Completion and Next Edit Suggestions * Cursor Tab * JetBrains AI Assistant Code Completion and Next Edit Suggestions On the other hand, using conversational generative AI for tasks such as understanding the problem, developing solutions or algorithms, creating or modifying code, debugging, testing, analyzing execution results or scores, or considering improvements is prohibited. For details, please refer to: [Generative AI Rules for Short AtCoder Heuristic Contests - 20260818 Version](https://info.atcoder.jp/entry/short-ahc-llm-rules-en) ## Differences from Algorithm Contests such as ABC, ARC, and AGC The use of generative AI during ongoing algorithm contests such as ABC, ARC, and AGC is also prohibited in principle. The new rules for short AHCs follow the same basic policy, but there is one major difference: **the specified code-completion features are permitted in short AHCs.** With code-completion features, comments may effectively function as prompts and cause code to be generated. For this reason, code-completion features are also prohibited in ABCs, ARCs, and AGCs. However, AHC solutions tend to require substantially more code, and we determined that prohibiting code-completion features as well would impose too great an implementation burden on participants. In addition, current code-completion features immediately return completion suggestions based on the code and comments being edited. Unlike conversational generative AI systems such as ChatGPT, they do not spend significant time considering an approach to the problem or performing deep reasoning. We therefore believe that, at present, the specified code-completion features do not substantially replace the ideation and trial-and-error that are central to AHCs, and that their impact on the competitive nature of the contest is limited. For this reason, we will permit them as an exception. ## Effective Date The new rules will take effect starting with the [13th Asprova Programming Contest (AHC070)](https://atcoder.jp/contests/ahc070), which will be held on **August 29, 2026**. ## Long AHCs This rule change applies only to short AHCs. For long AHCs, we plan to introduce two divisions in the future: a **Human Only Division**, in which generative AI may not be used, and a **With AI Division**, in which the use of generative AI is assumed. Because introducing these divisions requires changes to the contest system, we will announce the timing separately. The generative AI rules for each division will also be defined separately from the rules for short AHCs. ## Future Rules We expect generative AI and development environments to continue changing significantly. In particular, we expect it to become increasingly uncommon to write large amounts of code without using generative AI. For this reason, we do not believe that continually making the rules stricter by reducing the set of permitted features whenever generative AI becomes more capable is a desirable long-term approach. In the future, once generative AI becomes inexpensive enough to provide broadly, we are considering offering an official AI service through AtCoder with a fixed credit limit, and changing the rules so that only this official AI may be used during contests. This approach would allow us to incorporate the use of generative AI itself into the competition while reducing disparities caused by differences in participants’ access to generative AI services or the amount they are able to spend on them. We are considering such a transition for both short and long AHCs. We will continue to review the rules based on the capabilities, cost, and adoption of generative AI, as well as its actual impact on contests.