Background
In recent years, participants in the AtCoder Heuristic Contest (AHC) have increasingly utilized AI technologies, including conversational large language models (LLMs), in various ways. Common examples include:
- Delegating the translation of problem statements to AI
- Obtaining information from AI about unfamiliar knowledge or algorithms
- Asking AI to suggest strategies or algorithmic ideas
- Using AI to assist with implementation tasks such as code completion and bug fixing
- Describing the desired implementation in natural language and delegating code generation to AI
We view the above forms of generative AI usage positively, as tools to support creativity and trial-and-error. There is no policy of uniformly restricting all such usage.
However, recent research has shown that methods involving the automated generation of a large number of solution candidates using generative AI, followed by scoring and selection, can have an extremely strong impact in short contests where manual trial-and-error is inherently limited.
Such methods risk creating a new "pay-to-win" structure where results are influenced by financial resources or computational power, raising concerns about the fairness of the competition.
Moreover, we emphasize the nature of AHC as a competition between humans, and hold the view that human creativity and trial-and-error should remain central even when using generative AI.
Therefore, we do not prohibit the use of generative AI entirely, but impose restrictions only on highly automated usage that involves minimal human involvement and has a direct impact on results.
This policy aims to preserve the benefits of generative AI in supporting creativity and technical assistance, while preventing excessive resource usage from unfairly influencing competition results.
Rules Regarding the Use of Generative AI
Prohibited Uses
The following uses of generative AI are prohibited:
- Automatically generating a large number of outputs using APIs or scripts
- Techniques that involve automatically sending a large number of the same or different prompts to generate multiple candidates, then evaluating or selecting among them based on scores or other criteria, are explicitly prohibited.
- Example: Using a script to call the same code generation prompt 100 times, collecting the different outputs, and comparing or evaluating them for selection
Permitted Uses
The following uses of generative AI are allowed:
- Delegating the translation of problem statements to AI
- Obtaining information from AI about unfamiliar knowledge or algorithms
- Asking AI to suggest strategies or algorithmic ideas
- Using AI to assist with implementation tasks such as code completion and bug fixing
- Describing the desired implementation in natural language and delegating code generation to AI (as long as it does not fall under the prohibited uses above)
Additionally, the following cases are also permitted:
- The use of Chain-of-Thought (CoT) type LLMs (e.g., OpenAI o3) or Agent-type AIs (e.g., Claude Code) that autonomously carry out multi-step reasoning or planning in response to a single prompt is not prohibited. However, if they internally execute test cases and generate or select multiple candidates, such behavior falls under the prohibited uses
- Manually resending the same prompt to regenerate code when the AI output contains bugs or behaves unexpectedly is permitted
- The use of tools like Optuna for automated parameter tuning or score evaluation without using generative AI is also allowed. These are explorations based on strategies designed by humans and are distinct from the prohibited act of outsourcing the generation and selection of numerous solutions to generative AI
Supplementary Notes
Definition of Generative AI
In this rule, "generative AI" refers to artificial intelligence that generates new text, code, or similar content based on training data.
Specific examples include large language models (LLMs) such as GPT, Gemini, Gemma, Llama, and Claude.
Future Policy
Generative AI technology is evolving rapidly, and its usage patterns and impact are also continuously changing.
We will continue to review and adjust these rules as needed, striving to balance fairness and creativity based on actual trends and technological developments.
