Anthropic restricts Claude from creating clones of existing video games

New guardrails for Opus 5.5

Anthropic recently updated the rules for its Opus 5.5 model to address the growing trend of using AI to create software clones. The company specifically wants to prevent users from inundating platforms like Steam with low-effort recreations of existing titles. As reported by PCMag, the update was designed to steer the model away from infringing on the intellectual property of established creators.

According to Addy Osmani, a member of the technical staff at Anthropic, the company added two specific sentences to the guidelines. These instructions tell the AI to build original games or games within a specific genre, but explicitly forbid recreating a specific existing game. The policy clarifies that even if a user renames the project or calls it an original work, the AI should not assist if it is a recreation of a known title.

This shift marks a clear distinction in how Anthropic views different types of software generation. While the model is still encouraged to help with original designs and user-provided assets, it is now restricted from acting as a tool for direct duplication. The goal is to maintain a balance between providing a powerful coding assistant and preventing the automated theft of artistic work.

Modding and reverse engineering

Despite the new restrictions on cloning, the update does not appear to block many common developer workflows. Testing shows that Claude remains willing to help users decompile games or reverse engineer existing code. For example, users can still ask the model to break down a game's code to figure out how specific mechanics work or to help maintain existing mods.

The policy language suggests that Anthropic is not concerned with projects that breathe new life into old software or create community content. Users can still use the model to port games to different platforms or to develop achievement-tracker utilities that read progress from decompiled files. The AI even remains helpful when analyzing demos of unreleased games for the purpose of modding or porting.

This distinction is important for the hobbyist community that relies on these tools for preservation and customization. Because the rules do not mention modding or porting, these activities remain technically permissible under the current guidelines. The model continues to view these tasks as functional analysis rather than the creation of a competing commercial product.

Collateral damage and workarounds

The new restrictions have caused some unintended consequences for users working on non-commercial projects. Some people have reported that Claude now refuses to help with tasks that do not violate the new rules, such as creating a digital version of a personal board game. These instances suggest that the AI's guardrails might occasionally be overzealous in identifying what constitutes a 'recreation.'

To navigate these hurdles, some users are turning to careful prompt engineering to get the desired results. By breaking down a large project into smaller, more narrowly scoped tasks, users can sometimes bypass the model's refusal to engage with a broad goal. This method involves sharing less information about the ultimate objective while asking for specific, functional pieces of code.

Other developers are exploring more sophisticated methods, such as 'clean room' approaches or specialized apps. One example is an app that allows users to see how a website or game works and then have the AI build a version based on visual points of reference. These methods allow for the creation of new software without directly feeding the AI the code of a competitor.

The competitive landscape

While Anthropic has implemented these specific rules for Opus 5.5, other major AI players have not yet announced similar limitations. In testing, OpenAI's GPT-6 Astra model showed no issues helping to decompile, reverse engineer, or clone games, including those that only have demos available. This creates a different environment for developers who may prioritize unrestricted functionality over specific safety policies.

Chinese models like DeepSeek currently lag behind the frontier models in performance but historically have not shown qualms about taking on such tasks. As the AI landscape evolves, it is likely that other models will eventually adopt similar restrictions or find ways to balance the same issues. For now, the choice of model significantly impacts what a developer can achieve regarding software replication.

The situation highlights a growing tension in the AI industry between utility and safety. Anthropic is attempting to create a middle ground where the AI remains a powerful tool for creation without becoming a factory for piracy. However, the availability of other models means that the ability to clone software remains a possibility for those willing to use different platforms.

Protecting artistic labor

Anthropic's decision reflects a belief that video games are a unique form of art that deserves protection from unfeeling machines. While a digital tool is essentially the sum of its parts, a game involves human labor and artistic decisions that are harder to replicate perfectly. The company argues that while functionality might not be exclusive, the specific artistic direction of a game is worth protecting from automated ingestion.

The policy draws a line between copying a gameplay loop and copying a specific artistic work. For instance, a game can be inspired by a genre or a formula without being a direct clone. The goal is to prevent 'asset-flipping,' where a user buys premade assets and uses AI to slap them together into a replica of a successful title to siphon off profits.

Most casual users will likely not notice these changes in their daily interactions with Claude. The restrictions primarily affect those attempting to build direct clones or those working on projects that accidentally trigger the new safety filters. As Anthropic continues to refine these models, the goal is to minimize these errors while maintaining a firm stance against the automated recreation of existing commercial art.