Edit history

Not true, this does not occur frequently. This study and software LiCoEval from 2024 found 0.88% to 2.01% of code “strikingly similar to existing open-source implementations”. Afaik this is mostly textbook examples, snippets from stack overflow snippets or common github repositories, often replicated api examples and language boilerplate. How you prompt and refine also matters, and for generating novel code or business logic the LLM simply cannot use memorized snippets.

Presumably since then LLMs have worked to reduce that number of memorized code. Since LLMs cannot memorize all their training data, that number is limited. LiCoEval can find the often memorized examples and train to remove them, or suppress them, or they find other ways to reduce direct reproduction from memorization. For example it would be possible to do what malus.sh does with all the training data. Then it cannot memorize copyrighted code.

So for a model that came out 2026 this already small number might not be that relevant anymore.

Not true, this does not occur frequently. This study and software LiCoEval from 2024 found 0.88% to 2.01% of code “strikingly similar to existing open-source implementations”. Afaik this is mostly textbook examples, snippets from stack overflow snippets or common github repositories, often replicated api examples and language boilerplate. How you prompt and refine also matters, and for generating novel code or business logic the LLM simply cannot use memorized snippets.

Presumably since then LLMs have worked to reduce that number of memorized code. Since LLMs cannot memorize all their training data, that number is limited. LiCoEval can find the often memorized examples and train to remove them, or suppress them, or they find other ways to reduce direct reproduction from memorization. For example it would be possible to do what malus.sh does with all the training data. Then it cannot memorize copyrighted code.

So for a model that came out 2026 this already small number might not be that relevant anymore.

PS: My personal opinion as someone against copyright law in general is that if a computer can generate code for some function, no code can meet the standard of being worthy of copyright. It is not “creative” but “trivial”. Because a machine can do it already. And I believe that is at the heart of this whole matter, the deep insult to what makes us human. If what an LLMs generates is not creative, then any human work that can be generated by an LLM with comparable quality is not creative either.

Not true, this does not occur frequently. This study and software LiCoEval from 2024 found 0.88% to 2.01% of code “strikingly similar to existing open-source implementations”. Afaik this is mostly textbook examples, snippets from stack overflow snippets or common github repositories, often replicated api examples and language boilerplate. How you prompt and refine also matters, and for generating novel code or business logic the LLM simply cannot use memorized snippets.

Presumably since then LLMs have worked to reduce that number of memorized code. Since LLMs cannot memorize all their training data, that number is limited. LiCoEval can find the often memorized examples and train to remove them, or suppress them, or they find other ways to reduce direct reproduction from memorization. For example it would be possible to do what malus.sh does with all the training data. Then it cannot memorize copyrighted code.

So for a model that came out 2026 this already small number might not be that relevant anymore.