A Bloom filter became key evidence in a multibillion-dollar copyright lawsuit against OpenAI. Did not expect that combination ever to show up… ever.
Read this in TechCrunch… NYT is suing OpenAI, claiming ChatGPT was trained on its articles (that were behind the paywall) and sometimes reproduces them almost word for word. OpenAI’s defense was that it could not feasibly search its own data for this.
Here is the engineering problem. Checking every ChatGPT output against millions of NYT articles, in real time, with exact string matching, does not scale. This is where the Bloom filter kicks in.
LLM output is rarely identical to the source. So, if we reword one sentence, the hash will not match at all.
Hence, very likely, they shingle, i.e., break the copyrighted text into small overlapping chunks (5 to 10 words, sliding one word at a time), and hash each chunk into the filter. Do the same to the output. Even if most of a sentence is reworded, one surviving matching chunk still triggers a hit.
This is reportedly how OpenAI’s internal tool, Project Giraffe, worked. Their internal tool used a Bloom filter to cheaply flag outputs worth a closer look (on copyrighted text), instead of running exact matching on everything.