OpenAI's Codex Does Not Understand Code, yet Offers Both the Ingenious and the Mundane

ALLEGATIONS OF THE OPERATIVE COMPLAINT

A. OpenAI Develops A Generative AI Tool Called Codex.

OpenAI is a nonprofit organization that develops machine learning models, also referred to as “Artificial Intelligence.” Compl. ¶¶ 2, 126, 128. Such models are typically trained through exposure to a corpus of material called “training data.” Compl. ¶ 79. The patterns discerned from the set of training data become part of the model, which can then generate answers based upon those patterns in response to user prompts. Compl. ¶ 81; see Compl. ¶ 52.

The model at issue in this case is called Codex. Codex is a generative AI model trained on publicly available computer source code. Compl. ¶¶ 52, 135. The model embodies “inferred … statistical patterns governing the structure of code,” Compl. ¶ 52, which it has discerned from the training data based on “a complex probabilistic process.” Compl. ¶ 79. It is thus capable, in response to a prompt, of “predic[ting] … the most likely [coding] solution.” Compl. ¶ 79. “[E]ssentially [it] returns the solution it has found in the most [coding] projects when those projects are somehow weighted to adjust for whatever variables [the model] ha[s] identified as relevant.” Compl. ¶ 79.

Generative AI models are capable of “simulat[ing] human reasoning or inference,” engaging in the same sort of pattern recognition, synthesis, and prediction we do. Compl. ¶ 2. AI models like Codex also enable statistical analysis and prediction vastly more powerful, efficient, and sensitive than what the human brain can accomplish. Compl. ¶ 81. At the same time, Codex “does not understand the meaning of code,” Compl. ¶ 54, nor its “semantics and context the way humans do,” Compl. ¶ 81. Codex thus offers both the ingenious and the mundane, a powerful tool of invention for humans who supply the insight to direct its range of performance.