OpenAI got a feed, Stack Overflow got a model
The two companies announced a partnership giving OpenAI access to OverflowAPI, a feed of Stack Overflow's questions and answers, along with community feedback intended to improve model accuracy on technical topics. OpenAI committed to attributing Stack Overflow content shown inside ChatGPT. In return, Stack Overflow gained access to OpenAI models to build into OverflowAI, its own set of AI features. [1]
The exchange is a straight data-for-model swap rather than a cash deal with a disclosed price. Stack Overflow supplied the raw material an OpenAI model needs to answer coding questions well; OpenAI supplied the model Stack Overflow needed to ship a competing assistant on its own site rather than watch developers leave for ChatGPT. [1]
Contributors tried to pull their answers back
In protest, some users deleted or edited their own highly ranked answers so the licensed feed would not include them. Stack Overflow restored the original content and suspended the accounts responsible, saying the edits did not comply with its terms of service. [2]
The protest ran into a fact about how Stack Overflow works rather than a policy Stack Overflow invented for this deal. Answers are contributed under a Creative Commons license that survives after the fact; a user cannot unilaterally revoke text once it is posted, whether the downstream use is another programmer's search result or a training feed. That is what let the platform treat deletion after the announcement as a violation to reverse rather than a request to honor. [2]
Ten days before Reddit signed the same kind of deal
On May 16, OpenAI signed a similar data licensing and distribution agreement with Reddit, giving OpenAI access to Reddit's content feed and making OpenAI a Reddit advertising partner. Reporting on that deal noted the earlier Stack Overflow protest as a preview of the reaction a platform's own contributors have when their posts become licensed training data without separate compensation or consent. [2]
Both deals share a structure: a platform built on unpaid user contributions licenses that archive to a model maker, and the users who wrote the archive find out from a press release rather than a negotiation. Stack Overflow and Reddit each had a large forum of accumulated text and a declining reason to keep hosting it for free once OpenAI could pay to license it instead. [1][2]
Sources
- Stack Overflow and OpenAI Partner to Strengthen the World's Most Popular Large Language Models
Stack Overflow · May 6, 2024
- OpenAI inks deal to train AI on Reddit data
TechCrunch · May 16, 2024