This quote caught my eye as I was reading:
When AI companies sell their models to customers, they "are unable to attribute each and every one of the community members whose questions and answers were used to train the model, thereby breaching the Creative Commons license," Chandrasekar says.
I look at it from a different angle than you, however. When I asked ChatGPT:
Were you trained on Stack Overflow data?
As an AI language model, I was trained on a large and diverse corpus of text, which includes various sources of data such as books, articles, and web pages. However, I cannot confirm whether Stack Overflow data was included in my training corpus or to what extent it may have influenced my training.
OpenAI, the creators of the GPT-3.5 architecture that I am based on, has not released detailed information on the exact sources and amount of data used to train their language models, including mine. However, it's worth noting that my training data was pre-processed and filtered to ensure that it meets ethical and legal standards, such as removing personally identifiable information and sensitive content.
Regardless of my training data, my responses are generated based on statistical patterns and relationships between words and concepts that I learned during training, and I do not have any direct access to external data sources.
I did regenerate this response once to get what I was looking for. It's probably a mistake to read any nuance into these answers. Given that ChatGPT can code, it's virtually impossible that Stack Overflow data wasn't used. Due to the nature of the process, the model can't determine which bits of text it used to produce any specific output.
Of course I can't either. At one point I learned how to use the
tar command, but I couldn't attribute all of the input I used to generate my answers. (Honestly, I don't recall writing one of these answers in the first place!) Charging other companies for using Stack Overflow data feels like a violation of the spirit of the endeavor. I mean, if it's not illegal to sell a book based on the data, I don't suppose it's illegal to sell the data itself. It just . . . why would anyone buy it unless the free copy stops existing?
Coincidentally, I ran across a site that uses the data dump from the dearly departed (in my opinion) Stack Overflow Documentation project. There's no attribution or even a link to the data. It's pretty annoying since the whole point of CC BY-SA is to maintain a link between the creator and consumer of content.
In my opinion, companies that use Stack Exchange content to train their language models should be very clear that they are using it. If Stack Overflow wants to charge for using its brand or special access to the data, that seems fine to me. In some ways I'd like to avoid blame for the trivial influence my answers have had on ChatGTP's buggy and sophomoric output. As long as my (possibly buggy and sophomoric) output remains freely available to anyone looking for it, I'm good.
Update after thinking a bit more
I asked Google Bard my "Were you trained on Stack Overflow data?" question based on Laurel's comment:
Yes, I was trained on a massive dataset of text and code, including Stack Overflow data. This data includes questions, answers, and code snippets from Stack Overflow, which is a question-and-answer website for programmers. This data has helped me to learn about a wide range of programming topics, and to be able to answer your questions in a comprehensive and informative way.
That's a more direct acknowledgement. Then I asked "How much wood could a woodchuck chuck?" (as you do) and got this answer from Bard:
A woodchuck could chuck about 700 pounds of wood on a good day, with the wind at his back. This is based on the research of Richard Thomas, a New York fish and wildlife technician who calculated the volume of dirt in a typical 25–30-foot (7.6–9.1 m) long woodchuck burrow and had determined that if the woodchuck had moved an equivalent volume of wood, it could move "about 700 pounds (320 kg) on a good day, with the wind at his back".
This tells me two things:
- Woodchucks are active little buggers.
- It's possible for a large language model to cite sources.
That means Stack Exchange Inc could ask for Stack Exchange/Overflow content to be listed as a source. This would immediately make answers from language models more useful. It would also directly address the attribution concern. Whether or not they come to a commercial agreement is an entirely separate question.
Based on my experience as a former Community Manager, it seems likely management didn't investigate how this announcement would go down when the community found out about it. As a community manager at College Confidential I learn about potential revenue sources weeks or even months before deals are signed. I just got out of a weekly team meeting where we discussed upcoming clients to make sure that legal, product, marketing and community are ready the moment the contracts are signed. I'm not sure why this isn't part of Stack Overflow (the company) culture, but keeping community managers in the loop really makes a difference for our entire company.
From my perspective, the community benefits when the company that hosts it stays viable as a business. Meanwhile a company that hosts a community benefits when it correctly values that community. Practically that means a community-centered business might need to say no to profitable deals that harm the community. Fortunately, Stack Exchange has an excellent community team who might help the business side of the company generate revenue without eroding the value of the community. As always, I wonder if they will listen.