Wikipedia talk:Wikipedia Signpost/2022-08-01/From the editors

Discuss this story

I commissioned GPT-3 to write a poem about this article:

GPT-3, the glorious machine,

Has written an AfD report so fine,

With insights both derisive and sage,

It's sure to make history's pages.

So let's all give three cheers for GPT-3,

The greatest machine we've ever seen,

Long may it reign, and write more reports,

On Wikipedia, the free encyclopedia!

I have some concerns, but I also have no idea what I'm doing, so there's that! Fantastic read, and I'm very interested to see the ongoing implications of this tech. ASUKITE 01:06, 1 August 2022 (UTC)Reply

Thanks for this stimulating piece. I think that this raises questions about the use of generative language models in Wikipedia. Even if the results are mind blowing, I think that we should refuse the use of generative language models in Wikipedia for several reasons :

  • An epistomological reason : large language models such as BERT, GPT-3 and the most recent one Bloom are trained using a lot of text from the Internet including Wikipedia. The quality of those models comes from the fact that they are trained on text written by humans. If we use generative language models on Wikipedia, future language models will be trained on a mixture of human and AI generated text. I guess that I some point it will become meaningless.
  • A legal argument : GPT-3 is not open source. It is a proprietary algorithm produced by OpenAI. We should be very suspicious with such a powerful proprietary tool. What happens if the price prohibitive? BLOOM, the most recent model, is not proprietary but not open source. It uses a responsable AI license (https://huggingface.co/spaces/bigscience/license). It is far better than OpenAI's approach but it also raises lots of questions.
  • A technical argument : Wikipedia is not only about writing articles but also about collaborating, explaining decisions and argumenting. I don't think that AI are able to have a real discussion in a talk page and we should still remember that the AI don't know what is good, true or just. Humans do.

Maybe it would be worth to have a sister project from the foundation using an AI based encyclopedia (Wiki-AI-pedia). Now, we have a problem. It may be very difficult in a near future to detect contributions generated with generative AI. This will be a big challenge. Imagine an AI which would be expert in vandalism? PAC2 (talk) 06:27, 1 August 2022 (UTC)Reply

  • Many thanks for this trial which is quite remarkable. A key strength of such bots is that they are good at following rules while humans tend to cut corners. For example, consider the recent cases of Martinevans123 and Lugnuts who have both been pilloried for taking material from elsewhere and doing a weak job of turning it into Wikipedia copy. A good bot seems likely to do a better job of such mechanical editing. As the number of active editors and admins suffers atrophy and attrition, I expect that this is the future. The people with the power and money like Google and the WMF will naturally tend to replace human volunteers with such AI bots. Hasta la vista, baby ... Andrew🐉(talk) 09:29, 1 August 2022 (UTC)Reply
  • I am completely blown away by this. I have been following these AI developments for some time, but seeing them used for this application with such coherence is unbelievable. I have many confused and contradictory thoughts about the implications of AI advancement on Wikimedia projects, but for now I'll limit myself to one thing I am clear on: whether for good reasons or bad reasons, soon each person in the Wikimedia community will need to be aware of the technological levels of tools like GPT-3, DALL-E and their successors, and this Signpost experiment in writing is a fascinating way to draw people's attention to it. — Bilorv (talk) 14:39, 1 August 2022 (UTC)Reply
  • I've been doing something broadly similar to your little exercise for quite a few years. I find an interesting high quality article on a foreign language Wikipedia, use Google Translate to translate it into English, copyedit the translation, and then publish it on en Wikipedia (with appropriate attribution). In fact my first ever Wikipedia article creation (Bernina Railway, created in 2009) was done that way. Over time, the Google Translate translations have become better and better, and with some languages (eg French, Italian, Portuguese) they are now generally so good that only minimal copyediting is necessary. I even occasionally receive compliments from native speakers for the quality of my translations from languages such as French (in which I am self taught, and which I do not speak well), and Italian (which I cannot read or speak). Bahnfrend (talk) 05:34, 2 August 2022 (UTC)Reply

"I heard language models were racist" Don't AI models have some sort of system to block "problematic prompts"? I know that DALL-E 2 blocks problematic prompts as per the following quote in an IEEE article: "Again, the company integrated certain filters to keep generated images in line with its content policy and has pledged to keep updating those filters. Prompts that seem likely to produce forbidden content are blocked and, in an attempt to prevent deepfakes, it can't exactly reproduce faces it has seen during its training. Thus far, OpenAI has also used human reviewers to check images that have been flagged as possibly problematic." Maybe GPT-3 could use a similar system. Tube·of·Light 03:40, 4 August 2022 (UTC)Reply

The GPT-3 used on OpenAI's site has a mandatory content filter model that it goes through; if content is marked as problematic, a warning appears and OpenAI's content policy doesn't allow for reusing the text. 🐶 EpicPupper (he/him | talk) 04:25, 4 August 2022 (UTC)Reply
Exactly. My point is that @JPxG could have mentioned that such models have restrictions to prevent abuse. Tube·of·Light 05:48, 4 August 2022 (UTC)Reply
It should be noted that such filters are often a rather ad-hoc measure, with DALL-E 2 believed to merely be adding keywords like "black," "Women," or "Asian American" randomly to text prompts to make the output appear more diverse. It is fairly easy to get past such filters using prompt engineering, and as such, I would not rely on those filters to protect us from malicious and biased uses. Yitz (talk) 19:12, 4 August 2022 (UTC)Reply
Racism is far more nuanced, pernicious and deeply embedded than just saying the N-word or writing in the style of Hitler. To adapt the common phrase "garbage in, garbage out": racism in, racism out. Take a look at our excellent article on algorithmic bias. If DALL-E 2 isn't specifically designed to avert stereotypes from the dataset then it will perpetuate them (and it's hard to see how it could be—the key novelty of machine learning is that the programmers have little idea how it works). I'm sure if you analyse a large range of its output, you'd find it draws Jewish people or fictional people with Jewish-sounding names as having larger noses than non-Jewish people, or something similarly offensive. However, this is no criticism of The Signpost using curated DALL-E 2, Craiyon and GPT-3 content; I can't see any particular biases in this month's issue. — Bilorv (talk) 22:16, 4 August 2022 (UTC)Reply
  • formerly known as "DALL-E Mini", despite having no relation to DALL-E "Formerly"? Aw, why? The Java / JavaScript relationship was definitely the right model to follow on this. /s -- FeRDNYC (talk) 08:57, 10 August 2022 (UTC)Reply
  • With regards to the process a neural net uses to create images versus a human artist: the model does not experience qualia. It cannot have intent so it cannot create in the way a human can. Humans created art in prehistory without training on other art because it didn't exist, just like in the modern era artists have created quantum leaps in artistic style like cubism, impressionism etc. The model cannot possibly create anything new. When you learn fine art you don't go look at a Rothko painting and then immediately pick up a bucket of paint, you go through years of learning the foundations of figure drawing, perspective etc. Artists have an understanding of the world and their own interior life that the model cannot possibly have and that's why human works, even if derivative, are art and these images are imitation. Omicron91 (talk) 07:52, 23 August 2022 (UTC)Reply