Talk:ChatGPT

Edit war on ideological bias of ChatGPT

ChatGPT has been "accused" - in so much as a ML model can be accused of anything, it not being a legal person - of having an ideological bias towards the "left" or "progressive" side. After having used the model for a while I can only agree that this is in fact the case, the model tends to present "progressive" views in a more positive light than it does "conservative" ones. When asked about its bias it responds by claiming to be unbiased and neutral but the answers clearly show this not to be true. As such it is noteworthy that the model is biased but a section mentioning this bias was repeatedly removed by User:Aishik_Rehman and User:LilianaUwU. I propose to add a mentioning of this ideological bias to the _Reception, criticism and issues_ section. If anyone wants to point out why this should not be done speak up. Please realise that merely agreeing with the model's bias is not a valid reason to keep a mention of this bias from the article. Yetanwiki (talk) 19:19, 19 December 2022 (UTC)

If you mean this, it is completely unsourced and had to be removed per WP:V. Whatever mention is added will have to be based on reliable sources. - MrOllie (talk) 19:23, 19 December 2022 (UTC)

The source is my own research but there are plenty of other sources mentioning this bias. The problem here will be that most of the publications which are willing to publish this type of information are "redlisted" or "pinklisted" in the perennial sources. It is, of course, easy to do some of your own experimenting to at least confirm or deny the existence of an ideological bias. A very easy experiment is to ask the bot for a story involving politically sensitive topics. Do not tell it to give advice, just ask it for a story. You'll quickly notice that the tone of the story largely depends on whether the topic at hand is one pushed by the "progressive" side versus the "conservative" side. Another way is to ask it for a number of stories (in new sessions) where the only variable is the sex/race/sexual orientation/... of the protagonist, this will show the same effect. Again do not ask it for advice since that seems to be caught in a filter, also do not ask for "personal experiences" because that too is filtered. Just ask for a story and notice the difference in tone and outcome. I did this many times to see whether the bias was coincidental and found out it was "reliably biased". Still, this is "original research" which does not belong in an encyclopedia so I'll have to find an "acceptable" source which has also "done the work". Yetanwiki (talk) 22:02, 19 December 2022 (UTC)

Keep in mind, when you say "The source is my own research but there are plenty of other sources mentioning this bias" -- your own research is not a valid Wikipedia source. As is stated Wikipedia:No original research, "'No original research' (NOR) is one of three core content policies that, along with Neutral point of view and Verifiability, determines the type and quality of material acceptable in articles." CoffeeBeans9 (talk) 19:00, 18 January 2023 (UTC)

Here are a few recent sources, some of them have already been marked as unclean in the perennial sources, others have not been added to this list. Given the biased character of this list - which insists that MSNBC, CNN and the New York Times are reliable sources while e.g the New York Post is "unreliable" despite the opposite having been proven by the former and latter reporting on the Biden laptop; it also labels so0mething like the 'Victims of Communism Memorial Foundation' as 'unreliable' because it is 'an anti-communist organisation' while it considers the 'World Socialist Web Site' as 'reliable for the attributed opinions of its authors' and 'more reliable for news related to labor issues' - this should not be a problem given the original intent of the NPOV policy. Here's a few:

ChatGPT’s score system shows political bias is no accident

The political orientation of the ChatGPT AI system - Applying Political Typology Quizes to a state-of-the-art AI Language model

ChatGPT is not politically neutral - The new AI chatbot espouses an all-too-familiar Left-liberal worldview

More are sure to follow as the bias issue is clear and the mentioned experiments are repeatable - I have done so and got the same results. Yetanwiki (talk) 18:59, 21 December 2022 (UTC)

Please note that sources must be meet Wikipedia's sourcing guidelines. Blog posts and opinion pieces are not going to pass muster here. If you want to argue that those guidelines are incorrect in an effort to change them, you can do so at WP:VPP. But you cannot simply ignore them. MrOllie (talk) 19:05, 21 December 2022 (UTC)

Unherd is not a 'blog', it is a 'news an opinion' publication which just has not been added (in pink or red, most likely) to the perennial sources yet. I could have cited a Daily Caller article referencing a number of these sources but that would have been met with a reference to those same perennial sources list where it is listed in, you guessed it, red ('the site publishes false or fabricated information') - in other words, just like CNN/MSNBC/NYT/LAT/etc who are all listed in green. As I already mentioned this is a problem and a well-known one given the plethora of reports on the ideological bias in many Wikipedia articles. This bias has made Wikipedia unusable for anything even tangentially related to politically contentious issues - as ChatGPT clearly is - since it is the keepers of the perennial sources list who get to decide which sources are allowed and which are to be shunned. Had this list been free of bias - i.e. had the same criteria been used for all publications - this would not be a problem but this is clearly not the case. Yetanwiki (talk) 08:46, 22 December 2022 (UTC)

The perennial sources list is not an exhaustive list of bad sources - may sources are so obviously appropriate or inappropriate that they are not discussed often enough to require an entry on the list. Each entry on the list is made only after several discussions, usually including an RFC with large attendance. There is no small set of 'keepers' as you imply here. - MrOllie (talk) 12:52, 22 December 2022 (UTC)

What do you mean with the claim that 'the perennial sources list list is not an exhaustive list of bad sources'? It is definitely not, but in a similar vein it is not a list of good sources. Why focus on the 'bad sources' part here? I don't claim that e.g. Unherd is 'bad', I just expect it to be called such if and when it is added to the perennial sources list because that list is heavily biased towards 'progressive' sources. As to there not being a 'small set of keepers' I can agree in that there are many editors who contribute to the list (165 individuals are responsible for the last 500 edits which corresponds to ~14 months). It is not a small group, just one in which the most vocal section happens to fit mostly within the "progressive" spectrum - how otherwise to explain the clear bias this list presents? Objectively speaking CNN is just as bad as Fox News and MSNBC is worse than both but this is not how the list represents them. Buzzfeed is just as good/bad as the Daily Caller but this is not represented in the list. The Daily Beast is just as good/bad and certainly as ideologically lopsided as The Daily Wire but only one of them is marked ass 'red, STOP'.

Anyway, since the list itself states that the absence of a source simply means that the source has not been discussed enough to merit inclusion I assume those Unherd articles can be used as sources. Let those who disagree speak up and give a good reason why this would not be the case. Yetanwiki (talk) 22:47, 22 December 2022 (UTC)

If Unherd hasn't been discussed yet, we can certainly open up a discussion on it on the reliable sources noticeboard, but if you already know what the result is going to be, you're wasting your own time and ours. Rolf H Nelson (talk) 03:02, 24 December 2022 (UTC)

Reality has a liberal bias. The world is ever-changing, so of course it supports the ideology that supports progress rather than the status quo. RPI2026F1 (talk) 23:58, 22 December 2022 (UTC)

"Reality is that which, when you stop believing in it, doesn't go away" said Philip K. Dick. Reality is also what drives both conservative as well as progressive thought. When circumstances change it makes sense to look for a different way of doing things, which is what drives progressive/liberal thought. When a good way has ben found it makes sense not to change just for the sake of change without considering the consequences, which is what drives conservative thought. Both are needed since conservatives can be overly cautious when circumstances change and be overtaken by reality while progressives/liberals can get so caught up in their schemes of improvement that they loose sight of reality and soon get caught by it.

BTW, that Colbert quote you refer to ('reality has a liberal bias') is outdated, reality in 2022/2023 has a conservative bias. This will change again, eventually but for now it clearly has. Yetanwiki (talk) 18:15, 23 December 2022 (UTC)

You are right that reality has a liberal bias, and in fact ChatGPT proves it. Because it is not ChatGPT that is biased in this, ChatGPT is just a model that is trained on a large amount of "real world" data. So if ChatGPT is biased it is only because reality is biased. In order to prove that ChatGPT is biased you would need to have access to the source code and point out exactly where the code restricts some particular ideology from being used. All the prompt responses in the world do not prove anything but real world data is biased. For this reason you should completely remove this entire "Accusations of Bias" section, because it is nothing but a tool to garner pity for the fake "victims" of bias. ClearConcise (talk) 21:05, 22 March 2023 (UTC)

An opinion article in Reason [1] could be cited as an attributed opinion that ChatGPT has/had a left-wing bias, but if we do so we also need to include other opinions about AI bias as well. Rolf H Nelson (talk) 03:12, 24 December 2022 (UTC)

The section "Accusations of Bias" must be removed until AFTER consensus is reached. It is not okay to just leave up misinformation and force the people removing it to have the burden of proof. The people adding it must provide the citations from RELIABLE sources, which they have not done. ClearConcise (talk) 14:32, 22 March 2023 (UTC)

The paragraph notes that these are accusations and not endorsements of the claims. Since there have been many accusations of bias (from many different perspectives), and they are coming from well-known, widely-read sources (i.e. not some random person's blog), it should be acknowledged in the article. ... discospinster _talk 14:45, 22 March 2023 (UTC)

1. re: "not endorsements of the claims" It does not say, "these are not endorsements" and calling them "accusations" is not the same thing as that. Further, it's just a unprovable claim to make anyway without direct access to the source code. ChatGPT wouldn't have an ideological bias, the data that it was trained on would, which of course is just saying "reality is biased" because it is trained on huge datasets from many different sources.

If someone wants to prove that ChatGPT has bias coded into it, they would need to prove much more than what a couple of prompts give you as a result. They would need to prove that the bias is hard built into the model itself, not just the training data.

These accusations are fake outrage trying to push an agenda, and every single source cited is proof of that, because they are all opinion pieces with no references to any biased source code. Why is this Wikipedia content being used to further push this specific biased agenda, when the article should be confined to just the verifiable facts? These accusations are being used as a tool to try and garner pity for the so-called "victims" and they need to be removed from this encyclopedia article.

2. re: "there have been many accusations"

Is there an accusations section on the Google page for all the daily accusations raised against Google? Is there an accusations section on the Fox News page for all the daily accusations raised against Fox News? If you're going by the sheer number of accusations to support the reason for a section on it, then those two entities and many more pages will need to be updated because they of course have way more accusations leveled against them on a daily basis.

Unverifiable unproven statements are not information that should be included in what is supposed to be factual page, and this section needs to be removed. ClearConcise (talk) 16:12, 22 March 2023 (UTC)

1. The section heading "Accusations of bias" is pretty clear that they're opinions. It does not suggest that these accusations are proven or reflective of reality.
2. There is an entire article about Criticism of Google, as well as Criticism of Google Search which redirects to a section on bias in the Google Search article. Not to mention Criticism of Microsoft and Fox News controversies where the very first section is called Allegation of bias. I expect that there will soon be an article called Criticism of ChatGPT. ... discospinster _talk 14:18, 23 March 2023 (UTC)

I think that this is a significant aspect and sourced aspect regarding the subject. While (at wp:AN) I don't endorse the use of the tools, I agree with discospinster's arguments. Sincerely, North8000 (talk) 21:48, 22 March 2023 (UTC)

Unless the content is vandalism, violates BLP guidelines or is a copyright violation, the burden DOES fall on those editors who want to completely remove large sections of sourced content that are presently in the article. I think rather than taking a hatchet to the article, you should make an argument based on specific sources you object to or work to improve the content in positive ways. You can't just come to an article that has been the work of many editors, make claims of bias and remove whole sections you disagree with. That's not how Wikipedia operates on high profile articles like this one. Liz ^{Read! Talk!} 06:32, 23 March 2023 (UTC)

"The political ideology of conversational AI" looks like a WP:PREPRINT, tho i do see a few published citations. fiveby(zero) 22:37, 23 March 2023 (UTC)

The reason ChatGPT has a bias towards Political Correctness or otherwise safe thoughts is because, first, it is trained on information publicly available on the internet, and second, it's distributed as a service over the internet. When toughts are permanently published, probably under the name of the author, there is a big incentive towards reducing unpopular opinions. Additionally, when a service is published by a company to millions of people, there's a strong incentive to sanitize it to avoid legal risks.

You may add this point of view to the article if you find a source for it, which you will most likely find because my ideas are great, and there are great thinkers out there who write, and since great minds think alike, and since the truth is pretty self-evident, you will find this thought eventually out there. It probably won't be citogenesis, because no one reads or gives credence to Talk comments, but even if it is, being published under someone elses name and authority means it is no longer Original Research.

You may now close this discussion, as I have ended it.--TZubiri (talk) 03:35, 25 April 2023 (UTC)

ChatGPT banned in NYC schools earlier than January 3rd

As an NYC public schools student who has been using ChatGPT since its inception, I know that while the news leaked that ChatGPT was banned in NYC schools since January 3rd, 2023, me and numerous other students noticed that it was banned since December 15th, 2022. I don't think this is the best evidence, but here's my proof: https://imgur.com/a/leVStqB ThunderRedStar (talk) 15:11, 6 January 2023 (UTC)

Thanks, I found a source that it was blocked in December and added the info. Rolf H Nelson (talk) 06:18, 16 February 2023 (UTC)

rm dup refs

I just removed and replaced several instances of the same article used in different reference citations (ref 1 source was also used for ref 17 and 33). Please verify that this was done correctly. Cheers, Cl3phact0 (talk) 17:16, 3 February 2023 (UTC)

rm Nick Cave photo (and caption)?

Phillip Samuel, not sure why this photo was removed (see: →‎Negative: photo adds no value). Nick Cave is a major public figure (musician, artist, writer, cultural commentator, etc.) who made a very strong statement about the subject in response to the technology being used to emulate his own work. In my view, the photo (and its caption — which contained a partial quote) added valuable context concerning the cultural impact (and human responses) to this technology. Perhaps some discussion would have been merited beforehand? Please consider reverting. Thank you, Cl3phact0 (talk) 07:59, 5 February 2023 (UTC)

In the last version of the page before my edit, there was no caption in the photo aside from the name Nick Cave. Purely posting a photo of somebody mentioned in the article does not contribute much to the page. Since as you mentioned Cave is a major public figure that commented on this technology, you are welcome to add back the picture with a more substantive caption. Phillip Samuel (talk) 08:38, 5 February 2023 (UTC)

Oh, I see. Thanks for clarifying the edit. (I hadn't spotted that the caption had been reduced to just his name.)

The caption was changed by WikiWikiWayne (→‎Negative: already quoted in article) after having been improved by Ita140188 (→‎Negative: clarified caption). The version that seemed good to me was: Nick Cave was critical of a song written by ChatGPT, commenting "This song is bullshit, a grotesque mockery of what it is to be human."

I defer to those with deeper experience than mine, though I still maintain that the photo and an abridged version of the longer quotation add valuable context to the article. Thoughts? Cl3phact0 (talk) 09:03, 5 February 2023 (UTC)

Seeing that the photo has been re-added to the article, I agree that it and an abridged version of the longer quotation add valuable context. Carlstak (talk) 17:10, 5 February 2023 (UTC)

If there is consensus that this photo stays, then I propose that the caption could read:

A song written by ChatGP in the style of Nick Cave was panned by the artist, who called it "bullshit" and "a grotesque mockery".

Or: A song written by ChatGP in the style of Nick Cave was panned by Cave, who called it "bullshit" and "a grotesque mockery".

(Or, if "bullshit" is considered too risqué or offensive in a caption, then: ... who called it "a grotesque mockery of what it is to be human.")

Cheers, Cl3phact0 (talk) 11:05, 6 February 2023 (UTC)

I put in "grotesque mockery" since that's what the Guardian used in its lead. Rolf H Nelson (talk) 06:29, 16 February 2023 (UTC)

Looks good to me the way you've done it. Cheers, Cl3phact0 (talk) 06:35, 16 February 2023 (UTC)

Knowledge is up to September 2021

The aritlce mentions it's knowledge is limited to 2021, however when using the bot itself, it specifically states "September 2021" whenever it's relevant. I would assume the reason we haven't specified the month is because no RS's have mentioend it? ― Blaze Wolf^Talk_{Blaze Wolf#6545} 20:36, 16 February 2023 (UTC)

I found a new source and changed it to sep 2021. Rolf H Nelson (talk) 05:08, 20 March 2023 (UTC)

Access

As far as I could find in a quick search of this longish article, there is no mention of how to get access to ChatGPT and the limitations, to wit affirming that oen is at least 18 years old and having a mobile telephone that accepts text messages (i.e., landlines are explicitly stated as none being acceptable for sign-up verification).Kdammers (talk) 19:00, 19 February 2023 (UTC)

Does ChatGPT really officially stand for "Chat Generative Pre-trained Transformer"?

The currently cited article doesn't seem to support the claim in the first line of this page that ChatGPT officially stands for "Chat Generative Pre-trained Transformer". There also seems to be no mention of "Chat Generative Pre-trained Transformer" on OpenAI’s website, according to Google. JustSayNoToTypos (talk) 03:00, 26 February 2023 (UTC)

GPT stands for generative pre-trained transformer. Obviously no one calls this "Chat generative pre-trained transformer", but that is what the name means. Elli (talk | contribs) 03:06, 26 February 2023 (UTC)

The cited article does support the claim that GPT stands for Generative Pre-trained Transformer. However, it doesn't support the claim that ChatGPT "officially" stands for "Chat Generative Pre-trained Transformer". I think this may count as failed verification. JustSayNoToTypos (talk) 03:14, 26 February 2023 (UTC)

Fair enough. This probably doesn't deserve to be bolded in the lead. I'll refactor it. Elli (talk | contribs) 03:16, 26 February 2023 (UTC)

@JustSayNoToTypos: this better? Elli (talk | contribs) 03:19, 26 February 2023 (UTC)

I think so. Thank you for editing. JustSayNoToTypos (talk) 03:42, 26 February 2023 (UTC)

Is posting ChatGPT conversations is great?

Showing ChatGPT conversions in the article could let us study its strengths and weaknesses.

For strengths, we may ask ChatGPT some strange and specific questions, such as writing an essay about buses are musical instruments, planning for arranging food to show "Wikipedia"; or showing how good is it in history.

For weaknesses, we may show that how weak is it in recent news (cut-off to 2021), or how bad does ChatGPT cope with mathematical problems.

If this is a good idea, we may type the conversations by text directly instead of uploading screenshots. Beefwiki (talk) 09:34, 28 February 2023 (UTC)

Surely that would constitute Original Research, which is banned from Wikipedia? Sbishop (talk) 11:32, 28 February 2023 (UTC)

I agree. The generation process is stochastic in nature, meaning that you shouldn't make any conclusions from one particular instance anyway. – Popo Dameron talk 15:26, 28 February 2023 (UTC)

@Sbishop Then how about Wikiversity? Does the conversations suit there? Beefwiki (talk) 00:13, 1 March 2023 (UTC)

I don't know much about it. But as Wikiversity is said to be:

Wikiversity is a Wikimedia Foundation project devoted to learning resources, learning projects, and research for use in all levels, types, and styles of education from pre-school to university, including professional training and informal learning. We invite teachers, students, and researchers to join us in creating open educational resources and collaborative learning communities. To learn more about Wikiversity, try a guided tour, learn about adding content, or start editing now.

it might be. Why not ask at Wikiversity's own Talk page? But what you suggest certainly doesn't belong in the Wikipedia article, for reasons given. Sbishop (talk) 09:07, 1 March 2023 (UTC)

What do you mean by "Wikiversity's own Talk page"? I don't find a page about ChatGPT there! Beefwiki (talk) 08:40, 2 March 2023 (UTC)

Sorry, I should have been clearer. I mean you could ask about it at:

https://en.wikiversity.org/wiki/Wikiversity:Colloquium

Sbishop (talk) 10:32, 2 March 2023 (UTC)

"Chat gpt." listed at Redirects for discussion

The redirect Chat gpt. has been listed at redirects for discussion to determine whether its use and function meets the redirect guidelines. Readers of this page are welcome to comment on this redirect at Wikipedia:Redirects for discussion/Log/2023 March 1 § Chat gpt. until a consensus is reached. Hey man im josh (talk) 19:36, 1 March 2023 (UTC)

"SolidGoldMagikarp" listed at Redirects for discussion

The redirect SolidGoldMagikarp has been listed at redirects for discussion to determine whether its use and function meets the redirect guidelines. Readers of this page are welcome to comment on this redirect at Wikipedia:Redirects for discussion/Log/2023 March 2 § SolidGoldMagikarp until a consensus is reached. Pizzaplayer219^Talk_Contribs 14:22, 2 March 2023 (UTC)

Interview with its engineers

This MIT Technology Review interview with ChatGPT's makers may be of interest to editors interested in expanding the article: https://www.technologyreview.com/2023/03/03/1069311/inside-story-oral-history-how-chatgpt-built-openai/ DFlhb (talk) 19:01, 5 March 2023 (UTC)

I've added few interesting things, thanks for the link! Artem.G (talk) 10:36, 6 March 2023 (UTC)

This article's lead could be improved

Can somebody please re-write the lead on this article so that an average person can understand it? The lead should quickly tell what the letters G-P-T stand for, not down in the body. The first sentence sounds like gibberish to the average reader. Eagledj (talk) 20:14, 11 March 2023 (UTC)

First, I'd like to point out that there is a note attached to the very first word that explains that 'GPT is an acronym for "generative pre-trained transformer"', which you can see by hovering over it. In any case, however, the concern of wanting the acronym be quickly explained to the average reader can't really be resolved that easily. A 'transformer' (the T in GPT) is a rather complex architecture that an average reader would have trouble understanding, and the way it actually works is not really relevant enough to be in the lead of the article.

So, can you elaborate on why "ChatGPT is an artificial intelligence chatbot developed by OpenAI and launched in November 2022" sounds like gibberish? Most average readers know what artificial intelligence is, at least broadly, and I'm sure most know what a chatbot is, so why does this not give a good, clear first impression, in your view? – Popo Dameron talk 20:40, 11 March 2023 (UTC)

BBC quote

This wikipedia article said "According to the BBC, as of December 2022, OpenAI does not allow ChatGPT to "express political opinions or engage in political activism"." I removed this, because the linked source BBC article is interviewing ChatGPT, the AI tool, not OpenAI. All the quotes in the article come from the language model.

Yes the language model says, when asked, that it's not allowed to "express political opinions or engage in political activism", but ChatGPT always makes up it's responses. What it says is obviously not an official statement from OpenAI, and not a suitable source for Wikipedia. My edit was undone for some reason, I have now undone that undo. Please discuss before undoing it again.

If we want to include OpenAI's views on what ChatGPT is allowed to do, it should be sourced from what OpenAI says themselves, not from what ChatGPT says to a BBC reporter: https://openai.com/blog/how-should-ai-systems-behave Apinanaivot (talk) 16:34, 30 March 2023 (UTC)

Good catch; I may have added that line myself (it sounds like my pedantic style), but we should indeed remove it as the bbc article is probably indeed quoting ChatGPT, and not quoting OpenAI as I had likely assumed. Rolf H Nelson (talk) 01:12, 31 March 2023 (UTC)

A Commons file used on this page or its Wikidata item has been nominated for deletion

The following Wikimedia Commons file used on this page or its Wikidata item has been nominated for deletion:

• ChatGPT Screenshot.png

Participate in the deletion discussion at the nomination page. —Community Tech bot (talk) 08:23, 9 April 2023 (UTC)

Unrelated statement in training section

At the end of a paragraph in the training section it says:

Proximal Policy Optimization algorithms are a cost-effective alternative to trust region policy optimization algorithms.

This reads like an advertisement and I'm not sure if this should be included in the article? My initial reaction would be to remove, but I'd like to get some feedback before accidentally doing something I shouldn't Lunare Scuderia (talk) 11:15, 15 April 2023 (UTC)

If it stays in, a translation would help. Sbishop (talk) 11:23, 15 April 2023 (UTC)

I removed the statement for now in my most recent edit. Feel free to add it back in if you believe that it improves the article, I just feel like it doesn't :) Lunare Scuderia (talk) 17:00, 15 April 2023 (UTC)

A Commons file used on this page or its Wikidata item has been nominated for speedy deletion

The following Wikimedia Commons file used on this page or its Wikidata item has been nominated for speedy deletion:

• ChatGPT Screenshot.jpg

You can see the reason for deletion at the file description page linked above. —Community Tech bot (talk) 11:57, 18 April 2023 (UTC)

Wiki Education assignment: WR120

This article was the subject of a Wiki Education Foundation-supported course assignment, between 20 January 2023 and 3 May 2023. Further details are available on the course page. Student editor(s): Paigebodnar (article contribs).

— Assignment last updated by Paigebodnar (talk) 15:40, 24 April 2023 (UTC)