Wikipedia:Articles for deletion/AI Constitution

{{AFD help}}

:{{la|1=AI Constitution}} – (View AfDView log | edits since nomination)

:({{Find sources AFD|title=AI Constitution}})

Anthropic promotes its own ideas through Wikipedia. The concept of an "AI Constitution" is a construct originated by this startup company. It lacks widespread recognition within the academic community and has only been reported by certain media outlets such as Engadget. 日期20220626 (talk) 00:35, 22 August 2023 (UTC)

:The same deletion reasons were raised by the user:桃花影落飞神剑 on Chinese Wikipedia. I believe the reason he put forth are worthy of discussion. 日期20220626 (talk) 00:38, 22 August 2023 (UTC)

• Note: This discussion has been included in the deletion sorting lists for the following topics: Language, Philosophy, Products, and Computing. Skynxnex (talk) 01:41, 22 August 2023 (UTC)

:@日期20220626 Could you kindly provide some reliable source that suggest the concept of "AI Constitution" is "lacks widespread recognition within the academic community", thanks a lot ! Cloud29371 (talk) 10:00, 22 August 2023 (UTC)

::Isn't it that the person advocating for keeping the article should provide academic sources from the academic community to substantiate it? 日期20220626 (talk) 12:22, 22 August 2023 (UTC)

:::I am sorry, but it im afraid its the one who states something should provide evidence for their claims, not ask the opposing party to provide counter-claims for him. And as it happens i have already did provide the necessary evidence to disprove the claims made. Luroe (talk) 04:48, 23 August 2023 (UTC)

::::No. Editors of an article have to prove that it is generally notable, not anyone else. Cortador (talk) 06:49, 26 August 2023 (UTC)

:::::Well, I think that the burden of proof should lie on those advocating for deletion, as in not to.shift it from the accuser to the accused, they should provide a clear reason with sufficient evidence for why the article lacks notability (no related google results, no necessary links/sources provided, etc) and not the editors being forced to "prove that they are not a camel" as the saying goes.

:::::On a related note, I did already provide a list of necessary and related sources that prove the article's notability to a sufficient degree. Luroe (talk) 07:07, 28 August 2023 (UTC)

::::I am wondering what would constitute "some reliable source that suggest the concept [...] 'lacks widespread recognition within the academic community'": a list of publications that don't refer to it? a web search that doesn't return many results? Perhaps the burden of proof should be on those arguing for a change -- in this case, deletion -- but it is not obvious to me what would verify lack of recognition. (By the way, I have no argument for or against deleting this article.) Cnilep (talk) 03:18, 28 August 2023 (UTC)

:::::I understand that proving a negative can be incredibly hard or sometimes even impossible, but in such a specific case I think it should come to two things: lack of any results (as you already mentioned) and the fact that the writers of the article cant provide any of the reliable sources on the matter, all that in order to not shift the burden of proof onto the accused party from the accuser, otherwise we fear getting into the "prove that you are not a camel" territory. Luroe (talk) 06:58, 28 August 2023 (UTC)

:That is a very serious allegation that Anthropic purposefully promotes its own ideas through Wikipedia. Do you have any sources to back that up? It would be a very controversial thing, worthy of a news article if that allegation was proven to be true. Personally, I think that everything stated without evidence can be freely discarded without evidence, and I cannot find a single good reason as to why would they exactly want to do it. There are far better and honest ways to raise awareness if they were truly conspired to do so.

:Secondly, im sorry but can you provide evidence for lack of recognition in the academic community? Constitutional AI and OpenAI's rule-based reward model (later development, very similar conceptually to CAI, RBRM for short) are an important advancements in the subfield of AI Alignment, currently they, along with RLHF, are the only two techniques that are used to align major AI systems. I think it would be extremely unusual to have an article on only one of them, while the other one is being ignored for whatever reason. Having found such a glaring oversight i decided to correct it. According to [https://www.semanticscholar.org/paper/Constitutional-AI%3A-Harmlessness-from-AI-Feedback-Bai-Kadavath/3936fd3c6187f606c6e4e2e20b196dbc41cc4654 SemanticScholar] the Anthropic's scientific paper that describes the invention of the CAI method has 165 citations, which puts in the top 1% of scientific papers ever written. So the idea that it has no recognition in academic community is simply ungrounded and provably false.

:And thirdly, it has not been "only been reported by certain media outlets such as Engadget" . It has been reported by many non-tech reputable and authoritative sources such as [https://www.forbes.com/sites/lanceeliot/2023/05/25/latest-generative-ai-boldly-labeled-as-constitutional-ai-such-as-claude-by-anthropic-has-heart-in-the-right-place-says-ai-ethics-and-ai-law/?sh=3baf74d83064 Forbes], [https://www.nytimes.com/2023/07/11/technology/anthropic-ai-claude-chatbot.html New York Times], [https://www.reuters.com/technology/alphabet-backed-anthropic-outlines-moral-values-behind-its-ai-bot-2023-05-09/ Reuters], [https://time.com/6295523/claude-2-anthropic-chatgpt/ Time], [https://timesofindia.indiatimes.com/gadgets-news/google-backed-anthropic-draws-constitution-for-its-ai-chatbot/articleshow/100131106.cms The Times of India], Business Insider's [https://www.morningbrew.com/daily/stories/2023/05/23/constitutional-ai-gets-funding Morning Brew], [https://www.vox.com/future-perfect/23794855/anthropic-ai-openai-claude-2 Vox], [https://www.theaustralian.com.au/nation/new-ai-chatbot-claude-2-brings-ethics-to-artificial-intelligence/news-story/f388ac739752b922f7f702459943209f The Australian] and [https://observer.com/2023/07/anthropic-ceo-ai-safety/ Observer]. And even if it was primarily covered by tech-centered media, why would it be unacceptable that an innovation in a tech field is primarily covered by a tech-centered journals? I think it would only be fair not to expect CNN to report on video games and IGN not to report on the current presidential run of a certain candidate.

:I think all the evidence provided above are more than enough to factually dispel the presented concerns and prove them to be unfounded. Luroe (talk) 04:44, 23 August 2023 (UTC)

:Merge. There are some reliable sources mentioning the AI, but they all seems to be primarily about Anthropic, not the AI specifically. The article is also written entirely by a user named Luroe, who may be affiliated with Athropic.

:: Opinion : The Concept and technique of "AI Constitution" or "Artificial Intelligence Constitution" are very exciting and very important in the field of artificial intelligence and ethics, and has high potential to become a general research area. But the problem of this article focusing too much on Anthropic perspective make it look like advertising. However the AI Constitution is very new technique and concept, it has only less than 1 year, so we need to be patient and wait more other party research and develop on this area. Cloud29371 (talk) 07:01, 23 August 2023 (UTC)

:::Then the article should discuss a general problem and cover a range of proposals, rather than promoting a single concept from a start-up. Some exemplary articles can be Ethics of artificial intelligence and Existential risk from artificial general intelligence. 桃花影落飞神剑 (talk) 13:55, 23 August 2023 (UTC)

:Deletion. Firstly, the statement "165 citations, which puts in the top 1% of scientific papers ever written" is an over-promoted way to describe academic papers. There are tons of machine learning papers that have such numbers of citations. Secondly, the concept proposed by this start-up has not yet been recognized by any main-stream scientific society in their outlines, such as AAAI, IEEE, or ACM. Thirdly, it was not accepted by any countries as their regulations even laws. It received media reports mainly because it is a new start-up got lots of funds.--桃花影落飞神剑 (talk) 13:29, 23 August 2023 (UTC)

::Keep. This is an absurd, arbitrary and frivolous reason for deletion. It was stated that it supposedly lacks academic recognition, I have provided a counter-argument and proved it wrong by using the numbers of citations which is incredibly high (1% of scientific papers, which already disproves the notion that there are "tons" of papers like this, there aren't: it's 1%). Secondly, we cannot expect for AAAI, IEEE, or ACM to unanimously endorse or adopt this method this quickly, it does not happen this fast. To say that this method is not worthy of its own article considering its already an industry standard with two out of three frontier large language models using it: OpenAI with its use of RBRM and Claude with its use of eponymous CAI (we currently do not know what kind of alignment techniques Google's Bard, the third frontier model, uses and I give myself no liberty to assume use of any specific AI alignment technique without proper evidence of which we currently have none) would be incredibly cavil. Thirdly, we can say the same, or perhaps even substantially more, about laws and regulations of different countries, laws are infamously slow in their development, passing and adoption. As to my knowledge there are NO existing laws that concern the AI Alignment sphere and VERY few laws on AI in general in the whole world due to how new and innovative the whole field is, especially the field of AI alignment. All of the current proposals for AI laws are incredibly high-level and do not discuss or mandate/order the usage of any particular technique. I do not think that it is reasonable to demand for an article's subject to be mandated by law lest it be deleted from Wikipedia, that is an extremely high bar for a Wikipedia article of any kind.

::-

::Thirdly, the assumption that all of the articles written on Constitutional AI method are written so "because it is a new start-up got lots of funds" is simply baselessly untrue, more than that: provably not true. If you'd actually read the all of the sources that I have provided both here and in the article itself you would see that they do mention that it is a startup (which should not be a reason for immediate dismissal of the notability of their accomplishments) that got funded by Google, but the main part of the article after that backdrop is presented is either about the Constitutional AI method or about both Anthrophic and the Constitutional AI method that they pioneered. Let us substantiate our claims and prove this factually. Lets start with a [https://www.forbes.com/sites/lanceeliot/2023/05/25/latest-generative-ai-boldly-labeled-as-constitutional-ai-such-as-claude-by-anthropic-has-heart-in-the-right-place-says-ai-ethics-and-ai-law/?sh=3baf74d83064 Forbes] article on the Constitutional AI method of alignment. The fact of funding is mentioned once in the 10th paragraph as a quick sidenote. The rest of the article is strictly about the Constitutional AI method. You can go and check that yourself. Lets take another article: one from [https://arstechnica.com/information-technology/2023/05/ai-with-a-moral-compass-anthropic-outlines-constitutional-ai-in-its-claude-chatbot/ ArsTechnica], it doesn't mention the funding at all, and is also about the method of Constitutional AI alignment. Once again, feel free to fact-check my claims. And lets take a third article: the one from [https://www.nytimes.com/2023/07/11/technology/anthropic-ai-claude-chatbot.html New York Times]. The funding is mentioned, mentioned twice, first as background information, secondly as part of more of a detailed look at its past. The rest of the entire article is about Anthropic and its method of Constitutional AI, there are 38 paragraphs in this article, only two of them are about funding, 9 are about the method of Constitutional AI. It is clear which one is more important between the two. You can go and check all that yourself. To conclude. Constitutional AI has received substantive coverage in reputable publications for its technical merits - not just because of funding news, which was presented exclusively as a backdrop.

::-

::To conclude, the standards that you've set are incredibly high and would put around 80% of Wikipedia articles at risk of deletion. Your reasons for deletion are arbitrary, absurd and frivolous, for instance the fact that it hasn't been accepted as part of regulatory norm is true, but not a good reason to warrant deletion, as it does not prove the lack of notability. Please, read the pinned primers. None of your reasons align with Wikipedia's reasons for deletion. And secondly, the article in question DOES meet all the necessary criteria for existence. It IS notable enough, as proven by a number of reliable secondary sources reporting on it in detail. Moreover, in the Wikipedia primer "Help:My article got nominated for deletion!" it is stated that "On Wikipedia, the general inclusion threshold is whether the subject is notable enough for at least two people to have written something substantive (more than just a mention) about that subject that has been published in a reliable source." I believe i provided more than that. That should be more than enough to prove its notability. I believe if we were to follow your procedures the Wikipedia would be reduced substantially few article, which is not exactly a good thing considering its not a place for only extremely important things, but for things notable enough, and not notable to the highest degree there is.

::I think it is clear that the article should stay as there are no objective grounds for deletion. The grounds that have been presented here do not conform to Wikipedia standards, and there are more than enough notability and recognition for article's subject. Luroe (talk) 16:58, 23 August 2023 (UTC)

• Delete We're not a PR host; other companies have to take up and agree to the same thing, and it's not even an original work by any means, just taking things from other mission statements which are themselves PR creations. Nate • (chatter) 16:40, 23 August 2023 (UTC)
• :Keep. I'm sorry but I do not see how this is a PR article by any means at all, it describes the AI Constitution as a concept, including its background and technical details. The language itself isn't even alluding to any PR. Luroe (talk) 17:00, 23 August 2023 (UTC)

:::Comment It's one company putting out something to try to gain publicity for something they wrote with no input from anyone else. It's the very definition of public relations. Nate • (chatter) 18:11, 23 August 2023 (UTC)

::::Please stop implying and insinuating that I am working for or in any way related to Anthropic or that I wrote this article with Anthropic's direction. I am in NO way related or affiliated with Anthropic. I wrote this article because, as I already said previously here, there are currently two AI Alignment techniques that are used to align frontier LLMs: RLHF and CAI. I noticed a fact that there exist an article on only one of them, while the other technique is left unshown, I decided to change that fact, as I believe it is obviously notable enough and important enough in the AI Alignment subfield to be allowed to have its own article. That is it. No malice or foul play here.

::::-

::::And if you do want to accuse me of such foul-play please, provide at least some evidence that I have at least tangible or even circumstantial connection to that company or anyone who has ever worked there. I assure you you will provide none, I am more than certain of that, as certain as a man can be that you wont be able to do it because there is none. There is no malice here. It is just an important development in AI alignment methods that warranted its own article for that. That is it. I am not an Anthropic employee or contractor or PR manager or marketer. Just a person that is interested in AI safety and noticed a flaw and decided to change it. I have no other idea how can I prove that I am not affiliated with them or that I have no ill-intent behind my article. I just wanted to fill the gap that I saw in Wikipedia in the area that I am into, that's it. Luroe (talk) 14:38, 24 August 2023 (UTC)

:::@Luroe: please do not have multiple bold !votes in a single discussion. You already voted keep above, so I'd suggest debolding and striking out this second one. Comment as you like, within reason, but just a single !vote; see WP:AFDFORMAT for the general guidelines. Thanks. Skynxnex (talk) 21:33, 23 August 2023 (UTC)

• Delete This is just promotion by Anthropic, the idea hasn't caught on outside the company. Tercer (talk) 12:49, 24 August 2023 (UTC)
• :First of all, can you provide any evidence that it is just a promotion by Anthropic and not something else? That is a very strong and serious assumption, and I believe that such strong assumptions and accusations must be followed by at least some evidence backing them.
• :Like i said previously in this thread before, I DO NOT work for Antropic's PR team! NOT affiliated with them in any kind, have not met a single Anthropic employee in my entire life and have absolutely no, even circumstantial relation to them in any single way except maybe by being interested in the AI safety subfield of science as well. I made this article because as I said previously in this thread I have noticed that there exists a gap in the Wikipedia's list of AI alignment methods and decided to fill it. That is it, no foul play or any hidden agenda. That's it. I genuinely don't know how else I can prove that I don't work for Anthropic and that I really wrote this article because I really do believe that such an important and notable advancement should have its own, even if short, article. That I truly did it because I honestly thought that such a notable, at least in the AI safety sphere, method deserved its own article, and not out of some hidden commercial malicious intent.
• :-
• :Secondly, as I already mentioned in my previous reply in this thread, the idea of an AI Constitution have caught outside the company, especially considering the fact that it existed in its rudimentary and experimental form in Google DeepMind's Sparrow and now also exist in OpenAI's GPT model in the form of RBRM.
• :-
• :And thirdly, it doesn't have to catch on outside the company to be deemed notable enough to allow the existence of its own article. As I already shown before with the amount of sources there is enough notability for such an article to exist. That is an arbitrary requirement. Luroe (talk) 15:13, 24 August 2023 (UTC)
• Delete PR dressed up with a spurious footnote to Asimov is still PR. XOR'easter (talk) 17:18, 24 August 2023 (UTC)
• :How is it PR? It's a working and notable AI Alignment method as proven by the number of secondary reliable sources and the fact that it has been used by DeepMind and OpenAI. It definitely deserves its own article.
• :As I have previously stated here multiple times before, I am in no way related or working with the direction of Anthropic. I am an unrelated party that just noted the absence of this obviously quite a notable technique on Wikipedia and decided to change that fact, i had no intention of marketing something and you can see it in the article itself if you read it. The article itself is simply a description of what that method is, how it works, what problems and critiques it has, along with background. The core content of the article provides a neutral, factual overview of an emerging AI technique of Constitutional AI, all based on reliable sources. How is it necessarily PR? Is it PR because it describes a concept/method that is clearly notable and an important development in the sphere of AI safety? How exactly can I make it not PR? Not to write about it? But it's clearly notable enough to allow its existence on Wikipedia, and, in my personal opinion, notable enough not only to allow its existence on wikipedia, but to warrant it.
• :The reference to Asimovs three laws of robotics that are present in the background section are there to demonstrate that the concept of encoding principles into AI systems does have historical precedents and did not just originate from Anthropic. This is done to put the Constitutional AI method in the broader context of the field of AI safety research. The goal of the page was never promote Anthropic itself, but to document the academic concept and current research around Constitutional AI. The technique has received substantive coverage for its technical merits, not just PR. Luroe (talk) 07:41, 25 August 2023 (UTC)
• ::It's not a good idea to reply to every 'Delete' comment, especially if you are repeating arguments you made previously. See WP:BLUDGEON. HenryMP02 (talk) 02:39, 30 August 2023 (UTC)
• Leaning delete. This reminds me of the Tesla "Master Plan" article. While it has some substance (reinforcement learning using LLM feedback rather than human feedback) I'm not convinced it needs its own article, or that it should be presented as an Anthropic invention rather than a generic concept. DFlhb (talk) 08:06, 28 August 2023 (UTC)
• Merge to Anthropic. This article is clearly promotional to be honest, the arXiv ref it overly relies on seems WP:PRIMARY and is not considered a WP:RS anyway. Many of the other articles do not even refer to the subject, they discuss "Constitutional AI" instead which is basically a marketing term for the company. However some of the material and refs can still be salvaged to improve the company's article, which is the more appropriate place for this to be covered at present. - Indefensible (talk) 05:55, 29 August 2023 (UTC)

• Keep, remove promotional tone: Certainly seems to meet WP:GNG:{{source assess table

| user=HenryMP02

| src1 = [https://arxiv.org/abs/2212.08073 Constitutional AI: Harmlessness from AI Feedback]

| i1 = n | ij1 = Paper written by Anthropic

| r1 = | rj1 =

| s1 = | sj1 =

| src2 = [https://techcrunch.com/2023/05/09/anthropic-thinks-constitutional-ai-is-the-best-way-to-train-models/ Anthropic thinks ‘constitutional AI’ is the best way to train models]

| i2 = y | ij2 = Article by tech journalist; he doesn't appear to be affiliated with the company

| r2 = ~ | rj2 = TechCrunch; See Wikipedia:Reliable sources/Perennial sources#TechCrunch.

| s2 = y | sj2 = Whole article is about the topic

| src3 = [https://www.marketingaiinstitute.com/blog/anthropic-claude-constitutional-ai How Anthropic Is Teaching AI the Difference Between Right and Wrong]

| i3 = n | ij3 = Author uses "content marketing... to raise revenue for Marketing AI Institute"

| r3 = | rj3 =

| s3 = | sj3 =

| src4 = [https://www.theverge.com/2023/5/9/23716746/ai-startup-anthropic-constitutional-ai-safety AI startup Anthropic wants to write a new constitution for safe AI]

| i4 = y | ij4 = Independent reporter

| r4 = y | rj4 = The Verge; Green at Wikipedia:Reliable sources/Perennial sources#The Verge

| s4 = y | sj4 = Partially about the company, but covers the topic in detail

| src5 = [https://www.anthropic.com/index/claudes-constitution Claude's Constitution]

| i5 = n | ij5 = Anthropic

| src6 = [https://www.nytimes.com/2023/07/11/technology/anthropic-ai-claude-chatbot.html Inside the White-Hot Center of A.I. Doomerism]

| i6 = y | ij6 = Independent reporter

| r6 = y | rj6 = New York Times

| s6 = y | sj6 = Several paragraphs about the topic

| src7 = [https://www.forbes.com/sites/lanceeliot/2023/05/25/latest-generative-ai-boldly-labeled-as-constitutional-ai-such-as-claude-by-anthropic-has-heart-in-the-right-place-says-ai-ethics-and-ai-law Latest Generative AI Boldly Labeled As Constitutional AI Such As Claude By Anthropic Has Heart In The Right Place, Says AI Ethics And AI Law]

| i7 = ~ | ij7 = Forbes "contributor", {{strikethrough|appears to be independent}}. Edit: "AI Consultant", hmm...

| r7 = n | rj7 = {{strikethrough|Forbes}} Third-party "contributor", questionable oversight

| s7 = y | sj7 = Most of the article is about the topic

| src8 = [https://arstechnica.com/information-technology/2023/05/ai-with-a-moral-compass-anthropic-outlines-constitutional-ai-in-its-claude-chatbot/ AI gains “values” with Anthropic’s new Constitutional AI chatbot approach]

| i8 = y | ij8 = Reporter who appears to be independent

| r8 = y | rj8 = Ars Technica; see Wikipedia:Reliable sources/Perennial sources#Ars Technica

| s8 = y | sj8 = Article is about the topic

| src9 = [https://www.wired.com/story/anthropic-ai-chatbots-ethics/ A Radical Plan to Make AI Good, Not Evil]

| i9 = y | ij9 = Independent writer

| r9 = y | rj9 = Wired; see Wikipedia:Reliable_sources/Perennial_sources#Wired

| s9 = y | sj9 = Article is about the topic

}}

Stopped after the 9th source, when it seemed clear to me that the subject is notable. However, the article does seem like it written like an advertisement. This promotional tone should be fixed. HenryMP02 (talk) 02:24, 30 August 2023 (UTC)

:I think a problem with your assessment is that not a single reference you evaluated as meeting actually specifically calls out the subject by name, they are discussing "Constitutional AI" but not "AI constitution" which are seemingly related but different subjects. "Constitutional AI" is just a marketing term for Anthropic. None of these sources actually cover "AI constitution" which is what the article is supposed to be for. - Indefensible (talk) 05:05, 30 August 2023 (UTC)

::In my opinion, both names refer to the same topic. HenryMP02 (talk) 06:47, 30 August 2023 (UTC)

{{clear}}

:The above discussion is preserved as an archive of the debate. Please do not modify it. Subsequent comments should be made on the appropriate discussion page (such as the article's talk page or in a deletion review). No further edits should be made to this page.