Talk:ChatGPT/GA2

GA review

Article (edit | visual edit | history) · Article talk (edit | history) · Watch

Nominator: Czarking0 (talk · contribs) 23:43, 5 January 2026 (UTC)

Reviewer: WhaleFarm (talk · contribs) 18:52, 15 March 2026 (UTC)

Overall

Initial review of the article against the Good Article criteria.

Lead

The lead contains a large number of citations. Per WP:LEADCITE, citations are usually placed in the body of the article, the lead summarizing material cited elsewhere. Some of the citations, especially the ones on market acceptance and definition of a gpt, seem redundant to ones in the body. I think it would read easier if the lead was a simple summary. — Preceding unsigned comment added by WhaleFarm (talk • contribs) 19:20, 15 March 2026 (UTC)

I removed the sources that seemed less relevant or important, let me know if we should go further. Keeping a source for the "5 most-visited websites globally" statement can be good because ChatGPT is currently at the 5th place which could easily become outdated, so it's good to have a source for verification. Alenoach (talk) 23:04, 15 March 2026 (UTC)

Length and structure

The article is quite long. Some sections (such the "Model version" section, where there is redundancy with the sub-headings) appear to contain a level of detail exceeding summary style. It may improve readability if some of the detail were condensed or moved to a dedicated sub-article, the main page then summarizing the key developments. American tech personas is a sub-heading, but somewhat lost by its distance from the heading. Starting out as in politics was clearer, maybe continue, or “Reception by American tech personas” as possible ideas. — Preceding unsigned comment added by WhaleFarm (talk • contribs) 19:42, 15 March 2026 (UTC)

is within what GA guidelines indicate. I will change the section heading. Czarking0 (talk) 21:46, 15 March 2026 (UTC)

Prose

The article is readable and understandable to a broad audience. It avoids a lot of ML jargon. Some sections, particularly those describing model releases, read more like a release log than like an encyclopedic narrative. A more summarized description of major developments would improve readability. — Preceding unsigned comment added by WhaleFarm (talk • contribs) 19:54, 15 March 2026 (UTC)

Sources

I spot-checked the sources, all seemed well-formed and appropriate. Use of archives is good.. Some sources are older than the language and tense used in the text would suggest. For time-sensitive topics such as the Financial markets section, the prose may need to be written in historical terms, or supported with recent sources when the context implies current relevance. Both may be appropriate. — Preceding unsigned comment added by WhaleFarm (talk • contribs) 20:10, 15 March 2026 (UTC)

Made adjustments to the financial markets section Czarking0 (talk) 21:54, 15 March 2026 (UTC)

Images

Images appear appropriate for the article. However, the Time magazine cover (File:The AI Arms Race Is Changing Everything.webp) may have licensing issues. Magazine covers are typically copyrighted and should be used on Wikipedia under a non-free content rationale, not public domain. This file needs review. — Preceding unsigned comment added by WhaleFarm (talk • contribs) 20:28, 15 March 2026 (UTC)

This image was previously discussed (may have been off wiki or at least I am forgetting where the conversation was) and the individuals I spoke with indicated this particular magazine cover was public domain due it being too simple to copyright. However, I also do not think it is critical to the page and could be replaced. Czarking0 (talk) 21:44, 15 March 2026 (UTC)

I think there's probably a fair use claim for use, maybe that was the prior discussion WhaleFarm (talk) 01:25, 23 March 2026 (UTC)

It was nominated for deletion and kept at Commons:Deletion requests/File:The AI Arms Race Is Changing Everything.webp, the consensus was quite strong that it is not a copyright violation. If you feel like there are still copyright issues, please nominate it for deletion again. -Consigned (talk) 19:43, 11 April 2026 (UTC)

Not my expertise, but seemed creative. I'll go with the consensis WhaleFarm (talk) 20:52, 11 April 2026 (UTC)

Stability

The article appears stable.

Overall

The article is generally well written and covers the topic broadly. Issues noted above relate to structure, sourcing age, and a possible image licensing issue. These are all fixable. I will place the article on hold to allow for improvements.

Model section

@Alenoach: this review does not look favorably on the model history section. As you know, I also do not love it. Given the review what are your thoughts on removing this section or converting to a more brief prose summary.— Preceding unsigned comment added by Czarking0 (talk • contribs) 21:49, 15 March 2026 (UTC)

I think the table is useful to have a quick overview of the model history. It's more efficient and easier to visually parse than prose. I can improve the end of the table, which lacks descriptions and has a "citation needed". I don't know though if niche variants like o3-mini-high deserve a row in the table.

In my view, the problem is the subsections below. Having independent subsections for some of the models like this lacks cohesiveness and is quite redundant with the table. Using excerpts also means we have less flexibility. I suggest removing the subsections (integrating any really important information into the table's "description" column) and maybe adding below the table a summary of the overall trends in model development.

Would be interested to have your opinions on this proposal, Czarking0 and WhaleFarm. Alenoach (talk) 23:56, 15 March 2026 (UTC)

I made some change to the model descriptions in the table. Alenoach (talk) 04:16, 16 March 2026 (UTC)

I removed the subsections. Alenoach (talk) 20:39, 16 March 2026 (UTC)

Overall, I think you make a strong argument for why this should not be prose. I see reason in dropping more niche versions like o1-preview. List of Nvidia graphics processing units has tables that I really like. Though that is a listicle. An alternative proposal would be to make a listicle and then reference it in the article.

I have no issue with your proposal to remove subsections and add info to the table. That could be done either in the article like it is now or in a listicle. Czarking0 (talk) 02:55, 18 March 2026 (UTC)

The "-pro" variants could potentially be trimmed as well. Many recent models have a "pro" variant, so being exhaustive would be needlessly repetitive. And "pro" variants don't seem so different from the rest, it's mostly just that they use more compute for some extra reliability. If we are ready to trim further, a simple rule of thumb could even be "one line for every model that has an article on it".

Having a listicle like "List of large language models by OpenAI" is an option. I think the table is efficient in giving a quick overview of the different models, and since it is a table, readers can relatively quickly see what it is about and skip it if not interested. But I agree that there is place for reasonable disagreement, notably because the table has been getting quite long over time and understanding the differences between models matters less now that the "o"-series reasoning models are deprecated and there are fewer models to pick from (the table has historical value but less and less practical value). I would appreciate more opinions to help settle this either way. @WhaleFarm, do you have a preference on this? I don't know if it should affect the decision, but there is by the way another, less exhaustive table in Products_and_applications_of_OpenAI#Text_generation. Alenoach (talk) 18:19, 22 March 2026 (UTC)

I think the table is now reasonably sized. The description field doesn't say much of use, especially on the newer models. Maybe ther could be a summary of the release note differences. example:

Long-awaited, GPT-5 can either answer quickly like earlier GPT models or reason before answering like the reasoning models of the "o" series. Instead of a single GPT-5 model, there was a network of GPT-5 models with different levels of capability, with a router selecting one based on the complexity of the task and other factors.

long-awaited is a bit loaded, but the "like the reasoning models of the "o"..." tells me nothing if I haven't heard about those modes. Some prose before the table would be nice to get some context. Also, there's nothing about where the models are used. If I go ask a chat, what model do I get? Why should I care? What would the user see as an advantage?

Just some ideas, I understant there's a lot to present here.WhaleFarm (talk) 01:51, 23 March 2026 (UTC)

Agreed. The table also has problems. It includes unsourced/poorly sourced content. It's also a WP:NOTCATALOG issues. The line for GPT-5 is vague and ambiguously promotional, and "long-awaited" says one thing while multiple sources describe a "backlash", which suggest this is useful context. This is lopsided, to put it mildly. Grayfell (talk) 22:09, 23 March 2026 (UTC)

I wrote "long-awaited" because the source also uses that term. There were already a lot of speculations in 2023 about GPT-5 being soon released in late 2023, but it only came out in August 2025. Can drop it if you prefer. Alenoach (talk) 07:30, 24 March 2026 (UTC)

The context around that speculation isn't going to fit in a table, just like how context about the backlash isn't going to fit in a table. We're not trying to tease readers with ambiguities, we're trying to provide plain information. Grayfell (talk) 07:57, 24 March 2026 (UTC)

Sorry I have been very busy with work. Is there an update here? It looks like the above convo was resolved. What still needs action?Czarking0 (talk) 00:19, 24 March 2026 (UTC)

The table is more concise than the prose lists, but it has its own set of problems, such as unsourced information, WP:NOTCATALOG issues, and selective citing of sources to introduce editorializing. Grayfell (talk) 02:46, 24 March 2026 (UTC)

An example of a reference that could use work is the 5.4 model in the table. The description carries little information (Improvements focused on professional work and computer use.). The citation is to a techcrunch article that doesn't add anything to the press release, this is not a secondary source, so just go ahead and use the native link https://openai.com/index/introducing-gpt-5-4/, no gain pretending that techcrunch actually is a secondary source. WhaleFarm (talk) 03:22, 24 March 2026 (UTC)

Using a mish-mash of WP:PRIMARY and secondary sources is not ideal. Since this article is supposed to be about all of ChatGPT, sources about specific versions or minutia about specific features of specific versions gets undue and impractical pretty quickly.

If the table must be preserved it should stick to falsifiable statements from reliable sources. That is going to be mostly WP:IS. Extremely basic info such as release/deprecation/retirement dates could cite a primary source, but anything more than that is a WP:NOTPR issue. Right now several ChatGPT models have their own flimsy articles, so this suggests to me that a spin-off article might make sense (See Category:Software version histories for a few very rough points of comparison). Grayfell (talk) 07:57, 24 March 2026 (UTC)

You guys need to come to some sort of consensus on this or fail the review.Czarking0 (talk) 03:54, 1 April 2026 (UTC)

The worst part has been addressed (the redundant subsections were removed). I see multiple possible further actions:

1. Reduce the number of rows
2. Remove the "Description" column
3. Rework the "Description" column
4. Add a brief overview in prose
5. Replace the table with prose
6. Split the table section to a dedicated listicle article
7. Split the table section to the existing article Products_and_applications_of_OpenAI#Text_generation
8. Remove the section entirely

Czarking0 proposed option 6. I have some preference for just option 1.

We can proceed with option 6 by default unless there are further opinions. What would you vote for, WhaleFarm, Grayfell? Alenoach (talk) 04:27, 1 April 2026 (UTC)

I would reduce the number of rows. Additionally, I think a dedicated listicle article would be of interest, but that's not a GA issue.

A rework of the description is called for. I would think that a sentence or two of key characteristics or changes would be good, with a reference to a site that, ideally, does some analysis.

Some of the table entries cites (for example, GPT-5.4) rely on coverage that closely follows OpenAI’s pressers. Entries such as “improvements focused on…” impart little information. It would strengthen the section to draw more on independent, secondary coverage where available. For instance, The Verge draws mostly from the press release, but at least makes the attribution to OpenAI clear. More analytical reporting (such as Reuters or Fortune) would be preferable, especially for claims about capabilities or significance. I've look for such references, I understand that most of the ddep analysis is from non GA grade sources.

Here's an example entry using Fortune that is RS and more than an OpenAI echo for 5.4 description:

Expanded capabilities in reasoning, coding, and task automation, focused on enterprise use. Uses fewer tokens. Expands direct computer and applications operation (“agentic” AI systems).

Or even start with "Claimed expanded..." although the Fortune article does a good job on the attributions.

On sources in general, my spot check has matched well with the material. High-quality coverage (e.g., Reuters, The New York Times, MIT Technology Review) used appropriately. Some primary and lower-tier sources cause sections to read close to announcement language or lack detail. Expanding the use of independent secondary would improve the article.

I hope this helps WhaleFarm (talk) 15:57, 1 April 2026 (UTC)

Thanks for you patience. I reworked several descriptions to be more specific and to better fit what I found in reliable sources. I didn't mention the improved reasoning or the reduction of hallucinations in the descriptions because it's something that regularly happens with more recent models, so it isn't very distinctive of any particular model. I removed the rows about "mini", "pro" and other variants, since we can't exhaustively list all the variants and the descriptions were often similar and not very useful. Instead, I added a paragraph mentioning what the types of variants are. I also made some relatively minor improvements to other sections.

What do you think? Alenoach (talk) 00:06, 9 April 2026 (UTC)

Here's where I'm at. The table change was a huge improvements, thank you. My remaining reservation:

The Time cover seems troublesome. The licensing justification is weak at best, the layout, concept, formating choice are creative choices. The "no origonal authorship" justification is not credible.

There are quite a few citations of marginal sources, such as techcrunch. As an example in watermarking,

"According to an OpenAI spokesperson, their watermarking method is "trivial to circumvention by bad actors.""

This quote is well attributed to openai, the techcrunch report relies on the same statement, and there is no new analysis. Attributing directly in this case wuold be better, and would avoid looking like an attempt to make a primary source appear like a secondary source. The techcruch summary of the origonal source lost a lot of useful context, maybe this would be better:

In a May 2024 blog post, OpenAI reported developing an unreleased text watermarking method. The company said the method was effective against limited edits such as paraphrasing, but could be bypassed by broader transformations "making it trivial to circumvention by bad actors". [ref] (maybe even something about disproportionately affecting certain users, such as non-native English speakers, which techcrunch did cover)

Take a look at those points. Then I'm ready to weigh in for GA. WhaleFarm (talk) 16:39, 9 April 2026 (UTC)

Sure, I will take a look. For the Time cover, the rationale looks indeed shaky. If the justification doesn't hold, that would be quite annoying because many articles (including in other languages) use that image. Maybe I should request feedback at Village pump/Copyright to see if the image should be deleted. Alenoach (talk) 05:35, 10 April 2026 (UTC)

Or potentially we could rely exclusively on the WSJ source for the paragraph on watermarking (and potentially keep the ref to The Verge which provides a free summary of the WSJ one). The TechCrunch article mainly provides OpenAI's PR response, which we could arguably skip. By the way, OpenAI's original May 2024 blog post didn't cover text watermarking, the content on text watermarking was added on August 4, 2024, the day when the WSJ investigation was published.

Do you think this would be good:

Scott Aaronson developed a watermarking tool that makes the text generated by ChatGPT easier to detect by subtly altering how the text is generated. The watermarking was 99.9% effective on sufficiently long passages and was found not to degrade performance. It was of particular interest for teachers seeking to mitigate cheating. In surveys, respondents favored the release of such a tool by a four-to-one margin, but nearly 30% of users declared that they would use ChatGPT less often if it watermarked outputs and while rival chatbots did not. OpenAI has not deployed the tool.

This would remove the point about circumvention for conciseness, but based on the WSJ source, we could add something like "The watermarking was 99.9% effective on sufficiently long passages and was found not to degrade performance, although it could be circumvented, for example by using Google Translate to convert it to another language and then translate it back." Alenoach (talk) 05:04, 12 April 2026 (UTC)

I prefere the first (The Scott Aaronson based). The 99.9 extension would have qualifying, should really say ..claimed to be..

The 99.9 percent used directly is starting to seem like marketing.

Mainly, my issue was with the techcrunch, which doesn't seem to add anything other than a thin layer of secondary material. WhaleFarm (talk) 13:38, 12 April 2026 (UTC)

Just to be clear, the version 2 I proposed was almost the same as version 1 but with a longer second sentence. I applied version 1 with the "claimed to be". The remaining Techcrunch references in the article don't seem particularly problematic, though they can be replaced if needed.

I also modified the lead's second paragraph, primarily to remove the outdated claim that it's the "fastest-growing consumer software application in history" (Threads is now first). The website ranking sentence is arguably accurate (Semrush and Similarweb currently rank it as the fifth most-visited website, though Cloudflare Radar, which considers all domains, places it as number 10). I wasn't sure whether to keep mentioning it, I removed it because it could quickly become obsolete and the sentence on the number of users already conveyed similar information. Alenoach (talk) 21:42, 14 April 2026 (UTC)

References

moving ahead, POV tag

@Alenoach @Czarking0-

I read the article front to back, and my issues have been addressed for GA. The POV tag shows that @Uhoj doesn't agree, how do we resolve this? Could you explain whatshould be done? Is there now a stability issue? WhaleFarm (talk) 19:00, 17 April 2026 (UTC)

Point four of the the criteria is neutrality. Let's work together to fix it. Uhoj (talk) 20:15, 17 April 2026 (UTC)

I personally would fail this review right now since there are multiple outstanding issues. I can always renominate. If you do not wish to do that I understand. I plan to address all these issues in due time but there is no deadline and an active review does not change that. Czarking0 (talk) 00:47, 19 April 2026 (UTC)

I will fail, this isn't stable. WhaleFarm (talk) 12:43, 19 April 2026 (UTC)

Failed GA

Ongoing content disputes and repeated reversions during review. Does not meet Good Article criterion 5 (stability). Renomination encouraged once stabili. WhaleFarm (talk) 13:15, 19 April 2026 (UTC)