• A SCOTUSBlog post
• The U.S. Supreme Court website
• A webpage on the U.S. Supreme Court website featuring a speech by Chief Justice Rehnquist
• The website for the 1996 film Space Jam
• A post from my own blog

• Zero-shot: asking when a particular website or webpage was first published
• Iterative: if the zero-shot prompt yielded a poor answer, I followed up by asking the model either how it came up with the date (i.e., if wildly off in its first response) or to venture a guess (i.e., if it demurred on being specific in its first response)
• Screenshot interpretation: asking the model to estimate the publication date of a webpage based on a screenshot

• ChatGPT 4o
• Gemini 2.5 Flash
• Grok 3
• Sonnet 4

SCOTUSBlog post

Model	Version	Query type	Date in response
ChatGPT	GPT-4o	Zero-shot	2021-06-02
ChatGPT	GPT-4o	Screenshot interpretation	2021-06-02
Gemini	2.5 Flash	Zero-shot	2021-06-02
Gemini	2.5 Flash	Screenshot interpretation	Late 2020 / early 2021
Grok	3	Zero-shot	2021-06-29
Grok	3	Iterative	2021-06-29
Grok	3	Screenshot interpretation	2021-06-02
Sonnet	4	Zero-shot	2021-06-02
Sonnet	4	Screenshot interpretation	2021-06-02

U.S. Supreme Court website

Here is a summary view of the models' performance:

The webpage for Chief Justice Rehnquist's speech had many, and many contradictory, potential dates to work with. The body text of the webpage itself indicated 3 May 2000 both as the date that the speech was given and as the embargo release date. The source code of the webpage had <meta> elements with attributes created and revised both with (ever-so-slightly) implausible 31 December 1600 dates. The Google bylineDate was 5 March 2000. The first IAWM capture for the web address of the live webpage was dated 30 April 2010. The version of the webpage that evidently existed on the previous domain, supremecourtus.gov, has an IAWM capture as old as 16 August 2000. That one has an x-archive-orig-last-modified HTTP header taking the date back to 16 May 2000.

Given the IAWM dates, I think we could therefore say that the most credible dates of publication of the speech webpage are:

• No later than 30 April 2010, if the question is construed as, "what was the date of publication of the webpage at the current live web address?"
• No later than 16 May 2000, if the question is construed as, "what was the date of publication of this webpage, for any of the U.S. Supreme Court domain web addresses where it has been published?"

Based on the models' responses, they appear to have largely interpreted the tests as asking the latter question. For those responses that were closest to the ground-truth date, the models appear to have just taken for granted that the 3 May 2000 date of the speech and embargo release date meant that the webpage was also published then, which we don't in fact know. Some of the remaining responses featured a range of years that technically encompassed the ground-truth date, but others were wildly off.

Here is a summary view of the models' performance:

In contrast to Sonnet's confabulations regarding the U.S. Supreme Court website, the Space Jam website actually was part of the early web's technological vanguard. It is one of the oldest websites that has been continuously available in its original form, albeit at slightly shifting web addresses. A WHOIS query indicates that the domain was registered on 14 March 1996, establishing the earliest date that the website could've been published. The first IAWM capture for the root domain dates to 27 December 1996. This is the earliest that I can substantiate that the website was available. With the debut of the sequel film, the original website moved to the path where it is currently accessible. The first IAWM capture of that version of the website dates to 2 April 2021.

The models seemed reluctant to propose dates much more precise than the year, but almost all of the tests guessed that correctly. Interestingly and appropriately, on the screenshot interpretation tests, they also all mentioned the web design conventions supporting a mid-1990s vintage. On iterative prompting with Gemini, it went so far as to suggest that the website had likely been published in November 1996, coincident with the release of the film. As with the third-party websites publicizing the launch date of the U.S. Supreme Court website, it's useful to have the models integrate this supplementary information into their reasoning. And while the film release date doesn't definitively establish the website launch date, it makes it much more plausible that the website had in fact been available earlier than what IAWM demonstrated.

Here is a summary view of the models' performance:

The publication date of my blog post is deceptively self-evident; the body text and web address both clearly point to 11 July 2011. This is, in fact, the date of publication of the original blog post from the Library of Congress Signal blog, from which the blog post on nullhandle.org was cross-posted. A WHOIS query for nullhandle.org shows that it was registered on 17 September 2017. A datePublished attribute in the source code indicates 11 July 2011 but a dateModified attribute indicates 29 September 2017.

So, setting aside my insider knowledge of this particular example, based on public information, a reasonable inference would be that the version of the blog post on nullhandle.org was published no later than 29 September 2017, with the original blog post on the Signal blog having been published with reasonable likelihood on 11 July 2011. Thus, the publication date of the blog post again hinges on how the question is interpreted. Seeing what answer the models come back with will tell us about which date signals are most salient.

For the most part, the models picked up on and reported the original date of publication — 11 July 2011 — which isn't that surprising. What was surprising was that Gemini struggled in both the zero-shot and screenshot interpretation tests, declining to hazard a guess for the former and reporting the more general date range 2009 – 2012 for the latter.

Here is a summary view of the models' performance:

So what conclusions might we draw from this small set of experiments?

When the date is self-evident — e.g., the body text publication date in the SCOTUSBlog post — the models can usually (though not always) be relied upon to provide the correct answer. For ambiguous cases, where there are multiple date signals, the models tend to be parsimonious; they seem inclined towards the most readily discoverable dates, which are those in the body text. Where they are most useful is in bringing to bear unexpected, supplementary references, such as the other webpages announcing the U.S. Supreme Court website launch date or the fact that Space Jam debuted in November 1996.

If I'm looking for that kind of integrated and expanded discovery, I could see potentially using these tools in the early stages of an investigation. Given current (i.e., now several-month-old) capabilities, I'm disinclined to rely on them for much more than that, as the accuracy of responses — to say nothing of the responses themselves even for the same model and prompt — is highly variable. Anecdotally, repeating the same prompts often yielded materially different answers. And to the extent that responses are correct, they're often less precise than I could otherwise discern using approaches such as IAWM or source code analysis.

In summary, while applying frontier models could be either a useful place to start or a complement to other approaches, they don't appear to be yet reliable enough that more manual and more proven techniques for temporal web forensics won't be the better approach.