[SIGCIS-Members] On the Merits of ChatGPT and the like as Historian's Assistants

[Adam Hyland]

I found Donald Knuth’s reflections in this topic informative:
https://www-cs-faculty.stanford.edu/~knuth/chatGPT20.txt

My two cents:

The capacity for hallucination alongside what appears to be appropriate
identification of context or nuance casts a veil of fabulism over all the
output, which (IMO) makes models like these unreliable in ways that even
Procopius was not. Given that problem, a model which more often returns
correct information might well be considered *less* trustworthy overall as
an assistant.

On Tue, Jul 18, 2023 at 12:49 PM Brian Randell via Members <
members at lists.sigcis.org> wrote:

> Hi:
>
>
>
> Brian Coghlan and I have been pleasantly surprised by the interest aroused
> by, and the positive comments we have received on, the little paper
> "ChatGPT’s Astonishing Fabrications About Percy Ludgate" that we published
> in the April-June 2023 issue of the IEEE Annals of the History of
> Computing. The study [1] that we reported on was a temporary diversion from
> our main interests, but we are curious to know whether the startlingly high
> level of fabrications (so-called "hallucinations") in the answers we
> obtained from ChatGPT is typical of its performance when used as a
> historian's assistant. It would be great if someone with appropriate skills
> and resources (research students!) could organise and carry out a serious
> statistically-valid evaluation of the trustworthiness of the answers
> provided by ChatGPT and the like to a wide range of  questions on the
> history of computing, carefully checking (i) the existence of any citations
> listed, (ii) the accuracy of all the factual statements made, and (iii)
> whether the errors found were likely to be because of errors in the
> learning data, or due to ChatGPT's text selection and generation strategy.
>
>
>
> One of the motivations we had for our study was the enthusiastic and
> uncritical assessments of ChatGPT's merits as a historian's assistant that
> we found on the web. We have since been amused and embarrassed to find that
> the one article we quoted [2] in our paper was almost certainly generated
> by ChatGPT itself. It turns out that the alleged author, Martin
> Frackiewicz, has been posting on average close to twenty sizeable articles
> a day since mid-March on his company's website in praise of ChatGPT.
>
>
>
> A much more credible favourable account of the merits of ChatGPT and the
> like has been provided by Mark Humphries and Eric Story [3]. This is very
> optimistic about the ability of historians to use such systems in ways that
> achieve high levels of veracity, and to identify such errors as are
> nevertheless made. It would be good if such hopes could be validated - or
> debunked - before large amounts of ChatGPT-generated text on the history of
> computing was published and added to future sets of learning data.
>
>
>
> Cheers
>
>
>
> Brian Randell
>
>
>
> 1. ChatGPT’s Astonishing Fabrications about Percy Ludgate -
> https://treasures.scss.tcd.ie/miscellany/TCD-SCSS-X.20121208.002/ChatGPTs-AstonishingFabrications-aboutPercyLudgate-CoghlanRandellOBoyle-20230424-1434.pdf
> <https://urldefense.com/v3/__https://treasures.scss.tcd.ie/miscellany/TCD-SCSS-X.20121208.002/ChatGPTs-AstonishingFabrications-aboutPercyLudgate-CoghlanRandellOBoyle-20230424-1434.pdf__;!!K-Hz7m0Vt54!hwCqcmqZeYw_ZiA1gvEYeZcwiAuicJxRcaqmrqsshs1rFMcb_eTH4-FNydNlg1cciaPHtA6lKPrxlo51uk7E$>
>
>
>
> 2: ChatGPT-4: A Valuable Tool for Historical Research and Analysis -
> https://ts2.space/en/chatgpt-4-a-valuable-tool-for-historical-research-and-analysis/
> <https://urldefense.com/v3/__https://ts2.space/en/chatgpt-4-a-valuable-tool-for-historical-research-and-analysis/__;!!K-Hz7m0Vt54!hwCqcmqZeYw_ZiA1gvEYeZcwiAuicJxRcaqmrqsshs1rFMcb_eTH4-FNydNlg1cciaPHtA6lKPrxllUwtCeV$>
>
>
>
> 3. Today’s AI, Tomorrow’s History: Doing History in the Age of ChatGPT.
> https://activehistory.ca/2023/03/todays-ai-tomorrows-history-doing-history-in-the-age-of-chatgpt/
> <https://urldefense.com/v3/__https://activehistory.ca/2023/03/todays-ai-tomorrows-history-doing-history-in-the-age-of-chatgpt/__;!!K-Hz7m0Vt54!hwCqcmqZeYw_ZiA1gvEYeZcwiAuicJxRcaqmrqsshs1rFMcb_eTH4-FNydNlg1cciaPHtA6lKPrxln7isyWK$>
>
>
>
>
>
>
>
> —
>
>
>
> School of Computing, Newcastle University, 1 Science Square, Newcastle
> upon Tyne, NE4 5TG
>
> EMAIL = Brian.Randell at ncl.ac.uk   PHONE = +44 191 208 7923
>
> URL =  https://www.ncl.ac.uk/computing/staff/profile/brianrandell.html
> <https://urldefense.com/v3/__https://www.ncl.ac.uk/computing/staff/profile/brianrandell.html__;!!K-Hz7m0Vt54!hwCqcmqZeYw_ZiA1gvEYeZcwiAuicJxRcaqmrqsshs1rFMcb_eTH4-FNydNlg1cciaPHtA6lKPrxlgDUs0H9$>
> _______________________________________________
> This email is relayed from members at sigcis.org, the email discussion
> list of SHOT SIGCIS. Opinions expressed here are those of the member
> posting and are not reviewed, edited, or endorsed by SIGCIS. The list
> archives are at
> https://urldefense.com/v3/__http://lists.sigcis.org/pipermail/members-sigcis.org/__;!!K-Hz7m0Vt54!hwCqcmqZeYw_ZiA1gvEYeZcwiAuicJxRcaqmrqsshs1rFMcb_eTH4-FNydNlg1cciaPHtA6lKPrxlq_2gkLS$
> and you can change your subscription options at
> https://urldefense.com/v3/__http://lists.sigcis.org/listinfo.cgi/members-sigcis.org__;!!K-Hz7m0Vt54!hwCqcmqZeYw_ZiA1gvEYeZcwiAuicJxRcaqmrqsshs1rFMcb_eTH4-FNydNlg1cciaPHtA6lKPrxllcliWn4$
>
-- 
Adam Hyland (*he/him)*
adampunk.com
UW HCDE PhD Student
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.sigcis.org/pipermail/members-sigcis.org/attachments/20230718/48a63272/attachment.htm>