Brian, That's a wonderful demolition of ChatGPT as a source of knowledge. I did wonder whether some of the false facts could result from double machine translation. Also, if you want to find an association between "Ludgate" and "analytical engine" they are in adjacent sentences of the first paragraph of your own 'Ludgate's analytical machine of 1909', and machine learning algorithms have their own ways of forming associations. How the "No. 2" crept in is unclear, but again associations are not hard to find: "Difference Engine No. 2 and the Analytical Engine share the same design for the printer ..." [https://www.computerhistory.org/babbage/engines/]. So it's as if ChatGPT picks up such textual associations and turns them into logical associations. Actually it's quite easy to get false technology history out of ChatGPT. The first time I ever tried it, it said "IPv6 is designed to be backward-compatible with IPv4". That is not true and was never a design goal of IPv6. In fact, it seems almost trivial to make ChatGPT spout nonsense. This is a matter of considerable concern. Regards Brian Carpenter On 09-May-23 09:00, Brian Randell via Members wrote:
Hi:
Motivated by the extensive publicity and excitement about the new AI chatbot ChatGPT, a tool for answering natural language queries across numerous domains, I and colleagues undertook a brief assessment of its ability to answer questions on a computer history topic. To our astonishment ChatGPT responded with 48% fake facts to the twelve questions we posed it! A brief account of our assessment has just been accepted for publication by the IEEE Annals of the History of Computing. The full details of the assessment are to be found in the report “ChatGPT’s Astonishing Fabrications about Percy Ludgate” by Brian Coghlan, Noel O'Boyle and myself - available online at https://tinyurl.com/383pbuwb .
Cheers
Brian Randell
_______________________________________________ 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 http://lists.sigcis.org/pipermail/members-sigcis.org/ and you can change your subscription options at http://lists.sigcis.org/listinfo.cgi/members-sigcis.org