Re: [xsl] ChatGPT results are "subject to review"

["Michael Kay mike@xxxxxxxxxxxx" <xsl-list-service@xxxxxxxxxxxxxxxxxxxxxx>]

I find this all very scary. If it gets it right 80% of the time, that's fine,
people will test it carefully before putting it into production. It's when it
starts getting it right 99% of the time that we should start worrying: people
will get overconfident and we will have some nasty accidents as a result.

And the more it gets the easy things right, the more we will lose the skills
to see when it's getting the harder things wrong.

Michael Kay
Saxonica

> On 6 Jul 2023, at 16:45, Dorothy Hoskins dorothy.hoskins@xxxxxxxxx
<xsl-list-service@xxxxxxxxxxxxxxxxxxxxxx> wrote:
>
> I recently ran a small exercise in ChatGPT. I provided it a sample input and
output that would require it to put text from a footnote in XML input and
insert it a popup-enabled span in the HTML output where the footnote reference
number appeared in the source XML.
> The first stylesheet generated in the results returned was written in XSLT
1, as I forgot to prompt for an XSLT verion.
> I asked ChatGPT to rewrite in XSLT 2, taking advantage of any XSLT 2
features that would improve the efficiency of the transform, which it did and
explained what it did.
> Then I asked for a version 3 XSLT and got that with some new function, so
now I can compare how the same transformation concept was treated in the XSLT
versions.
> ChatGPT included comments in each XSLT and its own explanation for the
templates in the text thread about the XSLT. Seems like a good learning tool,
although all generated code must be reviewed and tested to see if the results
are indeed the same output.
> Along the way, I saw that ChatGPT had thrown in a concat() when building a
text string for an attribute. When I prompted ChatGPT to explain why it
included the concat(), it apologized and said the concat() wasn't necessary,
regenerated the previous XSLT and explained its reason for the change.
> Obviously, a person who doesn't look closely at the generated code, or
doesn't know much XSLT, might not have seen the concat() as an issue. So user
beware.
> From what I understand, I could try the exact same prompt with sample inputs
and outputs and might be given a different result, but I haven't done that
investigation yet.
> Other observation from trying out XML in ChatGPT: it can check
well-formedness but had problems discerning the XML declaration and the root
element when there wasn't a line break between them, so it incorrectly stated
that there was no root elements and produced a default <document> root and put
all the rest of the XML elements into it to retain the hierarchy. It also
generated a partial XML result with comments about where more XML elements
occurred, rather than providing the entire XML tree in the generated XML it
"corrected".
> It is capable of generating a schema, a schematron and an Xspec from sample
XML, all of which would require testing, but it sure is fast. A quick way to
create stubs for future development.
> I think if someone invested in a training set for a specific schema with a
bunch of examples, it would be a great tool for XML/XSLT development, always
"subject to review".
> Regards, Dorothy
>
> XSL-List info and archive <http://www.mulberrytech.com/xsl/xsl-list>
> EasyUnsubscribe <http://lists.mulberrytech.com/unsub/xsl-list/293509> (by
email <>)