The Old Diocesan Issue 13 - Magazine - Page 56
HOT TOPIC
company policy to ignore obviously
AI-generated job applications.
The existential questions will
not disappear. And they absolutely
shouldn’t. But while governments,
philosophers and investors argue
about where AI may eventually
take humanity, the practical
question that the rest of us have
to answer has evolved somewhat.
What are we still doing by hand
that a machine could do – so that
we can get back to the work that
only a human can do?
Tim Richman
(1995B) is a
publisher and
author, and is
the editor of
The Old Diocesan.
HOW WE WROTE THIS ARTICLE
Before Chris and I met up for the
first time, we agreed to use this
feature as a way to test the ability of
current AIs to write feature articles
that don’t read like AI slop.
I have had a $20 ChatGPT
subscription for 18 months, which
I’ve used to assist with writing and
editing fairly regularly, but based on
both Chris’s specific advice and its
recent surge in popularity, I decided
to try Claude as well. I subscribed to
Claude and the Granola transcription
app in advance of our meeting.
Our 70-minute interview at The
Mitre was transcribed by Granola.
As a backup, I also recorded it
on my iPhone’s Voice Memos app
(standard interviewer’s paranoia),
but Granola worked a charm.
It generated accurate meeting
notes in seconds, and stored a
full transcript for future reference.
Importantly, I had to link Granola
to both AIs so they could access its
data, a functionality point I had not
crossed before. Note that I didn’t
even read the interview transcript
– half an hour already saved.
A couple of hours later, having
spent some time thinking over our
conversation, I then gave identical
instructions to both ChatGPT 5.6
and Claude Opus 5 to write a first
draft of the article. Within each AI,
I created a “Hot Topic” project and
uploaded five context documents:
a recent LinkedIn post from Chris
outlining 10 thoughts on AI usage;
three samples of my own writing
from past issues of The Old Diocesan;
52 | THE OLD DIOCESAN
and the previous “Hot Topic” article
on Chinese cars by Brandon de Kock.
I then wrote a 300-word prompt
requesting a 2,000-word article,
and within minutes the AIs had
delivered. Claude’s attempt, at
around 2,200 words, was a passable
feature on par with many first-draft
articles I receive. It also included
an excellent sidebar – a first draft
of “Getting Started” on page 51,
which it thought to write based
on comments from the interview
transcript, not my prompt. But
the draft failed in one critical
area: it veered substantially
from my writing voice.
The ChatGPT version was
notably superior, a much closer
approximation of my tone and style.
This appears to be the result of my
history working with ChatGPT,
which, in its own words, gives it
“a general contextual familiarity
with [my] preferences and work”.
Its word count was also almost
exactly to brief, no doubt the result
of a specific training session six
months ago when I instructed it
to count words more accurately –
not a straightforward thing for AI.
The speed with which I’d managed
to create a first draft allowed me far
more time than usual to mull over
the story, and over the course of
a week I added various thoughts
to a Word document. These included
my impressions of the Kwanda
workshop, which took place after
the first draft had been created.
I then asked ChatGPT to integrate
my notes into its version of the
draft, again using a detailed prompt.
The result was seamless. And for
the next draft, I got stuck in myself,
reworking and rewriting in good oldfashioned human form. AI remained
up my sleeve to check the transcript
for a fact or a quote, which was
a fantastic time-saver – one I’ve
banked for the future.
The final, critical check was Chris
and Kieran’s review of the article for
facts and quotes. This is standard
procedure, but in this instance I was
more wary than usual that a glaring
factual inaccuracy or hallucination
may have slipped through. There
were, it turns out, no clangers.
So, was this article better or worse
than it might have been without the
extensive AI assistance? It’s certainly
a long way off the slop that currently
abounds online. And where it may
fall down occasionally line by line,
it has given me more room and
time for strategic thinking, which
has benefited both it – with the late
inclusion of the diagram on page 49,
for instance – and the overall issue.
As for the opening illustration,
that was 15 minutes on ChatGPT
and Claude, with each AI prompted
to produce one illustration and
one “photograph”. The ChatGPT
illustration was a clear winner;
our designer thought I’d sourced
it from an image bank. Note the
error, though: it used a table-tennis
bat rather than a tennis racquet.
We could have changed it, but
we felt it worked.
