IEEE Technology and Society Magazine - September 2023 - 81

segments of text containing specific keywords
and label them with respective codes. For
instance, when the keyword " privacy " is inputted
using the autocode function, MAXQDA will highlight
all the search hits of text " privacy. " One may
then manually determine whether to highlight
only the word " privacy " or segments of documents
containing the word " privacy. "
2) The autocode function only provided a coarse
screening as a first step of the coding process. It often
returned repetitive or mislabeled information.
For instance, a search for " race " would also return
text containing " embrace " and " trace. " In addition
to that, polysemous words also interfered with the
search results. A perfect example would be the
keyword " language. " Inputting " language " into the
autocode search engine would yield both results
related to " human language diversity " and " natural
language processing. " While those two types
of content could be relevant to AI ethics at the
same time, they by all means should not be categorized
under the same code label. Therefore, it
became necessary to manually check the search
results to ensure that all the keyword-containing
contents were coded correctly and accurately.
3) In addition to the coding system based on ethical
principles, another layer of coding mechanism
was also established to further categorize the
content. Eight modal verbs were used as search
keywords for this coding mechanism: shall,
should, can, could, will, would, may, and might.
The goal of this modal verb coding system was
to differentiate between segments of documents
that were calling for initiatives and those describing
existing actions. While utilizing the autocode
function, the search range was limited within
coded segments from Step 2. The purpose was
to find out the intersections of prelabeled coded
segments in Step 2 and segments containing the
modal verbs.
After the initial stage of coding, 659 coded segments
belonging to 21 subprinciples were generated
from the datasets. As additional documents
were added to the data pool, search terms were also
refined by modifying the tense and including more
synonyms to better describe their parent categories.
In the end, a total of 11,060 coded segments were generated
for the 21 subprinciples based on the coding
system. The process of generating theoretical coding
September 2023
using a GT framework involves several rounds of coding
and analysis. In the first round, initial coding, the
data is broken down into smaller segments and each
segment is given a descriptive code. These codes are
then sorted and grouped together into broader categories
in the next round as selective coding. This
involves identifying the key themes or concepts that
emerge from the data and grouping them into theoretical
categories or codes. These codes are then refined
and further developed through ongoing analysis and
comparison with new data.
To summarize, the central categories and main
codes of these rounds of analysis are illustrated in
Figure 3.
In the initial rounds of coding, we drew upon
literature and generated a few codes, including
" privacy, " " confidentiality, " " bias, " and " diversity. "
From these categories, selective coding was conducted
involving constant comparison throughout
the process with literature on AI principles in healthcare
through institutional and ethics lenses. The
initial categories were expanded to better describe
the six major ethical principles. More codes were
developed for each category with the goal of more
precisely extracting informative content from the
Figure 3. Selective coding process.
81
IEEE Technology and Society Magazine - September 2023

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - September 2023
