ARTICLE

https://doi.org/10.1057/s41599-024-02904-x OPEN

ChatGPT and the digitisation of writing

Xin Zhao 1, Andrew Cox1 ✉ & Liang Cai2

The aim of this study is to uncover how students’ practices of writing in higher education are
being impacted by ChatGPT. The use of ChatGPT and other generative AI needs to be set in
the context of a longer-term process of the digitisation of writing, where many tools are being
1234567890():,;

employed by students to support writing because it is a complex iterative process. Generative

AI appears to have had a large impact on how students write, and we propose a model of
generative AI literacy to assess their capabilities in doing so. Semi-structured interviews and
observation data were collected at a British University with 23 students from diverse back-
grounds, including the UK, USA, China, Japan, and Saudi Arabia. The data was analysed
thematically. It was found that students used ChatGPT alongside many other tools, and in
rather individualistic ways often to address specific challenges they felt they had with writing.
Their main concerns were around plagiarism, information inaccuracy and technology
dependence. There was a relatively weak understanding or interest in the ethical issues
around the exploitative and environmental impacts of generative AI. The social controversy
around ChatGPT can be seen as a useful opportunity to engage students in a discussion
about the digitisation of writing and promote AI literacy in this context.

1 Information
School, University of Sheffield, Sheffield, UK. 2 School of Foreign Language, NingboTech University, Ningbo, China.
✉email: a.m.cox@sheffield.ac.uk

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2024)11:482 | https://doi.org/10.1057/s41599-024-02904-x 1

ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-024-02904-x

T
Introduction
he use of AI in education (AIEd) has been a discrete area of ● Translation tools;
study for several decades, albeit the majority of studies ● Text generation tools which create bodies of text from a
have been from a technical development standpoint with short stimulus.
less involvement of educators (Zawacki-Richter et al., 2019).
We could add to this other tools and apps used during the
Development of AIEd has tended to be concentrated on Intelli-
research process at the beginning of writing such as for search,
gent Tutoring Systems (Guan et al., 2020). Use of such technol-
text summarisation (e.g., Scholarcy, iris.ai, summarisebot) and
ogies in an educational context has not been without its critics
literature reviewing (e.g., ResearchRabbit, Gecko, connectedpa-
(e.g. Selwyn, 2019). Meanwhile, AI has already made a relatively
pers). There are also well-established tools for referencing which
unheralded appearance in low-level features of much technology
link search and reading to producing a final reference list for a
supporting everyday knowledge work such as search, recom-
completed assignment (e.g., EndNote, RefWorks, Zotero, and
mendation, transcription and translation. It has also appeared
Mendeley).
increasingly within writing support tools, such as grammar
If writing as a process consists of the stages of “prewriting,
checkers, as well as in plagiarism detection.
planning, drafting, revising, and editing” (Strobl et al., 2019, p.
This picture of gradual change was dramatically disrupted in
38), then AWE and AWCF are mostly used in the latter two
November 2022 by the launch of ChatGPT. Particularly in edu-
stages. Our previous research revealed, however, that rephrasing
cation, generative AI has created excitement but is also a con-
tools such as Wordtune are used both in improving text at the
siderable concern (Kasneci et al., 2023, Trust et al., 2023; Lo,
revision stage, but also in breaking through mental blocks in the
2023). The usage figures of ChatGPT show an incredibly rapid
early stages of writing or even planning (Zhao et al., 2023).
rise in popularity and the potential benefits claimed for it are
Similarly, Malik et al. (2023) found Indonesian students also
wide-ranging. Much of the fear has revolved around its potential
using a wide range of AI-based tools in their writing. Translation
impact on academic integrity. What is lacking to date are in-
tools might also be used at various stages, such as in processing
depth studies that explore how ChatGPT is actually used and
reading and drafting text (Zhao et al., 2024). So just as writing is a
experienced by students. Since it is in writing text that generative
complex iterative process, the use of digital writing tools is
AI excels, and because writing is central to many forms of
complex. Our study of Wordtune also found it being commonly
learning, including assessment, this paper focuses on how gen-
being combined with other writing tools (Zhao et al., 2023).
erative AI is changing how students write. With the increasing
While many such tools have multiple functionalities, they tend to
use of AI in many domains of activity there is a growing interest
be used for specific tasks for which they are best known. For
in defining AI literacy (Long and Magerko, 2020). We build on
example, many tools will offer some support for translation, but
this work to propose a model of generative AI literacy as a fra-
users tend to have a preferred tool for this function.
mework to assess student use of ChatGPT in their writing.
Thus, it is important to recognise that many learners were
In this context, the present study had the aim of uncovering
already using multiple tools at different stages of the writing
how postgraduate students’ practices of writing were impacted by
process prior to the dramatic debut of ChatGPT, so the use of
ChatGPT, with the specific research questions for the study being:
generative AI appears in a pre-existing landscape of digital
(1) How were postgraduate students using ChatGPT and other writing. Yet the panoply of tools to support the writing process
digital writing tools for writing tasks in the summer has been little analysed, particularly from the perspective of how
of 2023? they are used in practice and in conjunction with each other,
(2) What do students consider the benefits and problems of during writing as a complex, iterative process.
ChatGPT’s use?
(3) What are the strengths and weaknesses in student Generative AI. ChatGPT’s launch has been a dramatic, poten-
generative AI literacy? tially paradigm-shifting intervention, influencing how writing as a
central aspect of learning is performed, but also the general
The digitisation of writing. The impacts of generative AI tools perception of AI in Education. It has seen an extraordinary
such as ChatGPT in education, need to be understood in the explosion of use, with a claimed 100 million users within two
context of long-term digitisation of writing. The digitisation of months of its launch (Trust et al., 2023). At the same time, it has
writing is a major shift in writing as a fundamental process of been deeply controversial, particularly within education, and has
expression and learning (Strobl et al. 2019). It carries particular been linked to many of the wider debates on the ethics of AI
significance because of the cultural value and status placed on around bias, privacy and impact on society. Whereas the wide-
good writing, particularly in some academic disciplines. spread use in the writing of tools such as Grammarly and Google
Through the introduction of word processors, then spelling, Translate seems to have been tacitly accepted with relatively little
grammar and style checking, then connectivity, and now gen- controversy, ChatGPT has drawn huge debate to the digitisation
erative AI tools, this digitisation process is having profound of writing (Adeshola and Adepoju, 2023; AlAfnan et al., 2023;
effects on writing, albeit they remain difficult to pinpoint Memarian and Doleck, 2023). This may be partly because tools
because they are primarily mental rather than directly visible such as Turnitin have at least partly given teachers the lead in the
changes (Kruse and Rapp, 2019). “arms race” with unfair means. From an educational point of
There has been an acceleration of this digitisation process in view, it is the impossibility of detecting generative AI’s use that
the last decade with the growing number of AI-powered writing makes it so controversial (Uzun, 2023).
assistants that are appearing. Godwin-Jones (2022) differentiates Yet it is hard to deny the power and user-friendliness of
four types of such tools: ChatGPT. While it remains essentially a form of narrow AI (as
opposed to a general AI that mimics the breadth of human
● Automatic writing evaluation (AWE) which provides intelligence), it does perform a wide range of tasks across the
feedback on completed work; writing process, potentially composing a complete essay, but also
● Automatic written corrective feedback (AWCF) which including the ability to: summarise readings or a topic, produce
offers synchronous feedback on spelling, grammar and or an outline for a text, draft text, rewrite text in different styles or
style as text is written; lengths, and check grammar and spelling (UNESCO, 2023). Thus,

2 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2024)11:482 | https://doi.org/10.1057/s41599-024-02904-x

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-024-02904-x ARTICLE

Table 1 Informational and ethical problems of ChatGPT.

Category Concerns/Impacts
Accuracy and Reliability • It “hallucinates” information, that is inaccurate, fails to acknowledge its sources and can even fabricate citations.
It only has data up to September 2021.
Transparency and Bias • It makes biased statements, e.g., studies have shown it has political bias but also reproduces sexist and racist stereotypes
(Deshpande et al., 2023; Motoki et al., 2023).
It is unexplainable because it is far from open about what data it is based on or how it works.
It is currently impossible to identify that the material was machine-generated.
Information Culture • It be used to create misinformation, fakes, or even harmful information.
It could also be used to accelerate the content creation explosion—leading to even more challenges of information overload
– but also potentially to increase the homogenisation of content.
It is “multilingual but monocultural” (Rettberg, 2022) because it is efficient in multiple languages but has American cultural
assumptions trained into it.
Better tools are available to people with money to subscribe, creating inequality in access to its benefits.
Impact on Learning • It could create lazy and superficial learning by making learning tasks like writing too easy.
Privacy is at risk if you share your data with it.
Ethical Concerns • It may violate intellectual property rights by using copyright material in its training without permission; training data sources
are not openly declared.
Very low-paid Kenyan workers were asked to view unpleasant material as part of the process of “detoxifying” data that was
being input to train ChatGPT (Perrigo, 2023).
GPT technologies have a huge environmental impact (Ludvigsen, 2022).

ChatGPT has the ability to write entire texts from a prompt or offered by Ridley and Pawlick-Potts (2021) when they suggest
support specific processes in writing. In addition, it can also write that:
computer code, solve math problems, etc. Yet ChatGPT, at least
in its early manifestation, poses many informational and ethical Algorithmic literacy is the skill, expertise, and awareness to,
problems (EPIC, 2023) (Table 1). understand and reason about algorithms and their
Some of these problems are being addressed in later versions of processes; recognise and interpret their use in systems
ChatGPT or in other text generation tools such as Bard or the (whether embedded or overt); create and apply algorithmic
new Bing. Moreover, they are not inherent to large language techniques and tools to problems in a variety of domains;
models but rather features of systems built by BigTech. Never- assess the influence and effect of algorithms in social,
theless, given the great AI capabilities of BigTech in terms of cultural, economic, and political contexts; position the
resources including data and so their power to define the individual as a co-constituent in algorithmic decision-
definition of AI, we anticipate that users need to be aware of such making.
potential issues. This is a concise expression of the key aspects, integrating
notions of algorithmic literacy, which focuses on the way that AI
AI literacy. In the last 5 years, there have been growing sugges- is often encountered indirectly through functions such as filtering
tions of the need to define AI literacy, because of the increasingly and personalisation on online platforms. Levels of student AI
pervasive presence of AI in everyday lives and work. A widely literacy have been much investigated, often using Long and
cited definition is offered by Long and Magerko (2020): Magerko’s (2020) framework (e.g. Kong et al. 2022). However,
most of this work was done before the advent of ChatGPT.
We define AI literacy as a set of competencies that enables The capability of ChatGPT and other generative AI to create a
individuals to critically evaluate AI technologies; commu- significant body of content from a short prompt has shifted
nicate and collaborate effectively with AI; and use AI as a concepts of what AI is. We suggest that this implies the need to
tool online, at home, and in the workplace. update our notion of AI literacy. In the light of the potential and
The authors break AI literacy down under five headings, with critiques of the technology (discussed above), we can suggest that
17 components under those headings: generative AI literacy might be defined under five headings:
● What AI is—this is knowledge such as how to recognise AI 1. Pragmatic understanding: The individual can use generative
when it is encountered and understanding distinctions AI effectively and interpret the information it produces
between general and narrow AI. critically
● What it can do—this consists of differentiating the tasks AI a. The individual can pick the right tool for the task, in the
is good at doing from those it is not good at, and also being context of the proliferation of writing tools (including
able to imagine future uses, reflecting the evolving alternative generative AI to ChatGPT)
nature of AI. b. The individual learns to use the chosen tool effectively for
● How AI works—includes ideas such as representation and a specific task
has an emphasis on data literacy, emphasising learning
from data and the need for critical interpretation of data. i. Deciding where in the writing process to use it, e.g.
● How it should be used—under which ethics is placed. for Search, brainstorming, structuring text etc
● How people perceive it. ii. Uses the tool effectively through prompt engineer-
ing), such as by
This is useful in contrasting to more technically oriented
definitions such as that of Pinski and Benlian (2023) which lack 1. Being CLEAR (concise, logical, explicit, adaptive
the ethical and critical dimension. Another useful definition is and reflective) (Lo, 2023)

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2024)11:482 | https://doi.org/10.1057/s41599-024-02904-x 3

ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-024-02904-x

2. Providing context for prompts posed similar range of concerns were also expressed such as about the
3. Defining what sort of answer is required reliability of information from generative AI, privacy, equity of
4. Rephrasing questions access and fears of becoming over-reliant on it. Students wanted
5. Asking for sources used educational institutions to have clearer policies and offer training
6. Iterating and synthesising results in the use of generative AI. Interestingly, they also wanted student
iii. Updating their knowledge as tools develop rapidly involvement in generating policy on AI.
If students are generally positive about ChatGPT’s use it is staff
c. The individual interprets generative AI outputs critically, who express more concerns. Cardon et al. (2023) conducted a
given an understanding of how they work and their limits survey of business communication instructors. The main concern
i. Information accuracy, currency, citeability of this group of educators was that students would use it to cheat.
ii. bias But they saw a range of negative impacts on learning, such as
● Less critical thinking/ creativity—itself seen as part of a
2. Safety understanding: The individual can use generative AI wider malaise, and the crisis of creativity
safely ● Less writing skills
a. Is aware of privacy risks ● Less authenticity
● Less agency because of dependence on such tools
3. Reflective understanding: The individual can assess and take ● Less commitment to authenticity in communication, such
action to manage the impacts of AI on their experience in the as valuing authorial voice and sincerity in communications
educational context
They also acknowledged that it can be helpful e.g. in the early
a. Impacts on own skills and learning stages of writing, and certainly enhanced the efficiency of writing.
b. Impacts on social connection, including the social However they did believe such tools would be used in the
aspects of learning workplace, so it was unavoidable that it had to be taught.
4. Socio-ethical understanding: The individual understands
the societal impacts of AI, including Methodology
a. IPR issues relating to how models are trained To answer the research questions, we employed a qualitative
b. Impact on information culture, misinformation and methodology within the interpretivist paradigm. We used a
disinformation combination of semi-structured interviews and observational
c. Social impacts such as through exploitative process of techniques to gain an understanding of how students selectively
creation, and the impacts on jobs/ job enrichment employed digital tools in their writing processes and to under-
d. Equity of access stand their experiences and concerns regarding the use of gen-
e. Environmental impacts erative AI. We recruited participants through an email invitation
f. Implications of the undue power of BigTech circulated to students asking for participants who were using
“digital tools” for writing. We also asked interviewees to suggest
5. Contextual understanding: The individual understands how other suitable participants. Our study included 23 participants of
to use generative AI appropriately in a particular context diverse nationalities, including students from the UK, China,
and make their own use explicit, as appropriate India, Thailand, Japan, Greece, Malaysia, the USA, and other
a. What is appropriate to context regions (Table 2). These students were pursuing a range of aca-
b. How to make use transparent and cite appropriately demic degrees, including postgraduate taught and postgraduate
research programmes. All the participants were in the process of
undertaking academic tasks, such as writing dissertations or
Generative AI in education. It has been education in particular theses. The interviews were conducted in the summer of 2023,
that has been disrupted by the potential and risks of generative AI this was before the university had issued its policy on AI use.
in 2023. Yet while much has been written about this in editorials At the beginning of the interview, participants were asked to
and opinion pieces (Kasneci et al., 2023; Trust et al., 2023) usually demonstrate their writing process for an academic essay/dis-
to inform educators about how to use it, we are only at the sertation and explain how they use digital tools to support their
beginning of learning its impact on student behaviour through writing. The second part of the interview participants were asked
empirical research. This is important because it seems likely that a series of questions, including about the tools they used during
students have taken up its use far more quickly than teachers. the writing process, how they had used ChatGPT, and what their
We do have a few early studies of use by students. A number of concerns were about it, such as data privacy, inclusivity, acces-
surveys by Best Colleges indicate that though US students had sibility, bias, ethics, and the potential impact of generative AI on
concerns about whether it was fair to use ChatGPT for education. Thematic Analysis served as our chosen method for
assessments, they were using it and saw it as soon to become analysing the qualitative data, enabling us to gain a nuanced
the norm (Welding, 2023). Chan and Hu (2023) found Hong understanding of students’ perceptions of digital writing and
Kong students positive about generative AI and willing to use it. ChatGPT in particular (Braun and Clarke, 2006). The research
This was partly because of its direct uses for brainstorming, received ethical approval from the University of Sheffield.
individualised assistance with questions, and help with literature Voluntary, informed consent was gained from participants. All
reviewing (such as summarisation). But their willingness to use it the data were anonymised for the purposes of analysis and
was partly because they saw it as representing long-term trends in reporting.
technologies. They had concerns about its accuracy, the
transparency of its working, the privacy of their data, the risk Findings
of becoming over-reliant, the impacts on employment and Many tools used in the complex task of writing. One theme that
conflicts with human values. This accords well with a study by emerged strongly from the data was that students were routinely
Attewell (2023) based on focus groups with UK students. This using a wide range of digital tools (many with an AI component)
again found generative AI being used in a wide range of ways. A throughout the academic writing process. The most commonly

4 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2024)11:482 | https://doi.org/10.1057/s41599-024-02904-x

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-024-02904-x ARTICLE

Table 2 Participant information.

Participant number Degree Nationality Self-perceived English proficiency

1 Postgraduate Taught China Professional working proficiency
2 Postgraduate Taught China Limited working proficiency
3 Postgraduate Taught China Professional working proficiency
4 Postgraduate Taught Saudi Arabia Professional working proficiency
5 Postgraduate Taught China Limited working proficiency
6 Postgraduate Taught China Professional working proficiency
7 Postgraduate Taught USA Native
8 Postgraduate Taught USA Native
9 Postgraduate Taught Thailand Professional working proficiency
10 Postgraduate Taught Japan Limited working proficiency
11 Postgraduate Taught China Professional working proficiency
12 Postgraduate Taught China Professional working proficiency
13 Postgraduate Research UK Native
14 Postgraduate Research India Native (Bilingual)
15 Postgraduate Taught India Native (Bilingual)
16 Postgraduate Taught India Native (Bilingual)
17 Postgraduate Taught UK Native
18 Postgraduate Taught Greece Professional working proficiency
19 Postgraduate Taught India Native (Bilingual)
20 Postgraduate Taught India Native (Bilingual)
21 Postgraduate Taught India Native (Bilingual)
22 Postgraduate Research China Professional working proficiency
23 Postgraduate Taught Malaysia Professional working proficiency

mentioned tools were grammar checkers (especially Grammarly), This is the topic and I first use ChatGPT to give me some
paraphrasing tools (Quillbot and Wordtune) and translation tools idea about this topic. And actually, when I saw this topic, I
(e.g., DeepL and Google Translate). Somewhat less commonly didn’t really understand it. I use ChatGPT to give me some
other types of tools were in use such as for managing references explanations. [9, Thailand]
(Endnote, MyBib) and plagiarism detection (Q-text).
Students who were native English speakers used more basic So I asked it how to organise an essay about this topic. And
tools such as Word’s grammar checking or very specialist tools then it gave me this structure. So I use this as a reference.
such as to manage references. In contrast, non-English speakers [2, China]
were using a wider range of tools, with considerable experience of Some also used it to check that the final text met the brief:
having done so built up over time.
Such services seem to be used in quite individualistic ways and Then in writing or after writing, I would ask if this
critically the impression was that their use had been learned from paragraph was in line with classwork requirements. I would
classmates, social media (such as Youtube or Little Red Book), send it all the requirements, and then ask it if I could write
and trial and error—rather than the institution and educators, like this? [1, China]
although one individual was using Grammarly on supervisory Some used it for searching for literature.
advice.
Sometimes students paid a subscription for such tools; I think the most difficult part of writing an essay is the
sometimes not. ChatGPT was the most frequently paid-for tool idea you need to make your own opinions and you need
among all those mentioned. to structure your essay but you need to find examples
and literature to support your argument. If you just
google it or search […] sometimes you can’t find much
ChatGPT: Used in many different ways. Interviewees talked relevant information or only a few. […] You can just
about other tools as much as about ChatGPT and for these find less relevant information and articles but with
interviewees, it was early days with ChatGPT. Many had used it ChatGPT, it will collect the most relevant information
to only a limited extent. Nevertheless, there was evidence of for you. [1, China].
ChatGPT being employed throughout the writing process. There For other tasks, such as summarising readings, rephrasing ideas
were frequent mentions of uses to: into more academic language and grammar checking participants
● understand difficult concepts including understanding used other tools as much as ChatGPT. Grammarly, Quillbot,
assignment briefs Wordtune and translation tools had heavy use in rewriting.
● summarise readings during the research process Students tended to use ChatGPT for just one or a few of these
● suggest structures for writing tasks, not all of them, linked to which aspect of writing they
● get words down on the page and break through a found hardest.
mental block Students’ individual explorations of the tool had produced
● rephrase text and check grammar quite distinct patterns of use, often linked to the areas of weakness
they perceived in their own writing. For example, participant 23
Central uses that ChatGPT was uniquely good at was gaining emphasised using it to generate analogies to help them under-
an understanding of an assignment brief and then structuring stand complex ideas in the context of moving to study computer
ideas: science from another discipline:

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2024)11:482 | https://doi.org/10.1057/s41599-024-02904-x 5

ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-024-02904-x

If I don’t understand, repeat, repeat, repeat, repeat, repeat Thinking about it, basically, there is no big change, because
until I understand and then proceed. And that is how my there are still steps in writing, mainly those steps, but the
process is. It’s a lot of repetition, a lot of reclarifying myself efficiency has been improved. [1, China]
and always detecting all this reading and fact checking. [23,
Malaysia] I think he saves me a lot of time when I’m reading like
Participant 17 described using in complex way to support reference that’s really long. Helps me summarise it. I think
reading: it’s really efficient. [11, China]
This discourse implied that it did not really change what they
I would copy literally like a whole article into it and then did or learned, just speeded the process up and saved them time.
say, summarise this in bullet points. I’ll then write it on a This may be a rationalisation and disguise the loss of learning that
piece of paper because that just helps me to like process it in the “efficiencies” might cause. A lot of the discourse around
my head as I write the words. Um, and then after that, and needing to use tools revolved around “laziness” but also a sense of
then I’ll read the paper like normal [17, UK] a lack of confidence and anxiety. Using ChatGPT was often
Participant 15 described using it directly in the writing process: justified for reducing stress.
Sometimes if I’ve written something, what I do is I ask for a I’ll say I feel less stressful because we get to work more
feedback and I ask or like, you know, if I want a efficiently by using the tools, by using AI tools compared to
paraphrasing also. So what I do is I prompt it by asking the traditional way. [11, China]
that the check for flow and paraphrase and then it will kind There was a sense of ChatGPT being the future and there is an
of give it suggestions and sometimes it adds its own things. inevitability to it overtaking “traditional” learning practices.
And then if I don’t want something, I can just not have it.
But most of the times what happens is you get a pretty good Ten years ago, study style, study style, and the current one
idea of like, you know, yeah, this is something good and is very different. [23, Malaysia]
then you can build up from there. So it gives you that initial Yet, while it was being used rather extensively our participants
kind of a boost and then it becomes easier to build your were wary of ChatGPT for a number of reasons. Three were
arguments or build your paragraphs. (15, India) mentioned by most interviewees. The first was the unreliability of
Participant 20 stands out as potentially using it in a way to the information it output and the need to “fact check” it. There was
simply write ideas for him. He talked about “delegating” tasks to less mention of specific issues such as failing to acknowledge
it, constructing himself as in control of the process: sources or making up references. A second major concern was that
its use would be detected by plagiarism detection tools (a concern
So when I break it down, so usually when I have an they held regardless of the fact that it is generally not detectable).
academic paper, say it’s 10 pages, I have one thought for More positively participant 19 wanted to preserve her “voice”
each page, like one heading for each page in my head. And
once I have that set, then I just start off with ChatGPT I feel like it doesn’t sound like me and I do want to write in
straight up. [20, India] my own voice or in my own words as much as I can.
Such complex uses, often involved quite interactive exchanges. [19, India]
Participant 23 used the term “bouncing” ideas off it. Similarly, participant 23 thought that ChatGPT created a text
that often read as “auto-generated” [23] so would not use it in
I always have sort of like a conversation with it. [23, many contexts, e.g. writing job applications because it would not
Malaysia] effectively differentiate them from other candidates. Implicitly,
Indeed, one interesting dimension of these interactions was the they did not see this as a problem in the context of academic
sense that the experience was something quite close to discussing writing, suggesting that they did not see individuality as valued in
with another human. that context.
There were certainly some students who did see using
Just directly ask a question. Yes. As if I’m asking a human ChatGPT as impacting their learning:
being. [22, China]
But it sometimes it stopped me from thinking. [12, China]
But always, always, always remember it is still a tool. It is
not a living being. That is that’s why I keep reminding It will reduce the ability to think independently, that is,
myself it’s not a living being. Always cross check your back. reduce the motivation to think independently. Some things
[23, Malaysia] need to be understood by yourself. [5, China]
The quote implies a genuine struggle not to see ChatGPT as Thus a third major concern was a fear that the use of tools in
human. Equally, the value of ChatGPT was often linked to not general produced dependence. There were also a few comments
needing to ask tutors or peers for help. Of course, it also gives that showed awareness of the risk to privacy of putting personal
immediate responses unlike them, but it raises the question of information into ChatGPT.
how ChatGPT is impacting the social dimension of learning. Nearly all the concerns reflected on personal impact.
ChatGPT was also being commonly used for non-academic Disappointingly, the societal impacts of ChatGPT were not often
writing, such as for job applications. Several interviewees acknowledged. One interviewee expressed concerns about the
mentioned using it for computer coding. There were also very impact on low-paid workers. But even when prompted the
specific uses, such as for converting text to Latex. participants showed little awareness or concern about bias, or
ethical issues, e.g., sustainability or the exploitative labour
relations used to create ChatGPT. Nevertheless, it seemed that
ChatGPT benefits and worries. The way that they described it, the controversy around ChatGPT had filtered through to
ChatGPT and the other digital tools they used, gave participants a problematising its use, in ways that were expressed far less in
sense of being more efficient and productive. relation to other tools such as Grammarly.

6 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2024)11:482 | https://doi.org/10.1057/s41599-024-02904-x

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-024-02904-x ARTICLE

Table 3 Participants’ Generative AI literacy.

Generative AI literacy criteria Summary assessment of interviewees’ generative AI literacy

1a The individual can pick the right tool for the task, in the context of the Students were aware of many AI writing assistants and tended to use
proliferation of writing tools (including alternative generative AI to them for specific tasks.
ChatGPT)
1b The individual learns to use the chosen tool effectively for a specific task A wide range of uses were being made to address individual needs.
The observations did not show sophisticated prompt engineering.
1c The individual can interpret generative AI outputs critically, given an There was awareness of problems of information accuracy, but less of
understanding of how they work and their limits bias.
2. Safety understanding: The individual can use generative AI safely There was good awareness of the privacy issue
3. Reflective understanding: The individual can assess and take action to Interviewees were concerned about issues such as the impact on their
manage the impacts of AI on their experience in the educational context learning and to a lesser extent the social dimension of learning
4.Socio-ethical understanding: The individual understands the societal Interviewees showed limited awareness and tended to make light of these
impacts of AI impacts
5. Contextual understanding: The individual understands how to use Concerns about appropriate uses were at the forefront of participants’
generative AI appropriately in a particular context and makes their own minds. Their call for institutional clarification reflected a desire to use
use explicit, as appropriate tools in contextually appropriate ways

Discussion early days with their use of it and almost none of them had
The findings confirm our earlier work from before the release of received support from their teachers or from their institution in
ChatGPT, that a wide range of tools is used during the writing understanding how to use it. The weakest area of development
process (Zhao et al., 2023). Perhaps ChatGPT will displace was probably appreciation of the societal impacts of generative
other tools, but at the time the study was conducted it was AI. As instructors catch up with students in understanding AI
reaching a community of users well-versed in using digital hopefully they can help students build up a more systematic
tools. It was finding a unique place in helping understand understanding of pragmatic use of AI, a more reflective approach
assignment questions, aligning answers to the question and and a much more critical awareness of the social implications
structuring ideas, but was also used to perform tasks such as of AI.
summarisation and proofreading that others performed with
pre-existing tools. Our findings suggested that students with
higher English proficiency levels tended to use AI tools less Conclusion
frequently. While ChatGPT appears as a threat to longstanding practices in
The study confirms previous research, that students are keen to education, especially to some genres of academic writing such as
use generative AI and see it as part of a general trend in tech- essays, it can also be seen as productively bringing to the fore the
nology development (Chan and Hu, 2023). Our respondents controversial nature of AI writing technologies which were
tended to use discourses around time-saving, efficiency, and stress already creeping unacknowledged into common use. This context
avoidance as justification for using the tools. produces an opportunity for educators to actively engage with
ChatGPT was used rather intensively and iteratively but in students in a discussion about how writing can be best supported.
highly individualistic ways. This differential use may reflect that Our analysis points to areas of weakness in generative AI literacy
there was little input at this time from the institution to help that need to be strengthened through this process, such as the
students how to use and where to draw the line in terms of understanding of the need to select between apps, to have more
appropriate use. Students wanted guidance from the institution sophisticated prompt engineering skills, to think more about bias
on ChatGPT, echoing Attewell’s (2023) findings. in results, to be more reflective about its use’s impact on learning
ChatGPT has brought advanced functions to digital writing and have a much stronger appreciation of the societal impacts of
but also intensified a sense of controversy in this area. For stu- generative AI.
dents, the worries focussed particularly on the unreliability of The paper is one of the first to explore student use of generative
information it produced, fear of being accused of plagiarism if AI in practice and discover in-depth their perception of its ben-
they used it and a concern about growing dependence on tech- efits and worries about its drawbacks. It has also developed a
nology. While often its use was claimed to be justified for its time framework of generative AI literacy as a way of assessing their
and stress reduction this may have underestimated the overall use. This can be used to plan both institutional policy and
impact on learning. Saving time on learning tasks may unin- instructor support by identifying gaps in AI literacy that need to
tentionally remove significant opportunities to learn. Many of be filled. For example, educational developers could use the fra-
the deeper ethical and societal issues such as around the mework to facilitate discussions with teaching staff, aiding in the
exploitative way ChatGPT was developed were not fully under- development of their AI literacy and enhancing their ability to
stood. Yet it was clear that how writing was done digitally had teach AI literacy to students. In addition, universities could apply
become controversial. This could be seen as a benefit of the dimensions of the AI framework to formulate policies and
ChatGPT, in that a gradual infiltration of digital tools into provide concrete examples that guide learning and teaching
writing was made more visible institutionally and the con- practices. The framework also has the potential in evaluating
troversial dimensions of technology use in education brought to student AI literacy.
the fore. The paper has a number of limitations, pointing to where
Using the model developed above we can point to strengths future research can build on its findings. Most participants in this
and weaknesses in students’ emergent generative AI literacy study were using the free version of ChatGPT (3.5) and only a few
(Table 3). used the paid version (4). We did not examine in detail the
Overall our student interviewees showed significant generative impact of using different versions. While ChatGPT was the main
AI literacy in most areas, particularly when considering it was generative AI tool in use at the time of the study, there were

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2024)11:482 | https://doi.org/10.1057/s41599-024-02904-x 7

ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-024-02904-x

others rapidly emerging in popularity. Future research would Kruse O, Rapp C (2019) Seamless Writing: How the Digitisation of Writing
need to examine how choices of different apps were made and the Transforms Thinking, Communication, and Student Learning. In: Looi CK,
impact of these choices on writing. Although the participants of Wong LH, Glahn C, Cai S (eds) Seamless Learning. Lecture Notes in Edu-
cational Technology. Springer, Singapore, pp 191–208. https://doi.org/10.
this study come from a variety of countries, it is focused on one 1007/978-981-13-3071-1_10
institution at a particular time. Given the speed of change in Lo CK (2023) What is the impact of ChatGPT on education? A rapid review of the
technology and educational policy and practice, it is likely that literature. Educ Sci 13(4):410
future research will identify rapid shifts in behaviour. But we Lo LS (2023) The CLEAR path: a framework for enhancing information literacy
emphasise the need to examine student writing practices in the through prompt engineering. J Acad Librariansh 49(4):102720
Long D, Magerko B (2020) What is AI literacy? Competencies and design con-
context of significant pre-existing use of digital writing tools. Our siderations. In: Bernhaupt R, Mueller F, Verweij D, Andres J (eds) Pro-
focus was on postgraduate students. We think less experienced ceedings of the 2020 CHI conference on human factors in computing
students, such as undergraduates may be quicker and less dis- systems, Association for Computing Machinery, pp. 1–16
criminating in adopting the technology. As generative AI evolves Ludvigsen K (2022) The carbon footprint of Chat GPT. https://
there will be a need to update our definition of generative AI towardsdatascience.com/the-carbon-footprint-of-chatgpt-66932314627d.
Accessed 21 Dec 2022
literacy and also to integrate it with notions such as algorithmic Malik AR, Pratiwi Y, Andajani K, Numertayasa IW, Suharti S, Darwis A (2023)
literacy, which point to the way that AI operates in rather hidden Exploring artificial intelligence in academic essay: higher education student’s
ways within the infrastructure. This research employed interviews perspective. Int J Educ Res Open 5:100296
and observations as its main data collection methods. These offer Memarian B, Doleck T (2023) ChatGPT in education: Methods, potentials and
depth of insight but have less power of generalisability. Future limitations. Computers in Human Behavior. Artificial Humans: 100022
Motoki F, Pinho Neto V, Rodrigues V (2023) More human than human: measuring
studies could usefully seek to validate our findings through ChatGPT political bias. Public Choice. https://doi.org/10.1007/s11127-023-
quantitative or mixed-methods approaches, such as surveys or 01097-2
experimental studies. Furthermore, future research could expand Pinski M, Benlian A (2023) AI literacy-towards measuring human competency in
the scope of this study from AI literacy to the broader concept of artificial intelligence. In: Proceedings of the 56th Hawaii International Con-
writing digitisation, exploring the issue from other perspectives ference on System Sciences
Perrigo B (2023) Exclusive: OpenAI used Kenyan workers on less than $2 per hour
such as psychology and second language acquisition. to make ChatGPT less toxic. Time Magazine. https://time.com/6247678/
openai-chatgpt-kenya-workers/ Accessed 18 Jan 2023
Rettberg J (2022) ChatGPT is multilingual but monocultural, and it’s learning your
Data availability values. https://jilltxt.net/right-nowchatgpt-is-multilingual-but-monocultural-
but-its-learning-your-values/ Accessed Jan 18 2023
The datasets analysed during the current study are not publicly Ridley M, Pawlick-Potts D (2021) Algorithmic literacy and the role for libraries. Inf
available, but are available from the corresponding author on Technol Libr 40(2) https://doi.org/10.6017/ital.v40i2.12963
reasonable request. Selwyn N (2019) Should robots replace teachers? AI and the future of education.
John Wiley & Sons, Cambridge
Strobl C, Ailhaud E, Benetos K, Devitt A, Kruse O, Proske A, Rapp C (2019)
Received: 30 December 2023; Accepted: 28 February 2024; Digital support for academic writing: a review of technologies and peda-
gogies. Comput Educ 131:33–48. https://doi.org/10.1016/j.compedu.2018.
12.005
Trust T, Whalen J, Mouza C (2023) ChatGPT: challenges, opportunities, and
implications for teacher education. Contemp Issues Technol Teach Educ
References 23(1):1–23
AlAfnan MA, Dishari S, Jovic M, Lomidze K (2023) Chatgpt as an educational tool: UNESCO (2023) ChatGPT and artificial intelligence in Higher Education https://
opportunities, challenges, and recommendations for communication, busi- www.iesalc.unesco.org/wp-content/uploads/2023/04/ChatGPT-and-
ness writing, and composition courses. J Artif Intell Technol 3(2):60–68 Artificial-Intelligence-in-higher-education-Quick-Start-guide_EN_FINAL.
Adeshola I, Adepoju AP (2023) The opportunities and challenges of ChatGPT in pdf
education. Interact Learn Environ 1–14 https://doi.org/10.1080/10494820. Uzun L (2023) ChatGPT and academic integrity concerns: detecting artificial
2023.2253858 intelligence generated content. Lang Educ Technol 3(1):100060
Attewell S (2023) Student perceptions of generative AI. JISC National Centre for Welding L (2023) Half of college students say using AI is cheating. BestColleges.
AI. https://beta.jisc.ac.uk/reports/student-perceptions-of-generative-ai https://www.bestcolleges.com/research/college-students-ai-tools-survey/.
Braun V, Clarke V (2006) Using thematic analysis in psychology. Qual Res Psychol Accessed 27 Dec 2023
3(2):77 Zawacki-Richter O, Marín V, Bond M, Gouverneur F (2019) Systematic review of
Cardon P, Fleischmann C, Aritz J, Logemann M, Heidewald J (2023) The chal- research on artificial intelligence applications in higher education—where are
lenges and opportunities of AI-assisted writing: developing AI Literacy for the educators? Int J Educ Technol High Educ 16(1):39. https://doi.org/10.
the AI Age. Bus Prof Commun Q 23294906231176517 1186/s41239-019-0171-0
Chan CKY, Hu W (2023) Students’ voices on generative AI: perceptions, benefits, Zhao X, Sbaffi L, Cox AM (2023) The Digitisation of Writing in Higher Education:
and challenges in higher education. Int J Educ Technol High Educ 20 (43). exploring the Use of Wordtune as an AI Writing Assistant. OSF preprint.
https://doi.org/10.1186/s41239-023-00411-8 https://doi.org/10.31219/osf.io/uzwy7
Deshpande A, Murahari V, Rajpurohit T, Kalyan A, Narasimhan K (2023) Toxicity Zhao X, Xu J, Cox, AM (2024) Incorporating artificial intelligence into student
in ChatGPT: analyzing persona-assigned language models. arXiv preprint. academic writing in higher education: the use of wordtune by Chinese
arXiv:2304.05335 international students. Paper presented at the Hawaii Systems Sciences
Electronic Privacy Information Center (2023) Generating Harms: Generative AI’s conference, Waikiki, Hawaii, 2–6 January 2024
Impact & Paths Forward. https://epic.org/wp-content/uploads/2023/05/
EPIC-Generative-AI-White-Paper-May2023.pdf
Godwin-Jones R (2022) Partnering with AI: intelligent writing assistance and
instructed language learning Language. Learn Technol 26:5–24. https://doi. Author contributions
org/10.10125/73474 The authors contributed equally to this work.
Guan C, Mou J, Jiang Z (2020) Artificial intelligence innovation in education: a
twenty-year data-driven historical analysis. Int J Innov Stud 4(4):134–147
Kasneci E, Seßler K, Küchemann S, Bannert M, Dementieva D, Fischer F, Kasneci Competing interests
G (2023) ChatGPT for good? On opportunities and challenges of large lan- The authors declare no competing interests.
guage models for education. Learn Individ Diff 103:102274
Kong SC, Cheung WMY, Zhang G (2022) Evaluating artificial intelligence literacy
courses for fostering conceptual learning, literacy and empowerment in Ethical approval
university students: refocusing to conceptual building. Comput Hum Behav Approval was obtained from the ethics committee of the University of Sheffield [054920].
Rep. 7:100223 The procedures used in this study adhere to the tenets of the Declaration of Helsinki.

8 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2024)11:482 | https://doi.org/10.1057/s41599-024-02904-x

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-024-02904-x ARTICLE

Informed consent Open Access This article is licensed under a Creative Commons Attri-
Participants received an information sheet prior to the interview. Informed consent was bution 4.0 International License, which permits use, sharing, adaptation,
collected from all participants before the interviews were conducted. Participants were distribution and reproduction in any medium or format, as long as you give appropriate
informed about anonymity and the right to withdraw. credit to the original author(s) and the source, provide a link to the Creative Commons
licence, and indicate if changes were made. The images or other third party material in
this article are included in the article’s Creative Commons licence, unless indicated
Additional information otherwise in a credit line to the material. If material is not included in the article’s Creative
Correspondence and requests for materials should be addressed to Andrew Cox. Commons licence and your intended use is not permitted by statutory regulation or
exceeds the permitted use, you will need to obtain permission directly from the copyright
Reprints and permission information is available at http://www.nature.com/reprints holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations. © The Author(s) 2024

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2024)11:482 | https://doi.org/10.1057/s41599-024-02904-x 9