1 Introduction
Teacher professional development literature has long advocated for informal teacher communities of practice that lead to unstructured social learning [
16,
71]. Prior work [
60] suggests that collaboration and relationships with other teaching community members help teachers shape their work. However, teacher communities in the Global South have infrastructural limitations, such as financial means, roads, and vehicles, preventing teachers from regularly meeting their teaching community in person, thus limiting community interactions. Today, the expansion of technology has allowed opportunities for creating virtual communities of practice for teachers using social media [
90,
102]. Notably, teacher support groups on WhatsApp have been instrumental in fostering virtual teacher communities in both Indian [
105] and African settings [
67,
79]. Unfortunately, connectivity issues prevent streamlined access and participation of rural African teachers [
67]. Therefore, it is still unclear if virtual communities of practice impact teacher professional development in rural African contexts.
Effective teacher professional development promises better-trained and motivated teachers, but it faces challenges when training educators in the Global South [
99]. The implementation of teacher professional development programs, especially in rural and isolated areas, encounters difficulties due to both infrastructural [
19,
48] and socio-cultural complexities [
61,
93]. For example, in African contexts a teacher may receive a one-week training and are subsequently expected to implement the program with limited mentoring [
19,
25]. While mentoring is provided through regular visits of pedagogical advisors acting as administrators [
48], infrastructural challenges and a shortage of mentors reduce the frequency of these sessions [
19,
68]. Recognizing technology’s potential in aiding teachers in the Global South [
47], both government and
non-governmental organizations (
NGOs) are investing in teacher training and technology integration [
67,
105]. While there’s initial evidence of technology impacting teachers’ professional development positively [
25,
26], the long-term impact on teacher motivation and pedagogical knowledge necessitates thorough assessment of such technological interventions.
Conversational agents, also known as chatbots, have seen initial success with teachers in rural African contexts [
26]. An exploratory deployment by the authors of a conversational agent on WhatsApp in rural Côte d’Ivoire [
25,
26] found initial evidence that such technology can provide personalized and context-specific support to teachers. Recent research has allowed chatbots to support virtual communities of practice using innovative AI techniques in Western settings [
45,
88]. These results are promising and motivate our choice of conversational agents. However, AI research is often disconnected from low infrastructure contexts due to socio-cultural nuances [
87]. Prior work has attempted to bridge this disconnect [
26,
49,
64,
116] through qualitative design studies with stakeholders. Despite these attempts, the longitudinal impacts of conversational agents and their ability to foster virtual communities of practice in rural African contexts are still unexplored.
To address the literature gaps in the impacts of virtual communities of practice and conversational agents for teachers in low infrastructure settings, we conducted a longitudinal study to evaluate the impact of community support designs in a chatbot on Facebook Messenger. Our study involved a year-long quasi-experiment with 313 teachers in rural Côte d’Ivoire to investigate the impact of two variations of a chatbot. One variation supported a virtual community of practice (vCOP); the other was a traditional agent that acted as a control. We evaluated the impact of our intervention on teacher motivation, teachers’ pedagogical knowledge, and technology adoption. Our article addresses the following research questions.
RQ1: How do teachers use a chatbot that supports a virtual community of practice (vCOP) in low infrastructure settings of Côte d’Ivoire?
RQ2:
How does a chatbot that supports a vCOP impact teachers in low infrastructure settings of Côte d’Ivoire?Our contributions to the COMPASS and HCI community include:
(1)
We ran a rigorous longitudinal quasi-experiment in rural African contexts with a vCOP and control for over a year. We reveal how teachers use such a system longitudinally to ask questions, share stories, and adapt their teaching practices.
(2)
We demonstrate moderate improvements in teachers’ motivation and knowledge components. We also show that active chatbot users improved a specific pedagogical knowledge component.
Our study extends prior work by the authors in Côte d’Ivoire [
25,
26,
27], which provided the initial need finding [
27] and exploratory deployment [
25,
26] to motivate this work. Through our article, we make a case for virtual communities of practice and conversational agents to bolster support for teachers in low-infrastructure settings.
6 Discussion
We will now explore the interpretations of our results in the context of our three hypotheses related to technology adoption, teacher motivation, and teacher knowledge. We examine our interpretations within the framework of previous research in ICTD to explain the phenomenon in our data.
We hypothesized in
H1, that chatbot adoption would increase in the community, but the results were not statistically significant. From qualitative data, we observed that teachers lost utility for some chatbot features over time, thus explaining the low difference in both conditions. Although teachers used the system, some teachers, especially in the control condition, did not see the utility in asking questions that could be obtained from the manual, indicating the redundancy of need-based features [
25] over time. For example, the information from the manual, tips, and questions were targeted solely at
user needs that, as Toyama mentioned, change over time [
101], thus reducing adoption. Unlike needs, leveraging
assets or users’ strengths paves the way for improved adoption. A prior study [
27] in Côte d’Ivoire describes children’s success and community as some of the core strengths or
assets [
77] in Côte d’Ivoire. Our data showed that teachers repeatedly appreciated the stories in the chatbot about children’s success and appreciated connecting with other teachers. Therefore, the synergy between vCOP designs and
assets may have improved technology adoption in the community condition.
However, external social media platforms allowed teachers in the control condition to form connections despite not having access to the features of the chatbot, leading to lowered adoption. Prior work [
67] that found that internal activity in their group intervention made it hard for some teachers to participate; similarly discovered that
external systems (WhatsApp groups) could also affect rural teachers’ participation in our Facebook
DIA. Although only one official WhatsApp group was set up by the NGO, we found that teachers organically set up groups independently. Some teachers mentioned they eventually moved to WhatsApp due to the myriad notifications and contacts they received from their colleagues. The self-motivation of the teacher community to set up and stay connected through WhatsApp groups further solidifies the importance of community for Ivorian teachers.
We found it surprising that teachers overcame financial barriers to access the community. Although Facebook was more accessible financially (due to the presence of Freebasics [
4], which provided free access to Facebook and Facebook Messenger for some networks), some teachers preferred Whats-App and were willing to pay for data to connect with their community on WhatsApp. Toyama describes self-motivation as amplification [
100], i.e., when the motivation is strong, users will find pathways to navigate technological challenges to succeed at their tasks. This finding suggests that users’ motivations can also help overcome financial struggles when the user is intrinsically motivated to form social connections.
We hypothesized in
H2 that motivation would increase in the community condition, and we observed an increase in some sub-scales. Referring back to Table
6, the community condition, enhanced by the chatbot, showed increased professional agency and perceived colleague support, aligning with our prior work on the positive impact of community support in Côte d’Ivoire [
25,
27]. For
agency combined, i.e., teachers’ professional agency in the community, there was a positive change over time for community users. We also saw a greater difference in active community users. The
professional agency in the community scale asked questions about community support (i.e.,
I ask my colleagues for support when facing exhausting work situations or
Our teacher community can care for our pupils together). The increase in these differences indicates the benefit of vCOP for the community condition by the chatbot. Prior work suggests that community support can be increased by a sense of encouragement [
15], provided by stories in our study. Additionally, the stories helped teachers reaffirm the benefits of the
NewMethod program for the students through the voices of other community members. Thus, the community-focused interactions may have helped reinforce existing strengths (assets [
77]) in the context [
27], leading to improved motivation. We observe greater differences in
perceived colleague support in active users with a marginally significant effect (7.4 %,p=0.079), hinting at the positive change over time in the community condition toward colleague support.
However, self-efficacy experienced no significant results, potentially influenced by conflicting demands from the
NewMethod program. Prior work has shown that teacher self-efficacy drops when there is more burnout in the context of novel teaching practices [
37]. From qualitative data, we learned that the
NewMethod program and traditional teaching were conflicting from a teacher’ perspective, leading them to have an additional workload with
NewMethod. For example, teachers mentioned that
NewMethod disrupted their classroom activities as it required time to prepare for
NewMethod. Thus, the conflict of
NewMethod between traditional teaching and additional workload may have increased burnout, reducing self-efficacy over time.
Teachers valued community-driven learning, akin to a prior peer support phenomenon [
54,
109] concept documented in HCI4D literature. In community support, teachers expressed that stories from their fellow teachers helped them learn the program’s importance. Prior work describes this concept as
social proof [
109], i.e., changing behavior by observing people in their circles. The benefit of community-driven behavior change is well-documented in HCI4D work [
54,
66,
73]. For example, early work in HCI4D demonstrated that farmers benefited from tips shared by peers through an interactive voice response system [
73], and recent work found that Youth with HIV found it beneficial to engage and learn from their peer WhatsApp groups [
54].
We hypothesized in
H3 that pedagogical knowledge will increase in the community condition, and we observed in Table 7 a positive change only on one knowledge component for active users. Further examination of log data does not give us conclusive evidence. The knowledge questionnaire tested teachers on the theoretical knowledge of
NewMethod (Knowledge question:
During the ‘jump on numbers’ activity, how likely is the student who jumps on numbers to find the correct number said by the other student?). The French and Math sections were theory-based, and the
NewMethod-
test and mentoring sections contained a mix of practical and theoretical knowledge. However, the questions log shows that teachers used the chatbot to ask practical questions about
NewMethod (chatbot question:
How to conduct the activity named jump on letters?). Teachers’ need for practical support can be explained by qualitative data where teachers said they referred to the chatbot specifically when they needed help with
NewMethod implementation. Therefore, the discrepancy between the theoretical questionnaire and the practical questions on the chatbot may explain the results. Furthermore, teachers mentioned the discrepancy in theory and practice, which could also mean that theoretical knowledge gained would not translate into successful implementation in rural Côte d’Ivoire.
Additionally, teachers mentioned using the chatbot offline through the offline PDF manuals. We also observed increased self-reports of usage in the control region, possibly hinting at more offline usage in the control region. Therefore, offline usage in control could have helped teachers gain theoretical knowledge from the manuals and helped them do well on the pedagogical knowledge questions. The increased knowledge from offline interactions in the control may have led to reduced results in difference-in-difference for the knowledge component.
6.1 Future Work
A Case for Virtual Communities of Practice and Chatbots in Low-Infrastructure Settings: During our study, vCOPs on social media were identified as an emerging research area with significant challenges in engaging teachers from rural areas and moderating content effectively [
67,
70,
78,
105,
106,
107]. Chatbots were seen as a promising solution to these issues [
41,
85,
99], though their impact in rural African contexts was uncertain. We addressed these gaps by developing a chatbot, inspired by our prior work in Côte d’Ivoire [
24,
25,
26] to support a virtual community of teachers. Although preliminary, our research contributes quantitative and qualitative metrics supporting the positive effects of virtual communities on teachers’ motivation and pedagogical knowledge. Therefore, future work can extend this work to create novel chat interfaces for communities. For example, an opportunity could be to work with
large language models (
LLMs), such as ChatGPT, to create novel interactions tailored to users’ contexts, such as generating personas of mentors [
72] using LLMs to help novices have early interactions with chatbots. Park et al. [
72] created personas of professors to help prospective grad school applicants reflect on their future choices. The limited training requirement (such as short surveys) to make these personas in LLMs allows for the extension of these systems to low-resource contexts.
A Case for Experimental Studies Motivated by Theory in ICTD: A challenge with ICTD research is the work is too contextual, thus failing to transfer the findings to other contexts [
33]. Although there have been experimental studies in ICTD [
66,
82,
96,
108,
115], the work focussed on practical and contextual impact, hence transferability in ICTD research [
33], i.e., translating the outcomes beyond the context of the experiment, is often challenging. Alternatively, there has been theoretical work [
34,
58,
101], often hard to connect with experimental work. Our contribution attempts to bridge this gap, and we make a case for running experimental studies toward bridging the gaps between theory and practice in ICTD [
22]. We have used our theoretical motivations in COPs [
78,
112] and practical understanding from prior work Côte d’Ivoire [
25,
26,
27] to design and conduct an experiment to understand the impact of vCOPs. We conducted a longitudinal study involving 313 teachers to assess the impact of the virtual community on teacher motivation and pedagogical knowledge. We categorized impact as technology use, pedagogical knowledge, and teacher motivation (as described by our Theory of Change, see Appendix
A.1. To evaluate our impact, we adapted and validated existing questionnaires on perceived social support, agency within the community, burnout, and self-efficacy for the Ivorian context. Future work can use our experimental approach to extend work on emerging theory in ICTD, such as aspirations [
101]. We also see the potential for chatbots and AI to play an active role in this direction due to their versatility [
41,
72,
85,
99] and potential for chat-based research in healthcare [
54] and education [
105].
Human-AI collaboration in Low-Infrastructure Settings One notable challenge in virtual communities is the necessity for moderation [
67], particularly as these systems involve teachers with low digital literacy. Our work shows a possibility to address this challenge through a humbot [
26,
41] architecture, a system where human operators and artificial intelligence collaborate to answer user queries. Our work shows that the humbot architecture is suited for constrained data availability, demonstrated by the chatbot’s ability to answer questions. Since this was an experimental setup, the NGO worker re-answered queries even if the chatbot answered them. The chatbot used a static database populated from prior work in the context. Hence, it could answer only a few questions beyond a greeting. Future work can focus on building a dynamic database for that chatbot [
45]. Huang et al. [
45] used responses from crowd workers to automate the chatbot responses over time; this concept could be adapted to build a dynamic database, i.e., a chatbot that learns from NGO workers every new answer, further reducing the workload of the NGO worker.
6.2 Limitations
6.2.1 Low Statistical Power of Analysis.
Although we observed significant results in the Motivation category (5 and 7 significant results out of 11 variables), we acknowledge that we had low statistical power for more rigorous tests, such as error correction. In our proposed future work, we will utilize the
False Discovery Rate (
FDR) for exploratory error correction, adjusting p-values to q-values via the Benjamini-Hochberg correction [
18]. FDR controlling procedures, known to be less stringent than the Bonferroni correction when handling multiple tests [
18], will be particularly relevant given the expected low statistical power in our data.
6.2.2 Challenges with Implementation.
Another limitation was the brevity of technology training sessions due to budget constraints and training resources. The Ministry of Education had a tight budget for training the teachers on NewMethod. The training occurred for a week, and teachers had to travel and stay in towns during the training. The intense training plan could accommodate short technology training for only an hour in 2–3 sessions. The fast technology training sessions diminished the technology training quality as teachers had lower digital literacy.
Another challenge was the Facebook regulations on interacting with users and timely answers. Facebook limited agent-initiated interactions to a 24-hour window after the user’s message to restrict spam messages through chatbots. This 24-hour limitation meant users could not be sent reminder messages or notifications beyond a day. Prior work [
57] has shown that notifications are integral to reminding users to continue using the system. Future work can consider requesting Facebook permissions that allow seven days of interaction for the human (NGO worker) interacting with the agent.