Trust Conditions and Regulatory Protection in AI Based Agricultural Information in Bangladesh

Trust Conditions and Regulatory Protection in AI
Based Agricultural Information in Bangladesh
Dr Kazi Abdul Mannan 
Center for Academic & Professional Career Development and Research (CAPCDR)
https://orcid.org/0000-0002-7123-132X
Md. Nasim 
Shanto-Mariam University of Creative Technology https://orcid.org/0009-0007-7312-2243
Research Article
Keywords: agricultural artificial intelligence, regulatory protection, trust, provider transparency,
accountability, Bangladesh
Posted Date: November 12th, 2026
DOI: https://doi.org/10.21203/rs.3.rs-11704076/v1
License:   This work is licensed under a Creative Commons Attribution 4.0 International License.  
Read Full License
Additional Declarations: The authors declare no competing interests.

1
Trust Conditions and Regulatory Protection in AI Bas ed
Agricultural Information in Bangladesh
Md. Nasim
Department of Law, Shanto-Mariam University of Creative Technology, Dhaka, Bangladesh
Email: nusratzahan2727@gmail.com
ORCID: https://orcid.org/0009-0007-7312-2243

Prof. Dr Kazi Abdul Mannan
Department of Business Administration, Shanto-Mariam University of Creative Technology,
Dhaka, Bangladesh
Email: drkaziabdulmannan@gmail.com
ORCID: https://orcid.org/0000-0002-7123-132X
Abstract
AI-supported agricultural information may provide useful guidance while raising concerns
about local suitability, accountability and oversight. This study examines stated trust conditions
and perceptions of regulatory protection in a Bangladesh survey. The 385 adult, AI-aware
records comprise law students, graduates, legal professionals and other respondents; 60 report
a location classified as upazila or rural. The file does not identify farmers or establish that
respondents use agricultural AI, so findings concern a mixed sample rather than representative
rural users. The analysis retains separate questionnaire items and compares paired endorsement
of statements on source/provider clarity, regulatory oversight and accountability. It also
explores correlations between AI awareness items and two trust-related statements. Of 385
respondents, 310 (80.5%) agreed that clear sources and a responsible provider would increase
trust, 255 (66.2%) agreed that regulatory oversight can increase confidence, and 317 (82.3%)
agreed that strong accountability increases willingness to use agricultural AI. Source/provider
clarity exceeded oversight endorsement by 14.3 percentage points; 210 respondents endorsed
both statements. These are differently worded hypothetical propositions, so their contrast is not
a ranking of interventions or evidence of causal effects. Associations between awareness and
trust conditions were small and do not establish that training changes reliance. The study
contributes an item-level distinction between provider transparency, institutional oversight,
local adaptation and accountability, and proposes a future design that evaluates specific
advisory systems with actual agricultural users.
Keywords: agricultural artificial intelligence; regulatory protection; trust; provider
transparency; accountability; Bangladesh

1. Introduction
   Agricultural information can influence crop selection, irrigation, pest management, marketing
   and other decisions with practical costs. Digital systems may make guidance easier to access
   and can support extension services, but their recommendations must be judged against local
   growing conditions, the quality and provenance of the information used and the availability of
   qualified human advice. A fluent answer is not evidence that a recommendation is accurate. A

2
recent review of AI and machine learning in agriculture identifies interconnected concerns
about data governance, reliability, digital access and explainability.1
Bangladesh is an important context for studying digital agricultural services because access,
infrastructure, language and trust may vary across communities and decision settings. Existing
digital-agriculture scholarship also highlights unequal control of data and the need to involve
end users in design. 2 The present records, however, do not identify respondents as farmers or
confirm that they have used agricultural AI. They provide evidence of the recorded mixed
sample’s stated perceptions, not national estimates for the rural agricultural population.
Trust in an AI system is often discussed as though it were a single attitude. A person may
instead distinguish knowing the responsible provider, having a route to report harm, believing
that public oversight is effective, and receiving locally appropriate advice. Each condition can
affect reliance differently. Regulatory protection in particular should not be inferred from
confidence in an individual provider. Conversely, endorsement of oversight does not establish
that an effective institution exists or that a particular service complies with law.
Earlier research on digital farming highlights that the service provider, the person receiving
advice and the institution responsible for a remedy can be different actors. An apparently
reliable recommendation can therefore raise questions at several stages: who supplied the
evidence, who checked its local suitability and who will respond if the advice is harmful.3 The
article focuses on differences among these stated conditions. Its distinctive analytical choice is
to compare how the same individuals respond to specific propositions on source and provider
clarity, oversight, accountability and local adaptation. These questions have different wording
and are not forced-choice alternatives. Their comparison can reveal different response
distributions, but cannot rank their causal effects on trust. The work uses Trust Theory and the
Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) as complementary
perspectives rather than asserting that the survey validates either model.4
The objective is to describe and compare reported trust conditions and governance preferences
for AI-supported agricultural information. Four questions guide the analysis: How often are
clarity, oversight, accountability and reporting endorsed? How do responses to source/provider
clarity and regulatory oversight coexist within respondents? How are selected self-reported AI-
awareness items associated with the two statements? What do the findings imply for
measurement and future research with agricultural users?

1 Richa Bhattarai and others, ‘Scale, Trust, and the Digital Divide: A Systematic Review of AI and ML for
Agricultural Applications’ (2026) 9 Frontiers in Artificial Intelligence 1798896.
2 Laurens Klerkx, Emma Jakku and Pierre Labarthe, ‘A Review of Social Science on Digital Agriculture, Smart
Farming and Agriculture 4.0: New Contributions and a Future Research Agenda’ (2019) 90–91 NJAS:
Wageningen Journal of Life Sciences 100315; Callum Eastwood and others, ‘Managing Socio-Ethical
Challenges in the Development of Smart Farming: From a Fragmented to a Comprehensive Approach for
Responsible Research and Innovation’ (2019) 32 Journal of Agricultural and Environmental Ethics 741.
3 Klerkx, Jakku and Labarthe (n 2); Isabelle M Carbonell, ‘The Ethics of Big Data in Big Agriculture’ (2016)
5(1) Internet Policy Review, DOI: 10.14763/2016.1.405.
4 Viswanath Venkatesh, James YL Thong and Xin Xu, ‘Consumer Acceptance and Use of Information
Technology: Extending the Unified Theory of Acceptance and Use of Technology’ (2012) 36 MIS Quarterly
157; Roger C Mayer, James H Davis and F David Schoorman, ‘An Integrative Model o f Organizational Trust’
(1995) 20 Academy of Management Review 709.

3
The analysis is exploratory. It does not measure actual system reliability, identify an
intervention, establish a causal chain or support a national estimate. Sections 2 and 3 explain
the theory and methods; Section 4 reports the results; Sections 5 –7 interpret them and set out
implications and limitations.
2. Literature Review and Theoretical Foundation
The literature links trust in AI to both the quality of an output and the institutions, people and
data arrangements surrounding a service. This section develops those links and distinguishes
hypothetical confidence from demonstrated reliability and enforceable protection.
2.1 AI advice and agricultural knowledge
AI-supported services can organise information, produce recommendations and interact with
users in natural language. Their value depends on whether outputs fit a particular crop, soil,
season, location and decision. A review of South Asian smallholder applications warns that
claims of transformation have often advanced faster than evidence of farmer outcomes. 5 A
service that presents a general recommendation as locally verified could encourage costly
decisions, even when its language appears authoritative. Because the survey does not specify a
named service or agronomic task, this study cannot assess the quality of actual advice.
Digital agriculture also raises questions about data control, access, power and the distribution
of benefits. Farm-level operational data and identifiable personal information are not
automatically the same thing; legal treatment depends on the nature of the data and the relevant
facts. Research on agricultural data ethics argues that ownership, privacy, security and unequal
bargaining power deserve attention alongside technical performance.6
2.2 Trust and regulatory protection
Trust involves a willingness to accept vulnerability based on expectations about another actor’s
competence, integrity or benevolence. 7 In a digital service, familiarity, performance, source
clarity and the ability to obtain help may all shape reliance, while a high reported trust level
does not prove that the advice is reliable. Trust should be calibrated: users need enough
information to decide when to rely, verify or seek qualified assistance. Studies of automation
distinguish trust in a system from its objective trustworthiness.8
Perceived regulatory protection is similarly distinct from legal compliance. It concerns what a
respondent believes oversight, rules or remedies can accomplish. The survey’s oversight item
asks whether oversight can increase confidence; it does not measure whether an identified
regulator currently supervises agricultural AI or whether a remedy works in practice. Likewise,

5 Amar Singh, Vinod Kumar Shukla and Chatter Singh, ‘From Pilots to Plots: A Critical Review of Artificial
Intelligence for Smallholder Agriculture in South Asia’ (2026) DOI: 10.1177/00307270261489284.
6 Carbonell (n 3); Klerkx, Jakku and Labarthe (n 2).
7 Mayer, Davis and Schoorman (n 4).
8 John D Lee and Katrina A See, ‘Trust in Automation: Designing for Appropriate Reliance’ (2004) 46 Human
Factors 50; Kevin Anthony Hoff and Masooda Bashir, ‘Trust in Automation: Integrating Empirical Evidence on
Factors That Influence Trust’ (2015) 57 Human Factors 407.

4
the accountability item asks about willingness under a stated condition and does not measure
actual behaviour.
Provider transparency and public oversight may complement one another. A user may know
who supplied advice without having a practical remedy; conversely, an oversight framework
may exist but be unfamiliar or difficult to access. A comparison of statements about these
conditions provides a starting point for more precise investigation rather than proof that one
form of governance is superior.
2.3 Technology acceptance and digital governance
UTAUT2 explains technology acceptance using constructs including performance expectancy,
effort expectancy and facilitating conditions. It gives a useful vocabulary for discussing
perceived usefulness and support, but the present questionnaire does not operationalise the full
model, and its data do not test UTAUT2’s structural paths. 9 In agricultural advisory services,
perceived benefits and an available support channel may be relevant while not resolving
whether advice is correct or responsibility is clear.
Trust Theory adds attention to vulnerability and reliance. A digital-governance perspective
adds transparency, allocation of responsibility, oversight, data practices and accessible
channels for addressing harm. These perspectives address different questions: a useful
interface, a trusted provider and a functioning remedy are not equivalent. Integrating them
conceptually encourages the design of distinct measurements rather than an unsupported
combined scale.10
An important distinction is between regulation as a public institutional arrangement and
accountability as a practical relationship. A user may recognise an accountable provider
without knowing the legal basis of a remedy, or may support government oversight while being
uncertain about whether a complaint will be resolved. Nor is oversight necessarily a single
instrument: audit, licensing, information requirements and complaint resolution operate
through different mechanisms. The AAR2 statement does not separate them. Treating it as
proof of confidence in a named regulator would therefore exceed the measure.11
The analysis also distinguishes perceived protection from a respondent’s legal knowledge. A
positive response might express a general expectation that oversight is desirable; it does not
demonstrate that the respondent knows which protections are currently available. Legal
research can complement the survey by examining a specific service, its data flows, applicable
legal provisions and actual complaint arrangements. That assessment cannot be substituted for
the responses analysed here.

9 Venkatesh, Thong and Xu (n 4).
10 Mayer, Davis and Schoorman (n 4); Mark Bovens, ‘Analysing and Assessing Accountability: A Conceptual
Framework’ (2007) 13 European Law Journal 447; Eastwood and others (n 2).
11 Bovens (n 10).

5
2.4 AI literacy and the measurement problem
AI literacy can include recognising incorrect outputs, verifying information and understanding
bias or data risks. 12 The supplied items reflect self-reported awareness and practices, not an
objective test of competence. Higher agreement with an awareness item may coincide with
either greater confidence or greater caution, depending on context. The survey cannot
determine the effect of education or the temporal order of these attitudes.
The AAR block includes source/provider clarity, oversight, local adaptation, reporting and
accountability. The AFR block includes usefulness, responsibility, limitations, human help and
rights information. These are conceptually related but do not necessarily reflect a single latent
trait. Mixing them can obscure differences between provider characteristics and governance
mechanisms. Item-level analysis avoids interpreting a composite as a validated measure;
further work should evaluate construct structure and item interpretation before latent
modelling.13
2.5 Analytical expectations
The primary comparison considers the paired responses to AAR1, which states that clear
sources and a responsible provider would increase trust, and AAR2, which states that
regulatory oversight can increase confidence. Secondary comparisons consider AAR5 versus
AAR2 and AFR3 versus AAR3. These are exploratory item comparisons and not pre-registered
hypotheses. The correlation analysis examines three awareness or verification items against
AAR1 and AAR2, with adjustment across six tests. It does not test whether regulation causes
trust or whether AI literacy moderates an intervention.
3. Methodology
The methodology describes the available records, measurement decisions, participant
safeguards and analysis. It preserves distinctions between the original intended rural focus and
the actual population represented by the recorded categories.
3.1 Design and population
The study analyses a cross-sectional questionnaire. The unit is the individual response; no
court, farm, provider or AI system is directly observed. Respondents are coded as law students,
law graduates, advocates or legal professionals, and other relevant respondents. This mixed
composition cannot establish that the sample consists of farmers, agricultural users or rural

12 Davy Tsz Kit Ng and others, ‘Conceptualizing AI Literacy: An Exploratory Review’ (2021) 2 Computers and
Education: Artificial Intelligence 100041; Duri Long and Brian Magerko, ‘What Is AI Literacy? Competencies
and Design Considerations’ in Proceedings of the 2020 CHI Conference on Human Factors in Computing
Systems (ACM 2020).
13 Godfred O Boateng and others, ‘Best Practices for Developing and Validating Scales for Health, Social, and
Behavioral Research: A Primer’ (2018) 6 Frontiers in Public Health 149; Klaas Sijtsma, ‘On the Use, the
Misuse, and the Very Limited Usefulness of Cronbach’s Alpha’ (2009) 74 Psychometrika 107.

6
inhabitants. These limitations determine the scope of the title, results and discussion. Cross-
sectional responses do not establish temporal sequence or causation.14
3.2 Sample and eligibility
The workbook contains 385 records, each coded as consenting, aged 18 or older and aware of
AI. Respondent IDs are unique. The available file does not document dates, recruitment
channels, sampling frame, response rate or completed duplicate-screening procedures.
Consequently, the sample is not described as randomly selected or nationally representative. A
sample-size planning target of 385 does not by itself establish the precision of estimates from
an undocumented sampling design.15
The Codebook labels the respondent identifier sequence and contextual comments as synthetic.
The researcher confirmed that the numerical survey responses were collected from actual
participants. The creation and purpose of the identifier and comments remain to be
documented; the optional contextual text was excluded. This distinction is retained in the article
because it affects the reproducibility and provenance of the dataset.
3.3 Questionnaire and analytical items
The Codebook records five-point response options from strongly disagree to strongly agree.
AAR1 combines clarity of sources and identification of a responsible provider; the components
cannot be separated. AAR2 concerns the potential effect of regulatory oversight on confidence.
AAR3 asks that recommendations be adapted locally; AAR4 concerns ways to report harm;
AAR5 links strong accountability to willingness to use services. The statements do not measure
the respondent’s current level of trust in a particular system, nor do they specify a regulator or
an intervention.
AFR2 concerns identifying responsibility, AFR3 concerns communication of uncertainty and
limitations, and AFR4 concerns human help. The exploratory awareness items are AIL1,
awareness of plausible inaccuracies; AIL2, reported verification; and AIL3, awareness of bias
or incomplete data. Exact wording is included in Appendix A. The records do not document
item-source adaptation, cognitive testing or a pilot; none is assumed.16
3.4 Data collection and screening
The dataset records eligibility and closed-ended responses, but the supplied materials do not
establish the mode of administration, locations of recruitment, dates, invitations, exclusions or
response rate. The focal numerical items contain no missing responses. We report current status
and recorded location as provided, without inferring agricultural occupation or agricultural AI
use. Optional contextual text is not used for analysis.

14 Erik von Elm and others, ‘The Strengthening the Reporting of Observational Studies in Epidemiology
(STROBE) Statement: Guidelines for Reporting Observational Studies’ (2007) 4 PLOS Medicine e296.
15 William G Cochran, Sampling Techniques (3rd edn, John Wiley & Sons 1977).
16 Boateng and others (n 13).

7
3.5 Statistical analysis
Item distributions are reported with counts and endorsement percentages, where endorsement
means a response of agree or strongly agree. Neutral answers are included in non-endorsement
and must not be classified as rejection. Pointwise Wilson intervals are descriptive uncertainty
summaries and depend on sampling assumptions not established for this dataset.17
The three paired contrasts compare endorsement proportions within respondents, retaining
discordant counts and using exact McNemar tests. Holm correction is applied across the three
comparisons. Whole-record bootstrap percentile intervals describe uncertainty in each paired
difference; they are pointwise, not multiplicity-adjusted. The statements differ in wording and
scope, so a paired difference is not a direct preference ranking.18
Spearman rank correlations describe associations between AIL1 –3 and AAR1 –2, with Holm
adjustment across the six exploratory tests. These are individual ordinal responses, not latent
scores; coefficients do not imply direction of influence. Pointwise bootstrap intervals and
adjusted p-values are presented separately. Lack of statistical evidence does not establish
equivalence or absence of an association.19
The contrast of AAR1 with AAR2 uses the same persons, avoiding the error of treating their
endorsement percentages as independent samples. The exact test concerns the numbers of
respondents who endorse only one of the two statements; it does not test equality of two latent
constructs. The bootstrap resamples complete respondent records to preserve that pairing. All
confidence intervals remain conditional on the recorded sample and cannot correct unknown
selection into the study. The six correlations similarly describe rank ordering of observed
responses and are sensitive to the questionnaire’s wording and response categories.
3.6 Measurement and data quality
No composite trust, AI literacy or regulatory-protection scale is claimed. The questionnaire
groups items for a shared project, but conceptual heterogeneity means internal consistency
alone would not establish a valid construct. Respondent categories may themselves be broad,
especially “other relevant respondent,” and the workbook does not show whether a person uses
AI for agriculture. The analysis therefore avoids occupation-specific and agricultural-user
claims.
3.7 Ethical considerations
Ethical approval for this study was obtained from the Center for Academic & Professional
Career Development and Research (CAPCDR), an external institutional body, under approval
reference CAPCDR: SMUCT/LLB/SS-0020260908-19. External ethical review was sought

17 Cochran (n 15).
18 Quinn McNemar, ‘Note on the Sampling Error of the Difference between Correlated Proportions or
Percentages’ (1947) 12 Psychometrika 153; Sture Holm, ‘A Simple Sequentially Rejective Multiple Test
Procedure’ (1979) 6 Scandinavian Journal of Statistics 65; Bradley Efron and Robert J Tibshirani, An
Introduction to the Bootstrap (Chapman & Hall 1993).
19 Holm (n 18); Douglas G Altman and J Martin Bland, ‘Absence of Evidence Is Not Evidence of Absence’
(1995) 311 BMJ 485; Ronald L Wasserstein and Nicole A Lazar, ‘The ASA Statement on p -Values: Context,
Process, and Purpose’ (2016) 70 American Statistician 129.

8
because Shanto-Mariam University of Creative Technology does not have an institutional
ethics committee responsible for reviewing this type of research.
4. Results
The results describe the 385 records and compare distinct questionnaire statements. Counts
reflect respondent endorsements rather than observed service performance or legally verified
protection.
4.1 Sample composition
Table 1 summarises recorded current status and location. Neither field identifies farmers or
confirms agricultural AI use.
Table 1. Recorded respondent status and location
Measure Category Count (%)
Status Other relevant respondent 140 (36.4%)
Law student 136 (35.3%)
Law graduate 58 (15.1%)
Advocate or legal professional 51 (13.2%)
Location Dhaka 177 (46.0%)
Other city corporation 78 (20.3%)
District town 66 (17.1%)
Upazila-rural area 60 (15.6%)
Other 4 (1.0%)
N = 385. The categories reflect the original questionnaire; agriculture-related work and use of
agricultural AI are not identified.
Of 385 respondents, 60 (15.6%) are in the upazila/rural location category, while 177 (46.0%)
are in Dhaka. The location category is a single broad classification and should not be treated
as a measure of rural residence alone. This distribution limits the original proposed focus on
rural users.
4.2 Statements on trust and governance
Table 2 reports responses to the five AAR items, each of which addresses a different condition
or safeguard.

9
Table 2. Trust-condition and governance responses
Item Counts 1 / 2 / 3 / 4 / 5 Endorsed n (%) 95% CI (%)
AAR1 0 / 4 / 71 / 187 / 123 310 (80.5%) 76.3–84.2
AAR2 0 / 18 / 112 / 182 / 73 255 (66.2%) 61.4–70.8
AAR3 1 / 30 / 119 / 180 / 55 235 (61.0%) 56.1–65.8
AAR4 0 / 9 / 113 / 179 / 84 263 (68.3%) 63.5–72.8
AAR5 0 / 0 / 68 / 173 / 144 317 (82.3%) 78.2–85.8
N = 385. Endorsement means agree or strongly agree; non-endorsement includes neutral
responses. Wilson intervals are pointwise. Appendix A gives exact wording.
Source and provider clarity was endorsed by 310 respondents (80.5%); the oversight statement
by 255 (66.2%); and strong accountability increasing willingness by 317 (82.3%). Local
adaptation and a route to report harmful or inaccurate recommendations were endorsed by 235
(61.0%) and 263 (68.3%). These values describe agreement with hypothetical or normative
statements and do not measure existing trust, actual oversight or service quality.
4.3 Agricultural-advice and safeguard statements
Table 3 presents the AFR items separately to clarify the relationship between views about
usefulness and preferred service safeguards.
Table 3. Agricultural advice and safeguard responses
Item Counts 1 / 2 / 3 / 4 / 5 Endorsed n (%) 95% CI (%)
AFR1 0 / 12 / 129 / 176 / 68 244 (63.4%) 58.5–68.0
AFR2 0 / 4 / 51 / 164 / 166 330 (85.7%) 81.9–88.9
AFR3 0 / 8 / 52 / 178 / 147 325 (84.4%) 80.5–87.7
AFR4 2 / 13 / 123 / 177 / 70 247 (64.2%) 59.2–68.8
AFR5 0 / 3 / 44 / 173 / 165 338 (87.8%) 84.1–90.7
N = 385. Endorsement means agree or strongly agree; non-endorsement includes neutral
responses. Wilson intervals are pointwise. Appendix A gives exact wording.
Useful information was endorsed by 244 respondents (63.4%); provider responsibility by 330
(85.7%); communicating limitations by 325 (84.4%); access to human assistance by 247
(64.2%); and inclusion of legal-rights information by 338 (87.8%). These different statements
are not equivalent tests of specific policy designs.

10
4.4 Paired comparisons and joint responses
Table 4 reports three within- respondent contrasts. Counts marked “first only” and “second
only” describe respondents who endorsed one statement but not the other; the latter includes
neutral responses.
Table 4. Paired endorsement contrasts
Contrast First only / second only Difference pp 95% CI pp Holm p
AAR1 − AAR2 100 / 45 14.3 8.3–20.0 <.001
AAR5 − AAR2 101 / 39 16.1 10.4–22.1 <.001
AFR3 − AAR3 123 / 33 23.4 17.4–29.1 <.001
N = 385; pp = percentage points. Exact McNemar tests, Holm adjustment across three
comparisons; 10,000 respondent-level bootstrap resamples, pointwise intervals. The statements
have different wording.
Source/provider clarity exceeded the oversight statement by 14.3 percentage points.
Accountability exceeded oversight by 16.1 points. Communication of uncertainty exceeded
local adaptation by 23.4 points. Each contrast remained statistically distinguishable after Holm
correction, but the paired statements have different wording and do not establish which
safeguard would be more effective in practice.
Table 5 displays the joint endorsement pattern for source/provider clarity and regulatory
oversight.
Table 5. Joint endorsement of clarity and oversight
Response configuration Count (%)
Both endorsed 210 (54.5%)
Clarity only 100 (26.0%)
Oversight only 45 (11.7%)
Neither endorsed 30 (7.8%)
N = 385. Neither includes neutral and disagreement; these are not fixed respondent types.
A total of 210 respondents (54.5%) endorsed both; 100 (26.0%) endorsed only clarity, and 45
(11.7%) endorsed only oversight. The remaining 30 (7.8%) endorsed neither. This descriptive
pattern indicates that both conditions can coexist in one respondent’s a nswers; non-
endorsement should not be described as opposition.

11
4.5 Exploratory correlations with awareness
Table 6 presents six rank correlations between three AI-awareness or verification statements
and the two trust-related conditions. The multiple-testing correction applies to the six p-values,
while bootstrap intervals are pointwise.
Table 6. Exploratory awareness and trust-condition associations
Items Spearman ρ 95% CI p Holm p
AIL1 / AAR1 0.187 0.095–0.277 <.001 0.001
AIL1 / AAR2 0.057 -0.042–0.156 0.263 1.000
AIL2 / AAR1 0.041 -0.062–0.144 0.426 1.000
AIL2 / AAR2 0.050 -0.044–0.145 0.331 1.000
AIL3 / AAR1 0.143 0.041–0.240 0.005 0.025
AIL3 / AAR2 0.045 -0.057–0.148 0.382 1.000
N = 385. Pointwise percentile intervals from 2,000 respondent-level bootstrap resamples; p-
values Holm adjusted across six tests. Intervals are not adjusted for multiplicity.
Awareness that AI can produce plausible inaccuracies (AIL1) had a small positive association
with the source/provider clarity item (ρ = .187; Holm p = .001), as did awareness of bias or
incomplete data (AIL3; ρ = .143; Holm p = .025). Neither item showed c lear evidence of
association with the oversight item after correction; the verification item was not clearly
associated with either outcome. These findings are exploratory and their interpretation is
constrained by the different content of the statements, self-report, simultaneous measurement
and the non-representative sample. An association does not establish that awareness causes
confidence or that training would change reliance. In particular, the questionnaire does not
specify an actual regulatory authority, service or decision outcome.
5. Discussion
The findings separate provider-level transparency from institutional oversight and other
governance preferences. They also show why hypothetical agreement needs to be interpreted
within the questionnaire and sample rather than as evidence of actual agricultural service
experience.
5.1 Source clarity and oversight
Most respondents endorsed the statement about clearer sources and a responsible provider, and
a smaller proportion endorsed the oversight statement. A paired contrast confirms the
difference in response distributions within this sample. The items, however, use different
language: one asks about a personal increase in trust under specified conditions, and the other
makes a general proposition that oversight can increase confidence. Different scope, familiarity
and wording may contribute to the difference. A survey cannot determine those explanations.
Nor can it establish whether identifiable providers or regulators actually protect users.

12
This finding suggests a measurement agenda rather than a governance ranking. Provider
identity, independently verified sources, effective oversight and accessible remedy should be
tested as separate attributes of a defined service. Research on digital agriculture emphasises
that trust depends on data practices and institutional arrangements as well as technical
properties.20
5.2 Accountability and redress
A large majority endorsed the statement linking accountability with willingness to use
agricultural AI. This is a conditional self-report, not observed adoption. The questionnaire does
not define who is accountable, what obligations apply or how a user would obtain redress.
Strong accountability could mean a responsive provider, independent oversight, enforceable
rights or a combination of mechanisms. A follow-up study should specify the organisation,
process, response time and remedy before testing their effects.
The reporting item addresses a necessary practical issue: a user must be able to identify an
inaccurate or harmful recommendation and reach a body able to act. Reporting alone does not
establish that a complaint will be resolved or that compensation will be available. In a
consequential agricultural decision, a meaningful process would require a clear record of what
advice was given, what limitations were disclosed and which human or institution is
responsible.21
5.3 Local adaptation and uncertainty
Respondents endorsed communication of uncertainty more frequently than local adaptation.
This difference does not show that one is more important: the first item describes what a service
should communicate, while the second specifies how recommendations should be adapted.
Both can be necessary. A statement of uncertainty is useful only if users can understand and
act on it; localisation is valuable only if the local data and agronomic guidance are appropriate
for the task. A real evaluation should therefore test accuracy and communication together
across crops, seasons and locations.22
5.4 Awareness and reliance
The two small positive associations between risk-awareness items and the source/provider
clarity statement suggest a question for further study: whether people who recognise possible
output failures give particular attention to identifying a source. Neither item was clearly
associated with the oversight statement after correction. These patterns do not establish that the
relationships differ across outcomes; a formal comparison would be required. The correlation
analysis examines self-reported awareness of inaccurate outputs and bias, and a reported
verification practice. Such items do not constitute a validated literacy test. Even an observable
association with a trust condition would not establish whether awareness preceded confidence,
whether a third factor shaped both, or whether a training programme would alter decision-

20 Bhattarai and others (n 1); Klerkx, Jakku and Labarthe (n 2).
21 Bovens (n 10); Eastwood and others (n 2).
22 Singh, Shukla and Singh (n 5); Bhattarai and others (n 1).

13
making. The sample includes respondents with diverse legal and other backgrounds, whose
interpretations of agricultural AI may differ.23
Trust in a hypothetical agricultural service is not a demonstrated willingness to rely on advice
about pesticide application, planting or a costly input. Experimental vignettes and observed
verification tasks could provide clearer evidence about how users respond to source
information, local evidence and the possibility of human review. A comparative design should
include farmers, extension workers and affected rural communities, with sampling and
recruitment documented.
5.5 Bangladesh law and institutional context
Bangladesh’s Personal Data Protection Act 2026 supplies a potential legal context when a
digital service processes identifiable personal data. Its applicability depends on the data, actors,
legal provisions and commencement arrangements; the survey does not establish compliance,
a breach or a specific obligation for a particular agricultural AI service. 24 Farm-related
information should not automatically be assumed to be personal data, and the legal
consequences of a given processing activity require a fact-specific assessment.
Respondents may use the word “regulatory” to mean rules, oversight, professional guidance or
a complaint mechanism. The questionnaire does not identify a particular authority, and no
specific protection can be inferred from agreement with AAR2. Future work should explain
the relevant law and institutions to participants and measure their ability to use available
mechanisms, rather than relying only on a general statement of confidence.
Comparing the two source-related awareness correlations with the two oversight correlations
also illustrates the value of defining an outcome. Awareness that AI may give inaccurate
information has a clearer descriptive connection with the source/provider statement than with
the broad oversight statement, but the study has not tested the difference between these two
coefficients. A difference in their p-values alone cannot demonstrate that the correlations differ.
Further studies should compare them directly and specify the causal or predictive question
before analysis.25
The location data require equally careful use. Upazila and rural areas form one combined
category, so membership does not by itself demonstrate farming status, residence in a village
or use of an AI advisory application. The remaining categories include Dhaka, other city
corporations, district towns and other locations. Sampling respondents from varied locations
can be valuable, but the design must explicitly state whose perceptions the study aims to
represent. In this file the occupation categories are legal and other rather than agricultural, so
rural service recommendations are prospective rather than direct findings from farmers.

23 Ng and others (n 12); Wasserstein and Lazar (n 19).
24 Personal Data Protection Act 2026 (Bangladesh Act No 63 of 2026), s 1.
25 Altman and Bland (n 19); Wasserstein and Lazar (n 19).

14
5.6 Integrated interpretation
UTAUT2 highlights perceived usefulness and support, while Trust Theory directs attention to
reliance under uncertainty. Digital governance draws attention to mechanisms that make
responsibility and correction possible. These perspectives complement each other, but the
present study does not validate the integrated model or test moderation. Its empirical
contribution is to show that clear-provider, oversight, accountability and local-adaptation items
produce distinguishable distributions in a mixed sample.26
6. Theoretical and Practical Implications
The implications below are grounded in the reported item-level responses and relevant
research. They are proposals for service design and future evaluation, not demonstrated effects
of the stated safeguards.
6.1 Theoretical implications
First, trust in agricultural AI should be separated from the conditions a respondent believes
could increase trust. AAR1 and AAR5 ask conditional questions; neither gives a baseline level
of trust. Second, provider transparency and regulatory oversight should be measured separately
because they describe different institutions and responsibilities. The paired distribution
supports that analytical distinction but does not prove that the two mechanisms have different
effects.
Third, the results illustrate why technology-acceptance constructs must be adapted to a concrete
task. A general belief that agricultural AI could be useful may coexist with demands for
accountability and human help. A validated follow-up instrument could measure perceived
competence, integrity, institutional safeguards, trust and intended reliance separately, then
examine their relationships without treating a heterogeneous item block as a single scale.27
6.2 Practical and policy implications
Providers could identify the organisation responsible for the service, disclose relevant sources
and the date or limits of the information, and present uncertainty in a language and format
suitable for the intended audience. They could also test recommendations against local
agronomic evidence and provide a functioning route to qualified human help. These actions
require local testing and should not be represented as guaranteed to increase trust or improve
farm outcomes.
Policymakers and extension organisations can assess the accessibility of complaints, the
allocation of responsibilities and data protection in the relevant service context. A digital form
alone may not be usable where connectivity, literacy or language is a barrier. Testing should
include real users with different access conditions, and any legal compliance assessment should
identify the applicable rules and facts.28

26 Venkatesh, Thong and Xu (n 4); Mayer, Davis and Schoorman (n 4).
27 Mayer, Davis and Schoorman (n 4); Boateng and others (n 13).
28 Eastwood and others (n 2); Carbonell (n 3).

15
An operational evaluation could begin with several concrete questions: whether an advisory
answer identifies its source and publication date; whether a user can recognise its limitations;
whether an appropriately trained person can review a high-consequence recommendation; and
whether an incorrect answer can be reported without excessive cost or delay. These questions
translate the article’s conceptual categories into observable features. Service -specific audits
would still need to assess accuracy, inclusion and local relevance rather than assume that
disclosure alone makes advice safe.29
7. Conclusion, Limitations and Future Research
In this 385-record mixed sample, 80.5% endorsed the statement that clear sources and a
responsible provider would increase trust, 66.2% endorsed the statement that regulatory
oversight can increase confidence, and 82.3% endorsed the accountability statement. The
source/provider clarity item exceeded oversight by 14.3 percentage points, with 210
respondents endorsing both. Because these are different hypothetical statements, the findings
do not establish the comparative effectiveness of governance measures.
The study is limited by cross-sectional self-report, incomplete recruitment and fieldwork
documentation, uncertain sampling design and heterogeneous questionnaire content. Only 60
respondents are recorded in the broad upazila/rural location category, and agricultural
occupation and actual use of agricultural AI are not identified. The Codebook labels identifiers
and contextual comments synthetic, while the research team confirms actual numerical survey
responses; the source and transformation of these fields should be documented. No national,
farmer-specific or causal claim follows from this dataset.
Future work should recruit and identify agricultural users transparently, document location and
fieldwork procedures, and distinguish actual agricultural service experience from general AI
awareness. A scenario experiment could vary provider identity, source documentation,
oversight, local validation and an accessible complaint mechanism. Interviews and behavioural
tasks could examine whether users understand and act on these features. Independently
evaluated agronomic outcomes would provide evidence about accuracy and suitability that a
perception survey cannot supply.
Declarations
Ethical approval and oversight
Ethical approval for this study was obtained from the Center for Academic & Professional
Career Development and Research (CAPCDR), an external institutional body, under approval
reference CAPCDR: SMUCT/LLB/SS-0020260908-19. External ethical review was sought
because Shanto-Mariam University of Creative Technology does not have an institutional
ethics committee responsible for reviewing this type of research.

29 Bhattarai and others (n 1); Eastwood and others (n 2).

16
Informed consent
Informed consent was obtained from adult participants before questionnaire completion. The
questionnaire explained the study purpose, voluntary participation, anonymity and the option
to discontinue. All analysed records contain affirmative consent and adult-eligibility responses.
Data availability
The numerical data supporting the reported analyses are held by the research team. Requests
for de-identified data and analytical specifications may be directed to the corresponding author.
Any sharing is subject to participant confidentiality and applicable institutional requirements.
The contextual-text field was excluded from the analysis.
Conflict of interest
The authors declare that they have no conflicts of interest.
Author contributions
Md. Nasim: Lead student researcher; conceptualisation; preparation of study materials;
organisation and curation of survey data; manuscript development. The project was conducted
under the supervision of Prof. Dr Kazi Abdul Mannan.
Kazi Abdul Mannan: Conceptual and methodological guidance; supervision; interpretation
of statistical findings; critical review and editing of the manuscript.
Generative AI disclosure
Generative AI assisted manuscript drafting, language editing, organisation and the
development of analytical code. Statistical estimates were computed programmatically from
the numerical survey data. Responsibility for the study’s analysis, interpretation and reporting
rests with the authors.
Bibliography
Altman DG and Bland JM, ‘Absence of Evidence Is Not Evidence of Absence’ (1995) 311
BMJ 485
Bhattarai R and others, ‘Scale, Trust, and the Digital Divide: A Systematic Review of AI and
ML for Agricultural Applications’ (2026) 9 Frontiers in Artificial Intelligence 1798896
Boateng GO and others, ‘Best Practices for Developing and Validating Scales for Health,
Social, and Behavioral Research: A Primer’ (2018) 6 Frontiers in Public Health 149
Bovens M, ‘Analysing and Assessing Accountability: A Conceptual Framework’ (2007) 13
European Law Journal 447
Carbonell IM, ‘The Ethics of Big Data in Big Agriculture’ (2016) 5(1) Internet Policy
Review, DOI: 10.14763/2016.1.405
Cochran WG, Sampling Techniques (3rd edn, John Wiley & Sons 1977)
Eastwood C and others, ‘Managing Socio-Ethical Challenges in the Development of Smart
Farming: From a Fragmented to a Comprehensive Approach for Responsible Research and
Innovation’ (2019) 32 Journal of Agricultural and Environmental Ethics 741
Efron B and Tibshirani RJ, An Introduction to the Bootstrap (Chapman & Hall 1993)

17
Hoff KA and Bashir M, ‘Trust in Automation: Integrating Empirical Evidence on Factors
That Influence Trust’ (2015) 57 Human Factors 407
Holm S, ‘A Simple Sequentially Rejective Multiple Test Procedure’ (1979) 6 Scandinavian
Journal of Statistics 65
Klerkx L, Jakku E and Labarthe P, ‘A Review of Social Science on Digital Agriculture,
Smart Farming and Agriculture 4.0: New Contributions and a Future Research Agenda’
(2019) 90–91 NJAS: Wageningen Journal of Life Sciences 100315
Lee JD and See KA, ‘Trust in Automation: Designing for Appropriate Reliance’ (2004) 46
Human Factors 50
Long D and Magerko B, ‘What Is AI Literacy? Competencies and Design Considerations’ in
Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (ACM
2020)
Mayer RC, Davis JH and Schoorman FD, ‘An Integrative Model of Organizational Trust’
(1995) 20 Academy of Management Review 709
McNemar Q, ‘Note on the Sampling Error of the Difference between Correlated Proportions
or Percentages’ (1947) 12 Psychometrika 153
Ng DTK and others, ‘Conceptualizing AI Literacy: An Exploratory Review’ (2021) 2
Computers and Education: Artificial Intelligence 100041
Sijtsma K, ‘On the Use, the Misuse, and the Very Limited Usefulness of Cronbach’s Alpha’
(2009) 74 Psychometrika 107
Singh A, Shukla VK and Singh C, ‘From Pilots to Plots: A Critical Review of Artificial
Intelligence for Smallholder Agriculture in South Asia’ (2026) DOI:
10.1177/00307270261489284
Venkatesh V, Thong JYL and Xu X, ‘Consumer Acceptance and Use of Information
Technology: Extending the Unified Theory of Acceptance and Use of Technology’ (2012) 36
MIS Quarterly 157
von Elm E and others, ‘The Strengthening the Reporting of Observational Studies in
Epidemiology (STROBE) Statement: Guidelines for Reporting Observational Studies’ (2007)
4 PLOS Medicine e296
Wasserstein RL and Lazar NA, ‘The ASA Statement on p-Values: Context, Process, and
Purpose’ (2016) 70 American Statistician 129
Appendix A. Analytical questionnaire items
Table A1 reproduces the statements used in the analysis so that each estimate can be read
against the original question wording.
Table A1. Exact wording of the analytical items
Code Statement
AAR1 I would trust AI agricultural information more if its sources and responsible provider were clear.
AAR2 Regulatory oversight can increase confidence in digital agricultural services.
AAR3 AI agricultural recommendations should be adapted to local conditions rather than applied blindly.
AAR4 Users should have ways to report harmful or inaccurate recommendations.

18
Code Statement
AAR5 Strong accountability increases my willingness to use AI agricultural services.
AFR1 AI agricultural advisory services can help farmers access useful information.
AFR2 Farmers should know who is responsible for important advice delivered through AI systems.
AFR3 AI advisory services should clearly communicate uncertainty and limitations.
AFR4 Farmers should have access to human assistance when AI advice is inadequate.
AFR5 Legal-rights information should be incorporated into responsible digital agricultural services.
AIL1 I understand that generative AI can produce plausible but inaccurate information.
AIL2 I verify important AI-generated information before relying on it.
AIL3 I understand that AI outputs may reflect bias or incomplete data.
Source: Group 16 Codebook. Responses use five categories from strongly disagree to strongly
agree.
The items are analysed separately; their presence in one block does not establish a single latent
construct.
Table of legislation
Personal Data Protection Act 2026 (Bangladesh Act No 63 of 2026)