Source-linked AI summary
Human-AI Collaboration in Requirements Engineering: Evidence of the Negative Effect of LLMs on Requirements Inspection
Giovanna Broccia, Julian Frattini, Chetan Arora, Maurice H. ter Beek, Alessandro Fantechi, Andreas Vogelsang, Alessio Ferrari
TL;DR
Requirements inspection research has provided limited evidence about how LLMs affect human inspectors rather than model output quality alone. Using a controlled crossover experiment with novice inspectors, this paper finds that LLM support reduces smell-detection accuracy without significantly changing classification accuracy or task duration, and may slow learning when introduced first.
Problem
Empirical evidence is limited on LLMs as collaborative assistants in human-performed requirements inspection, despite their growing use in requirements engineering.
Method
A controlled crossover experiment evaluated novice inspectors performing requirements inspection with and without ChatGPT across smell identification, severity classification, and inspection duration.
Results
LLM support negatively affected smell detection accuracy, had no significant effect on smell classification or task duration, and reduced the learning effect when used first.
Takeaways & Limitations
LLM support does not necessarily improve novice RI and may hinder systematic inspection-skill development when relied on from the outset.
Takeaways & Limitations
The experiment used bachelor students, simpler requirements, and a simplified RI task, so observed effects may differ from those for experienced industrial practitioners.
Abstract
from arXiv · showhide
Background. Requirements inspection (RI) is a well-established practice for detecting potential defects in requirements artifacts early in the software lifecycle. Recent advances in large language models (LLMs) have stimulated interest in their potential to support requirements engineering (RE) tasks. However, empirical evidence on the effects of LLMs when used as collaborative assistants in human-performed RI remains scarce. Aims. We aim to investigate the impact of LLM support on human-performed RI, considering inspection effectiveness in terms of smell identification and severity classification (i.e., nocuous vs innocuous), as well as inspection duration. Method. We conducted a controlled crossover design experiment with 34 participants, who inspected textual specifications with and without LLM support, identifying and classifying requirements smells while recording inspection time. We analyzed the data using one Bayesian regression model per outcome variable, accounting for validity threats induced by the crossover design as well as covariates and mediators. Results. Results show that LLM support negatively affects smell detection accuracy but has no significant effect on smell classification or task duration. A learning effect is present across experimental periods, but reduced when RI is first performed with LLM support. Conclusions. Our findings provide empirical evidence that LLM support does not necessarily improve performance and may, instead, hinder it for novice inspectors. Moreover, the results suggest that learning RI with LLM-support from the beginning may slow down the skill acquisition process, implying threats for LLM-supported learning.
1 Introduction
Requirements inspection applies systematic software inspection to requirements artifacts, where early defect detection can prevent costly downstream problems. This study addresses limited evidence about LLMs as collaborative assistants in human-performed RI and reports negative effects on novice inspectors’ smell detection.
- Requirements inspection systematically examines requirements artifacts to identify potential defects before they propagate downstream.Requirements smells include indicators such as ambiguity, vagueness, and unverifiability.
- Existing LLM research mainly evaluates model outputs, leaving limited empirical evidence about human–LLM collaboration during requirements inspection.
- The controlled crossover experiment examined ChatGPT-supported and unsupported RI by novice inspectors, focusing on smell identification, nocuous-versus-innocuous classification, and inspection duration.
- ChatGPT support reduced smell detection accuracy but had no significant effect on smell severity classification.
- Participants generally improved on their second inspection, but this learning effect was reduced when ChatGPT support was used first.
- The study contributes empirical evidence about LLM effects, learning and order effects, and practical use of LLMs in novice RI.
2 Background and Related Work
Requirements inspection seeks early detection of quality problems, with requirements smells guiding attention and nocuousness distinguishing likely misunderstanding risks from benign issues. Prior work demonstrates LLM potential but largely evaluates model outputs rather than human-centered inspection processes, which this study investigates.
- Background: Requirements inspection examines requirements artifacts for defects, inconsistencies, and quality indicators that can cause downstream rework, delays, and quality problems.
- Background: Requirements smells are observable indicators of potential quality issues that guide inspectors toward requirements needing closer scrutiny.
- Background: Nocuous smells may produce divergent interpretations and incorrect implementations, whereas innocuous smells are typically interpreted consistently.
- Related Work: Prior research includes LLM-based detection, categorization, explanation, and transformation of requirements quality issues across varied evaluation settings.
- Related Work: Some studies report improved requirements-analysis performance or classification with LLMs, while others caution that comparisons with human inspection are constrained by necessary adaptations.
- Related Work: Existing work largely assesses LLM outputs, whereas this study evaluates LLMs as support tools within human-centered RI and examines effectiveness, learning, order effects, and duration.
3 Study Design
The study investigates how LLM support affects novice inspectors’ RI time, smell identification, classification, and learning behavior. It uses a controlled experimental design with balanced participant groups and defined smell categories.
- Goal and Research Questions: The experiment aims to understand LLM support for novice RI and its impact on learning behavior.
- Goal and Research Questions: The research questions address inspection duration, smell-identification accuracy, and classification accuracy under LLM support.
- Selection of Subjects: Participants were BSc Computer Engineering students recruited from an Industrial Computer Science course.
- Selection of Subjects: Two experimental groups contained 17 participants each and were optimally matched on self-reported RI and LLM proficiency.
- Requirements Smells: The study distinguishes syntactic smells, such as vague pronouns and passive voice, from semantic smells, such as inconsistencies and ambiguities.
3.3 Experimental Objects and Tasks
Participants inspected comparable requirements documents using reference materials and a structured spreadsheet, identifying predefined smells, classifying their severity, and recording inspection time. The documents and annotations were designed to support controlled performance measurement while simplifying analysis.
- Inspection Tasks: Participants inspected each requirement, identified predefined smell types, classified detected smells as nocuous or innocuous, and recorded task start and end times.
- Requirements Documents: The two documents covered Arkanoid and Snake implementations and contained 40 requirements with 21 smells and 39 requirements with 19 smells, respectively.
- Requirements Documents: Each document covered nine smell types, with each requirement containing at most one smell to simplify participant annotation and performance analysis.
- Ground Truth: Three authors reconciled annotation discrepancies through iterative consensus-oriented discussion because requirements smells are subjective and context-dependent.
- Study Materials: Participants received definitions and examples for smell types and severity classes before using the test spreadsheet.
- Study Materials: The test spreadsheet used one row per requirement, smell-type columns, a severity field, and start- and end-time fields.
- Questionnaires: Pre-test questionnaires collected demographics and self-reported RI and LLM proficiency, while post-tests gathered condition-specific feedback and preferences.
3.4 Choice of Design and Variables
The study used a two-period crossover design to compare requirements inspection with and without ChatGPT support, while distinguishing exposure, outcomes, crossover factors, and covariates.
- Experimental design: The experiment used two periods separated by a one-week washout, reversing ChatGPT-support conditions between groups.Group 1 inspected without ChatGPT first and with support second; group 2 followed the reverse order.
- Variables: The study’s main exposure was LLM support, with detection performance, classification performance, and duration as dependent variables.
- Variables: Crossover-design factors and additional covariates were included alongside the exposure and outcome variables.
3.5 Experimental Procedure
Participants received training and then performed requirements-inspection tasks that measured smell identification and severity classification against predefined ground truth.
- Training: The training covered requirements inspection, requirements smells, nocuous-versus-innocuous distinctions, and basic prompt engineering.Gamification framed the activity as a simulated hiring scenario.
- Smell identification: Participants marked the presence or absence of predefined smell types for each requirement.Performance was quantified by comparing annotations with ground truth.
- Smell identification: Identification performance used participant-level F1-scores computed for each smell type from aligned annotations and ground-truth labels.True positives, false positives, and false negatives were derived using requirement identifiers.
- Severity classification: Severity classification was evaluated only for correctly detected smells, treating the nocuous label as a binary classification task.This separation avoided conflating detection errors with classification errors.
- Severity classification: Classification performance was summarized using F1-scores.
3.6 Data Analysis
The analysis combined Bayesian data analysis with statistical causal inference, using DAG-based adjustment and one Bayesian regression model per research question to estimate controlled effects.
- Analysis framework: Bayesian data analysis was used within a statistical causal-inference framework to analyze the collected outcomes.The approach was selected because the outcome variables were not normally distributed.
- Causal modeling: A separate DAG was specified for each research question around ChatGPT use and the relevant outcome.The outcomes were duration, detection performance, and classification performance.
- Causal modeling: D-separation identified adjustment variables needed to de-confound effects, while LLM usage was modeled as a mediator and other variables as covariates.For duration, period and sequence were identified as confounders.
- Regression modeling: One Bayesian regression model was designed per research question using the variables in each adjustment set.
- Outcome distributions: Duration was modeled with a negative-binomial distribution, while detection and classification F1-scores used zero-one-inflated beta distributions.The beta variant accounted for observed scores at 0.0 and 1.0.
- Regression modeling: The regression model included population- and participant-level intercepts, main effects, and interactions in which one variable moderates another’s effect.
- Prior specification: Uninformative priors were centered at 0 with variance 1 because strong empirical evidence for the modeled factors was lacking.The priors were expressed as N(µ = 0, σ = 1).
- Effect interpretation: Marginal and conditional effects were calculated by varying one or two factors while holding other factors at representative levels.This produced interpretable outcome-space estimates.
4 Study Results
ChatGPT support did not significantly change inspection duration or smell-classification accuracy, but it reduced smell-detection accuracy. Performance improved across periods, with a stronger learning effect when participants began without ChatGPT.
- ChatGPT support did not significantly change task duration, whose 95% credibility intervals overlapped.
- Control-group participants achieved about 8% higher average F1-scores for detecting requirements smells.The 95% credibility intervals still overlapped.
- Detection accuracy improved by more than 10% in Period 2 across both experimental sequences.Participants who started without ChatGPT improved by approximately 12%, compared with approximately 7% for those who started with ChatGPT.
- Longer task durations were associated with lower detection F1-scores, while other covariates showed no effect.
- ChatGPT had virtually no effect on classification accuracy for correctly detected smells.Classification accuracy nevertheless improved between periods, with a weaker carryover pattern than for detection.
- Overall, participants took equally long with LLM support but detected smells less precisely, with ΔF1 ≈−8%; classification accuracy remained unaffected.Starting with LLM support reduced the maturation-based improvement to 6% instead of 12%.
5 Discussion
The discussion interprets reduced detection accuracy and weaker learning after early ChatGPT use as evidence that LLM assistance can hinder novice inspectors’ engagement and skill development. It recommends cautious, staged use while emphasizing that findings require further validation with professionals and industrial requirements.
- 5.1 Key Observations: Reduced detection accuracy may reflect novice inspectors delegating inspection effort to ChatGPT instead of engaging deeply with requirements.The authors identify over-reliance and automation bias as a plausible explanation.
- 5.1 Key Observations: Detection and severity classification are cognitively distinct activities because ChatGPT affected detection but not classification after successful smell identification.
- 5.1 Key Observations: Repeated inspection improved performance, but beginning with LLM support reduced learning and may hinder development of inspection skills.The discussion links this pattern to potentially short-circuiting reflective processes used to internalize strategies and quality heuristics.
- 5.1 Key Observations: Longer inspection duration was associated with lower performance rather than greater accuracy.The authors interpret longer durations as possible signals of uncertainty, weak strategy, or difficulty understanding requirements, while treating the association as non-causal.
- 5.2 Implications for Practice, Research, and Teaching: Practitioners should avoid uncritical LLM adoption in RI and consider restricting assistance to post-inspection review or justification refinement.
- 5.2 Implications for Practice, Research, and Teaching: LLMs may support novice onboarding when used for learning and orientation rather than as substitutes for human reasoning or task execution.The discussion still emphasizes sustained human mentoring and supervision.
- 5.2 Implications for Practice, Research, and Teaching: Future research should test constrained or staged LLM use and replicate the study with professional inspectors and more complex industrial requirements.These studies would assess generalizability and whether expertise moderates LLM-support effects.
- 5.2 Implications for Practice, Research, and Teaching: Teaching should establish manual inspection skills before introducing AI assistance, which can later support discussion, reflection, or comparison.
6 Threats to Validity
The study identifies construct, internal, external, and conclusion validity threats affecting how its results should be interpreted. These include measurement limitations, unmeasured confounding, novice-student and simplified-task boundaries, and a limited sample with subjective analytical choices.
- Construct Validity: Self-reported proficiency measures were coarse proxies, and ground-truth annotations lacked formal inter-rater agreement metrics.The authors mitigated annotation concerns through collective review and consensus discussions.
- Internal Validity: Unmeasured general skill and motivation may confound the observed association between inspection duration and other outcomes.The authors therefore assume the association between longer duration and lower detection accuracy is non-causal.
- External Validity: The experiment used bachelor students, simpler-than-industrial requirements, at most one smell per requirement, and a constrained inspection task.These differences from real RI practice limit generalizability to experienced practitioners and complex industrial settings.
- Conclusion Validity: The sample included 34 participants, and the Bayesian analysis involved subjective choices about distributions, model parameters, and prior probabilities.Bayesian analysis preserves uncertainty through credibility intervals and makes analytical decisions explicit for review.
7 Conclusions
The paper reports a controlled crossover experiment showing that LLM support does not necessarily improve requirements inspection for novice inspectors. It recommends investigating constrained or staged integration strategies and validating the findings with professionals and more complex requirements.
- Conclusions: LLM support negatively affected smell detection accuracy but had no significant effect on smell classification or inspection duration.The study also found a reduced learning effect when LLM support was used from the outset.
- Future Work: Future research should examine constrained, staged, or post-hoc LLM usage to mitigate the negative effects observed in this study.The paper also proposes studying interaction designs, prompting strategies, and explanation mechanisms.
- Future Work: Replications with professional inspectors and more complex industrial requirements are needed to assess generalizability and the role of expertise.The paper further calls for evaluating how expertise moderates LLM-support effects.
8 Data Availability
The paper makes its replication package publicly available.
- Data Availability: The replication package is publicly available.