Mehrdad Arashpour

Purpose: This study presents one of the first time-series, cross-industry analyses of the artificial intelligence (AI) job market for the construction sector. It benchmarks construction against both digital and traditional industries to interrogate prevailing assumptions of linear diffusion and temporal catch-up in AI workforce adoption. Design/methodology/approach: Drawing on 51 months of Lightcast job postings (January 2021–March 2025), with construction defined as ANZSIC Division E, the study employs an interpretable analytical toolkit comprising Seasonal-Trend decomposition using Loess decomposition (period = 12), month-of-year bootstrap tests on detrended series, cross-correlation analysis with bootstrap confidence bands, Newey–West trend estimation on sectoral share ratios and conservative one-step Exponential Smoothing State-Space (ETS) forecasts. All results are re-estimated excluding the partial final month (March 2025) as a sensitivity check. Findings: Construction's AI hiring is very sparse, accounting for under 1% of AI postings and is more volatile than hiring in peer sectors, but it grew fastest: annual construction postings rose more than fourfold between 2021 and 2024, against increases of 13% and 9% in the digital and traditional sectors, whose postings peaked in 2022. Construction's share of AI postings relative to both comparators therefore rises (converges) rather than falls. No delayed-follower pattern is detected: no lag at which the comparator sectors lead construction is statistically reliable, and the short-run associations that are detected are weak, near-contemporaneous and sensitive to the sample end. Unlike the comparator sectors, construction shows no statistically significant within-year seasonality and short-horizon ETS forecasts do not improve on a seasonal-naïve benchmark. These findings are robust to the exclusion of the partial final month of data. Construction thus appears as a low-base, fast-growing entrant with its own hiring dynamics, not a “laggard” following digital-sector cycles with a delay. Practical implications: The findings suggest that construction firms should prioritise internal capability building and domain-specific AI skills where external hiring is thin, support flexible, year-round upskilling, and design policy and vendor strategies tailored to the sector's operational realities rather than importing digital-sector playbooks. Originality/value: This study provides among the first data-driven, time-series evidence that qualifies the dominant “late adopter” narrative in construction. It shows that construction's low level of AI hiring coexists with the fastest growth and a rising share, without a delayed-follower timing pattern, distinguishing a low starting point from a lagging trajectory and offering an empirically grounded alternative to technologically deterministic accounts of cross-industry AI adoption.

Timely and accurate recognition of workplace hazards is critical for ensuring safety in dynamic and high-risk environments such as construction sites. However, existing video-based approaches often rely on extensive annotations and fully supervised training, and many are tailored to specific hazard types, which limits scalability to rare or diverse scenarios. This study investigates whether a data-efficient vision–language framework can support early-stage hazard recognition from short preincident video segments under limited supervision. The proposed method builds on a video-adapted contrastive language-image pre-training (CLIP)-based vision–language backbone and introduces a parameter-efficient adaptation strategy that combines learnable visual prompting with low-rank adaptation (LoRA) applied to the text encoder. Learnable visual prompts capture global, summary, and local spatiotemporal cues from short video clips, and LoRA enables lightweight semantic adaptation without fine-tuning the full backbone. This design preserves most pretrained parameters and introduces only a small number of additional trainable parameters, making it well suited to few-shot learning in data-scarce safety scenarios. To evaluate early recognition capability, this paper further examines performance under shorter temporal observation windows by reducing the amount of video available for prediction. Experiments on a curated real-world hazard video data set show that the proposed method substantially improved few-shot recognition compared with a prompt-free baseline across five hazard categories and 5-, 10-, and 15-shot supervision. Overall, the findings suggest that combining temporal visual prompting with parameter-efficient text adaptation offers a scalable engineering pathway toward assistive hazard recognition in data-scarce environments, with promising but preliminary performance.