1. ESTABLISHING CREDIBILITY
### 1.1 AUTHOR BACKGROUNDS
The paper is authored by Taryn Wilson, Hafeni Mthoko, Yusra Adnan, and Sarina Till, all of whom are affiliated with The Independent Institute of Education's Varsity College in Durban, South Africa's largest private higher education provider. Their source of authority stems from their academic affiliation and their direct, hands-on experience as the researchers who designed, deployed, and continuously monitored the AI and IoT-enabled hydroponics systems in the field alongside rural subsistence farmers.
### 1.2 FUNDING AND CONFLICTS OF INTEREST
Funding information is provided in the Acknowledgments section, which states that the study was funded by Willie Scheepers from The Independent Institute of Education's (IIE) Varsity College. The funding source is institutional and aligns with academic and community development objectives rather than corporate or commercial interests. The paper does not report any conflicts of interest, and no hidden conflicts are indicated in the provided text; the researchers released their system designs as open-source resources and permanently donated the deployed grow tents to participating rural communities at no financial cost to the farmers.
### 1.3 PUBLISHER
The publisher of the paper is the Association for Computing Machinery (ACM). Papers bearing ACM publication rights for this conference are double-blind peer-reviewed prior to being accepted for presentation and publication in their proceedings. Therefore, this article is an academic, peer-reviewed conference paper published in the proceedings of the ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies (COMPASS '24).
### 1.4 PURPOSE
The paper is an original research article. The primary objective of the study is to test the real-world viability of an AI and IoT-enabled hydroponics system by deploying it alongside rural subsistence farmers in South Africa over a seven-month period, moving beyond the theoretical or lab-based environments typical of previous research.
The authors set out to find evidence for and convince the reader that the true complexities of agricultural technology, such as extreme weather, equipment failures, and pest infestations, only emerge during actual field deployments (hence the title, "the devil is in the deployment"). Ultimately, they argue that rather than trying to fully automate the farming process, researchers must prioritize real-world deployments and co-design systems that foster an "AI-human symbiosis," where technology assists rather than replaces the crucial manual interventions and local knowledge of the farmers.
## 2. SUMMARY OF THE ARTICLE
### 2.1 RESEARCH METHODS
The study employed a qualitative co-design methodology using "technology probes," deploying six AI and IoT-enabled hydroponics grow tents (four with farmers, two as controls) in KwaZulu-Natal and the Eastern Cape, South Africa. The primary sources of qualitative data included observations from field site visits, transcribed and translated 30-minute focus groups, and continuous WhatsApp conversations with the participating farmers. This qualitative data was subsequently evaluated using inductive thematic analysis. For the technical system, the authors built an IoT network utilizing sensors for pH, Electrical Conductivity (EC), temperature, light, and humidity, which connected to a custom Android application. Notably, the data source for the system's Artificial Intelligence was entirely synthetic; due to a lack of publicly available hydroponic datasets, the Random Forest Classifier (RFC) models used to automate the tents were trained and tested (using an 80:20 split) on synthetic data generated via Python's Numpy random module.
The methodology is based on three primary assumptions. Firstly, the authors assumed that using synthetic data to train the ML models would not adversely affect real-world accuracy, basing this on prior literature and validating it via high F1 scores and a successful control crop prior to deployment. Secondly, they assumed that deploying a pre-built, fully functional "technology probe" would overcome the "blank page problem" (ideation fatigue) and successfully stimulate co-design conversations with farmers who had no prior experience with these technologies. Finally, they assumed that their physical infrastructure choices, specifically, a gravity-fed hydroponic stand, would allow the plants to safely survive up to 48 hours without electricity, mitigating the impact of South Africa's severe rolling blackouts (load-shedding).
### 2.2 MOST IMPORTANT POINTS
The deployment revealed that AI alone is insufficient for successful hydroponic farming in real-world settings; rather, a symbiotic relationship between farmers and technology is essential. While the Random Forest Classifier models successfully automated parameter adjustments, farmers had to manually intervene to manage issues the AI could not detect, such as manually inspecting for root rot, pruning yellowing leaves, and treating aphid infestations. For example, when strong winds shook power cords loose or disconnected extractor fans, it was the farmers' manual inspections and subsequent communication with researchers that saved the crops, highlighting that the technology acts as a co-grower rather than a complete replacement for human labor.
Additionally, the authors found that unpredictable, extreme environmental conditions frequently outpaced and overwhelmed the AI's ability to maintain a controlled environment. During a severe storm in KwaZulu-Natal, heavy gusts of wind blew an entire grow tent over, flooding the system, damaging the electronics, and destroying the crop before the system could be relocated. In another instance, a sudden temperature spike from 20 degrees Celsius to 40 degrees Celsius the next day caused the plants to rapidly absorb water, which concentrated the nutrient solution in the reservoir and resulted in nutrient (tip) burn, requiring farmers to manually intervene and dilute the water.
Finally, the study demonstrated that the physical hardware and infrastructure choices are just as critical to the survival of the crop as the AI and IoT software. The researchers discovered that smaller 300-liter-per-hour water pumps frequently failed because they became clogged with plant matter, whereas larger 620-liter-per-hour pumps were far more resilient. They noted that hardware reliability ultimately supersedes software optimization, as a failure in the water pump or the physical damage of EC sensors from storms will immediately result in a failed crop, regardless of how accurately the AI models classify the tent's status.
## 3. WEAKNESSES
A major limitation identified by the authors is the high cost of the technology. At approximately 13,000 ZAR (around 694 USD) per tent, the system is prohibitively expensive for the rural subsistence farmers it aims to assist, necessitating full subsidization by the researchers. The authors also acknowledged that the physical system is unforgiving. Any hardware malfunction, whether caused by severe weather or human error (like accidentally unplugging a cord), leads to rapid crop failure within days. Furthermore, they noted that older participants (over the age of 60) initially struggled to use the mobile application and connect to the local WiFi network. This network itself was cited as a limitation, as its 100-meter range prevented true remote monitoring, a feature the farmers explicitly requested.
Beyond the authors' own acknowledgments, there are several additional methodological weaknesses. First, the sample size is extremely small. The paper reports on a deployment involving a handful of farmers in the KwaZulu-Natal province, which limits the generalizability of the findings across different demographics or agricultural zones. Second, due to a lack of publicly available data, the Machine Learning models were trained entirely on synthetic data generated via Python's Numpy random module. While this was validated in a control setting, simplistic synthetic data often fails to capture the complex, compounding anomalies of real-world environments. Finally, the AI architecture relies on four separate Random Forest Classifiers predicting individual parameters (temperature, humidity, EC, pH) in isolation. This approach is a shortcoming because it fails to account for the highly interdependent nature of these variables, such as how extreme heat directly causes plants to absorb more water, which subsequently spikes the EC (nutrient concentration) levels.
## 4. PERSONAL REFLECTION
AI in Nigerian agriculture is rapidly transitioning from theory to large-scale application. Driven by the Federal Ministry of Communications, Innovation and Digital Economy's National AI Strategy (NAIS) launched in 2025 [2], the sector is heavily targeting AI for food security. Significant advancements are already underway. For example, the AI-powered National Agro-Productivity System (NAPS) was launched in July 2026 to monitor crop yields and climate risks using satellite imagery [3]. On the ground, startups like Hello Tractor use data and AI to coordinate farming activities and predict equipment demand, connecting over 2.5 million smallholder farmers to mechanized services [4]. However, as the South African hydroponics deployment demonstrated, sophisticated software cannot function in a vacuum [1]. Nigeria faces severe infrastructural deficits, including unreliable power grids, low broadband penetration in rural areas, and extreme weather vulnerabilities. If a sudden power outage or storm physically destroys a smart farm's sensors or pumps, the AI's predictive accuracy becomes entirely irrelevant [1].
This suggests that the real constraint to adopting AI in Nigerian agriculture is not a lack of innovative software, but the fragility of our physical infrastructure. For Nigeria to truly benefit from the dividends of AI in farming, she must build robust, climate-resilient hardware systems and reliable power networks to support it. If I could change one thing to ensure maximum benefit with the least harm, I would mandate that all government-backed agricultural AI initiatives include subsidized, off-grid renewable energy solutions (like solar mini-grids) as a prerequisite for deployment. Overall, I think the technology itself is highly promising. The hydroponics paper shows that AI can drastically accelerate crop growth and reduce physical labor when conditions are ideal [1]. What is missing in the Nigerian context is the infrastructural safety net. We have the data and the ambition, but until the hardware and power realities catch up with the software, the devil will always remain in the deployment.
## 5. REFERENCES
[1] T. Wilson, H. Mthoko, Y. Adnan, and S. Till, "The Devil is in the deployment: Lessons learned while deploying an AI and IoT-enabled hydroponics grow tent with rural subsistence farmers in South Africa," in Proc. ACM SIGCAS/SIGCHI Conf. Comput. Sustain. Soc. (COMPASS '24), New Delhi, India, 2024, pp. 318-329, doi: 10.1145/3674829.3675090.
[2] OECD.AI, "National Artificial Intelligence Strategy (NAIS)," OECD.AI Policy Observatory, Apr. 22, 2026. [Online]. Available: https://oecd.ai/en/dashboards/policy-initiatives/national-artificial-intelligence-strategy-nais
[3] Africa Sustainability Matters, "Nigeria launches AI-powered national agro-productivity system to strengthen food security and climate-smart agriculture," Africa Sustainability Matters, Jul. 21, 2026. [Online]. Available: https://africasustainabilitymatters.com/nigeria-launches-ai-powered-national-agro-productivity-system-to-strengthen-food-security-and-climate-smart-agriculture/
[4] Future Africa AI, "Interview: A Nigerian Startup Using AI for Agriculture," Future Africa AI, Aug. 28, 2026. [Online]. Available: https://futureafricaai.wordpress.com/2026/08/28/interview-a-nigerian-startup-using-ai-for-agriculture/