Farm Security’s Evidence Gap—and What AI Gate Monitoring Has Yet to Prove

A farm gate can be a point of entry, an escape route or simply a place where nobody knows what happened overnight. Yet the evidence needed to answer basic questions—how often gates are left open, who or what passes through, and what consequences follow—is remarkably sparse, according to a new review of research and public reports on UK farm security.

The review asks five related questions: what is known about farm theft, livestock predation and unauthorised access; how reliable the available records are; whether security measures work; what camera-based AI can detect; and what evidence would be needed to judge continuous monitoring.

The authors searched Web of Science and Scopus in November 2026, then added grey literature and hand-searched sources. From 1,183 database records, they screened 761 unique items and included 96 in their synthesis, alongside 24 grey-literature sources. The resulting review covers crime, animal attacks, technology, reporting and farmer wellbeing—not a trial of a monitoring system.

A costly problem with an incomplete record

The review describes a mismatch between the apparent costs of rural crime and the quality of the data behind them. NFU Mutual estimated UK rural crime at £41.5 million in 2025, including about £4.5 million in livestock theft. These are insurer-based estimates, not totals from a comprehensive national crime register. The review notes that the Home Office says agricultural crime cannot be separately identified in its central data.

Other estimates come from surveys with different populations and methods, so they cannot simply be combined. For example, the Home Office’s 2017 survey covered agriculture, forestry and fishing businesses, while some national rural-crime surveys relied on people choosing to respond. In one UK study of farmer wellbeing, 80 people completed a survey; 72 said they had been crime victims. The authors caution that its low response rate means it cannot establish how common crime or its effects are among farmers generally.

The review also finds little evidence about gates themselves. One database study concerned livestock-related fatalities, not gate use. No study identified by the authors measured how often gates are left open, what happens afterward, or whether an intervention changes that pattern.

What security measures have—and haven’t—shown

The review finds more consistent evidence for physical barriers and farm management in some predator-control settings than for security devices. Evidence on cameras, trackers, sensors and marking is especially thin: the authors found no independent evaluation showing that these measures reduce farm losses. Some sensor prototypes report strong detection performance in developer-led tests, but those results do not demonstrate fewer thefts or attacks.

Research on predator deterrents offers a caution against assuming that a device will keep working. A systematic review of anti-predator interventions found that light and sound deterrents could lose effectiveness within one to five months. Some other measures had mixed or context-dependent results. These findings concern particular interventions and settings; they do not establish that every deterrent fails or that gate monitoring would fail.

AI camera research has demonstrated detection on ecological image sets and other non-farm settings, but the review found no study evaluating an AI camera system at a farm gate. Performance can depend on conditions such as distance, obstruction, vegetation and time of day. In one comparison cited by the review, machine learning detected about 41% of wildlife, compared with about 81% for human analysts. That result is a warning about generalising from one dataset or model, not a score for every modern camera system.

A proposal, not a proven remedy

The authors argue that continuous, AI-assisted gate monitoring could create time-stamped records of people, vehicles and animals, helping farmers investigate events and test other security measures. They present this as an inference from gaps in the evidence—not as a demonstrated reduction in theft, predation or livestock loss. A record of detections could improve visibility, but detection alone does not prevent crime, and cameras can miss events or generate false alarms.

The distinction matters because the paper’s authors are not only reviewing technology in the abstract: they mention a companion case study of a 14-camera deployment using the Fieldmark AI platform. That deployment is not evidence in this review, and the review says it had not yet measured accuracy or false-positive rates. The authors’ positive assessment of potential benefits should therefore be read as their interpretation, not an independently established outcome.

The review has important methodological limits. It searched only two bibliographic databases, restricted the literature to English and 2000–2026, and did not formally assess risk of bias or grade certainty. Screening and extraction used a language model under author direction, with author checks; a post hoc rule was used to decide which uncertain records to include. The authors also acknowledge likely publication bias and the use of self-selected surveys and commercially interested sources. The paper is a preprint, so its findings have not necessarily been peer reviewed.

To establish whether gate monitoring works, the next step is an independent, pre-registered evaluation that measures accuracy, missed events, false alarms, costs and actual changes in loss. Until then, the review’s strongest conclusion is not that AI security has been proven, but that UK farm security lacks the reliable event-level evidence needed to judge what works.