The camera that records is not the camera that complies: Singapore's WSH Act after the 2024 enforcement update
Fifty thousand Singapore dollars. That is the WSH Act fine ceiling for an individual after the June 2025 amendment, with corporate exposure up to half a million per breach and personal imprisonment available to the courts. A safety officer at a Changi Airport contractor was sentenced to two months' jail in 2020 for failing to recommend reasonably practicable measures. The personal liability is not theoretical. The MOM enforcement update in 2024 sharpened the test of what counts as a reasonably practicable measure on the monitoring side, and the gap that update opened is what most facilities have not closed yet.
The shorthand version of the gap: a CCTV system records what happened. A monitoring system detects that something is happening. The first satisfies an evidence-after-the-fact question. The second is what the WSH Act asks of the dutyholder under Section 12 — that the workplace, so far as is reasonably practicable, be safe and without risk to health. Recording an injury on tape is not a control. Detecting the precursor to the injury, alerting the worker, and logging the response is.
This post is about the architectural difference between those two systems, what the regulator is testing for now, and what the deployment looks like for facilities that need to close the gap before an enforcement visit.
What the 2024 enforcement update actually shifted
The WSH Act has always required dutyholders to take reasonably practicable measures. The text has not changed. The interpretation has. MOM's 2024 enforcement guidance reflects what inspectors are now testing during site visits: not whether the dutyholder filed a hazard register, conducted a risk assessment, and posted PPE signage, but whether the controls those documents claim are functioning at the moment the inspector walks the floor.
The shift is from documentation compliance to operational compliance. A facility with a fully populated hazard register and a binder of risk assessments can still fail an inspection if the controls are not demonstrably running. The test is no longer "did you write the plan." It is "is the plan in operation right now, and can you prove it without a paper trail that depends on the worker who took the shift you are inspecting today."
The HDB precedent reinforces the direction. Since 2023, HDB has required AI video analytics for safety monitoring on all public housing construction sites under its purview. The government has signalled in its own contracting practice what suitable monitoring looks like when it is paying the bill. Private-sector dutyholders making the same compliance argument have an existence proof that AI video analytics meets the standard, because the government wrote the standard for itself first.
CCTV is evidence; detection is control
Practitioner discussions on Singapore safety forums repeatedly name the same architectural confusion: the safety manager points to the CCTV system as evidence of monitoring. The lawyer points to the same CCTV system as evidence of recording. The two roles are different. The CCTV is admissible at an inquiry. It is not a control measure under the Act.
A control measure has to operate before the harm event. The ladder fall, the forklift-pedestrian collision, the unprotected zone entry, the missing face shield at the chemical line — each of these has an observable precursor visible to a camera and detectable by an AI model trained on those precursor states. The control function is the alert that fires in the few seconds between precursor and injury, to the worker who can move and to the supervisor who can intervene. A camera that produces footage for a post-incident review has done nothing about the incident itself.
This is the same architectural distinction we covered in the post on what HyperQ AI Safety is and the moment-before window the system is built for. The WSH Act test maps onto that architecture cleanly: the dutyholder demonstrates a reasonably practicable measure when the system reaches the worker before the harm, not when the system reaches the file after it.
The four detection categories MOM inspections are looking for
The categories that currently dominate enforcement attention are not exotic. They are the ones the construction- and manufacturing-incident statistics have repeatedly named, and they are also the categories where AI vision detection is mature enough to deploy in production today.
PPE compliance at zone entry. Hard hat, hi-vis, eye protection, hearing protection, and chemical PPE where the zone requires it. The detection has to be specific to the PPE class — generic body-keypoint models do not satisfy a chemical-zone face-shield-in-position check, and the inspector will know the difference if asked.
Forklift and pedestrian proximity. The detection is the approach event, not the collision. The intervention window is the few seconds in which the alert reaches the pedestrian's wearable, the forklift driver's HUD, or the line supervisor's terminal.
Zone breach into restricted areas. Energised electrical, confined space, hot work, chemical storage. The control is the alert at the entry boundary plus a record of the entry event with timestamp and worker identification, so the dutyholder can demonstrate a functioning permit-to-work system, not just a posted permit.
Lone worker presence in a hazardous zone. The control combines the camera-side detection of a single worker in a zone that requires a buddy with the wearable-side biometric monitoring that flags incipient distress. The architectural pattern is the same one applied in the AI safety monitoring approach for chemical and process environments and in adjacent high-risk environments.
These four categories are the surface area where most current enforcement findings land. They are also the surface area where deployment is fastest because the camera infrastructure usually already exists. The architectural addition is the inference layer and the alert path.
The audit trail is the compliance position
The MOM-side question that follows any monitoring claim is documentary: show the inspector that the controls were operating at the time of the incident. A continuous monitoring layer with timestamped video, classification confidence, model version, and alert routing produces an evidence base that does not depend on the worker's recollection or the supervisor's after-the-fact reconstruction. The inspector receives the same evidence the dutyholder used to manage the shift in real time.
This is the regulatory exposure the companion post on the Korean Serious Accidents Punishment Act regional precedent covers in detail. SAPA in Korea, WSH Act in Singapore, CDM 2024 and OSHA Amendment Act in Malaysia — the regional direction of travel is consistent. The penalties scale to executive liability. The standard for evidence is moving from filed documents to demonstrable continuous controls. The dutyholder who has not closed the gap before the inspection is the one who answers the question after.
What HyperQ AI Safety does inside this regulatory frame
The architecture is purpose-built for the four detection categories above. The Visual Language Model with PEFT fine-tuning is trained on the precursor states for falls, fires, intrusion, and PPE non-compliance. ONVIF auto-recognition picks up existing CCTV in roughly one hour of deployment. The IP68-rated smartband at 250 US dollars per worker (4G/WiFi model with firmware and app) is the worker-side alert channel — vibration cue at the wrist when the precursor is detected, plus continuous monitoring of heart rate, SpO2, skin temperature, and blood pressure to catch the biometric anomalies that visual detection alone cannot see.
Hardware footprint runs 30 to 50 percent lower than hardware-locked safety platforms. The architecture is air-gappable when the facility's data-sovereignty position requires it, which in practice covers most defence, pharmaceutical, and semiconductor environments operating in Singapore. Dashboard data is downstream of the worker-first alert routing — managers receive the aggregated trend, the worker receives the directional signal first.
The product does not make permit decisions, draft incident investigation conclusions, or replace the qualified safety professional. The professional makes the calls. The system provides the data and the alert. The boundary is intentional and aligns with the practitioner consensus on what AI is accepted for in EHS work and what it is not.
What you can verify before any commitment
Send the floor plan for one zone — assembly line, packaging hall, chemical area, restricted access, lone-worker zone — and the inventory of the existing camera and sensor coverage. Within two weeks, we map the zone against the four detection categories, identify where the current detection-to-alert window leaves workers stranded, and produce a written hazard register tied to the WSH Act's "reasonably practicable" test for that specific layout.
The deployment to validate the architecture on a single zone is a one-hour install on existing ONVIF-compatible cameras, with the smartband layer added per worker as the worker-side alert channel. The retraining workflow is owned by the customer's safety team after handover, which keeps the audit trail under the dutyholder's direct control rather than at a vendor's discretion.
The difference between a facility with cameras and a facility with controls is not a hardware difference. It is the architecture decision about who receives the alert, when, and what the system has logged about whether the alert produced a response. In Singapore after the 2024 enforcement update, only one of those positions is defensible.