An independent audit has forced WA Police to withdraw roughly 2,000 infringement notices issued by an AI-powered road safety camera network, after the system repeatedly misread innocent driver behaviour as illegal phone use.
Acting Police Commissioner Col Blanch confirmed on Monday that all affected fines will be revoked and refunds processed for drivers who have already paid. The notices were generated by Acusensus's 'Heads-Up' system, an Australian-built machine learning platform marketed to governments for detecting mobile phone use and seatbelt violations at scale. The audit found the AI was flagging drivers for actions like singing, gesturing while speaking to passengers, or scratching their face. For the roughly 2,000 drivers who received those notices, the financial and demerit point consequences were entirely unwarranted.
Critics were quick to reach for the word "outrageous," and it is hard to argue with them. The Heads-Up system was deployed as a fixed-camera trailer setup designed to operate with minimal human review before notices were issued. That architecture, automated capture feeding directly into automated adjudication, left very little room for the kind of common-sense check that would have caught obvious misclassifications before they became formal penalties. When algorithmic confidence scores substitute for human judgment on matters with legal consequences, the margin for error has to be near zero. Clearly it was not.
What makes this episode particularly damaging for the broader GovTech sector is not the error rate in isolation but the opacity that allowed it to persist. Machine learning models trained to detect specific physical behaviours face a well-documented challenge: the real world produces edge cases far messier than any training dataset. A driver singing along to the radio produces hand and head movements that can visually resemble phone interaction, especially at highway speed and camera angle. Without explainability baked into the system, no administrator watching a queue of flagged images can easily understand why the model fired, which makes quality assurance genuinely difficult.
Acusensus has built a credible commercial profile across multiple Australian jurisdictions and has contracts in other countries. The WA outcome will now follow every future procurement conversation the company has. Government clients evaluating automated enforcement tools will demand documented accuracy benchmarks, independent validation, and clearer liability frameworks before signing. That is not necessarily a bad outcome for the market long-term, but it will slow the sales cycle considerably.
A liability question governments cannot ignore
Beyond the vendor relationship, the WA case surfaces a liability structure that state governments have not fully stress-tested. When an AI system issues an infringement, who carries legal responsibility for a wrongful notice? The vendor who built the model, the agency that deployed it, or the statutory framework that authorised automated enforcement in the first place? WA Police's decision to withdraw the fines en masse is the pragmatic response, but it sidesteps the deeper question. If even a fraction of those 2,000 drivers paid their fines without contesting, then a government entity collected revenue on the strength of a demonstrably flawed algorithm. That is territory most legal systems are not yet equipped to adjudicate cleanly.
Public trust in automated enforcement is fragile. Speed cameras took years to gain legitimacy, and they operate on physics, not machine learning inference. AI-based behavioural detection is a considerably harder sell, and an early, high-visibility failure in a major Australian state will give advocacy groups and defence lawyers in every jurisdiction fresh ammunition to challenge similar systems.
For Acusensus and its peers, the path forward almost certainly involves independent third-party auditing before deployment rather than after, human review thresholds built into the adjudication workflow, and transparent reporting of false positive rates as a standard procurement deliverable. For governments, the lesson is that contracting an AI system does not transfer accountability. The agency whose letterhead appears on the fine owns the consequences when the algorithm gets it wrong. WA learned that at the cost of 2,000 notices. Other jurisdictions would be wise to build the safeguards before they face the same reckoning.
Also read: ChatGPT's new assertive persona is alienating the users paying $200 a month for it • OpenRouter data shows most AI token consumption is now driven by everyday users not developers • OpenClaw and its clones are impressive toys that serious developers stopped needing before they ever arrived
