208 Million Checks, 114,000 Fines: Auditing Queensland’s AI Road Cameras

In 2024, AI cameras on Queensland roads made 208.4 million assessments looking for drivers on their phones or without seatbelts, leading to about 114,000 fines. The state’s auditor found real safeguards, including two layers of human review, but no full…

FairByDesign The Receipts Room title card: “208 Million Checks, 114,000 Fines: Auditing Queensland’s AI Road Cameras”

The Receipts Room

A letter arrives. Inside is a photo of you at the wheel, taken through your windscreen, and a fine for a seatbelt that was not worn properly. You were sure it was. Who decided you were wrong: a machine, a person, or both? (That letter is illustrative. The system behind it is real.)

If you drove in Queensland in 2024, an AI may well have looked at your car. Fixed and portable cameras photographed passing vehicles, and image-recognition software checked the pictures for two possible offences: illegal mobile phone use, and seatbelts not worn properly. That year the system made 208.4 million assessments using 9 portable devices and 12 fixed camera lanes. About 114,000 fines followed (Queensland Audit Office [QAO], 2025, Figs. 5A, 5C).

In September 2025, Queensland’s Auditor-General tabled a report in the state parliament on how the government manages the ethical risks of AI, using that camera program and a staff chatbot as its two case studies. The findings are more interesting than most AI audit headlines. The cameras came with real safeguards, including two separate layers of human review before any fine. But the department running them had never done the full ethical risk assessment that the state’s own AI policy required, and it did not even have a list of the AI systems it was using.

So here is my proposition: the cameras are not the weak point. The weak point is that nobody had done the full check, required by the state’s own policy, of what the whole system could get wrong. The state gets a filter that can look at hundreds of millions of images. Drivers carry the risk if the system around it fails. Careful at the point of use and thin at the level of governance is a common combination, and this audit lets us see it in detail.

From 208 million to 114,000

The Mobile Phone and Seatbelt Technology program is run by the Department of Transport and Main Roads, with fines handled by the Queensland Revenue Office. An external vendor owns and operates the AI and the cameras under contract (QAO, 2025, Fig. 5A). The audit’s figures for 2024 show how the work is divided between machine and people:

Stage (2024) Volume
Assessments made by the AI 208.4 million
Assessments reviewed by a person at the external vendor 2.7 million
Potential offences reviewed by the Queensland Revenue Office 137,000
Fines issued (including fines later withdrawn) 114,000

Source: QAO (2025), Figure 5C. Figures rounded by QAO. One photo can produce more than one assessment, because the AI checks for both offences.

The AI’s job is to filter. It cut the number of assessments needing a human look by 98.7 percent (QAO, 2025, Ch. 5). By our arithmetic, about 5 percent of what the vendor’s reviewers saw went on to the Revenue Office, and about 83 percent of those became fines. The audit lists the controls around that pipeline: automated cropping to hide irrelevant details and protect identities during the vendor’s review; a person at the vendor checking every photo the AI flags; “two humans” at the Revenue Office reviewing every potential offence before a fine is issued; weekly audits by the department to check the AI’s results against human assessments; quarterly performance meetings with the vendor; and the right of anyone fined to contest it in court (Fig. 5B). The vendor’s contract includes performance targets for accuracy (Ch. 5).

The auditor credited this. The program has an oversight board, committees, defined roles and named risk owners, and its governance is “well positioned” to oversee ethical risks (Ch. 5).

The receipt: the assessment nobody did

Queensland introduced an AI governance policy in September 2024. It requires government bodies to set up governance for AI and to assess the ethical risks of each AI system they use (QAO, 2025, Report summary). The camera program predates the policy, and its risks had been assessed mainly through privacy impact assessments, which covered privacy, data storage and security, discrimination and accountability (Ch. 5).

What the department had not done, a year after the policy took effect, was a full ethical risk assessment of the system. In the auditor’s words, that “means it does not know whether all ethical risks for the MPST program are identified and managed” (Ch. 5). Nor had it checked whether the program’s governance matched the new policy. Across the department, there was no AI-wide governance and “no central record or AI inventory” of where AI was used, what for, or with what risks (Ch. 5). The overall conclusion: the department “is not effectively identifying and managing aspects of ethical risks” for both of the AI systems examined (Ch. 1).

Key idea: Ethical risk assessment, including for systems already running

An ethical risk assessment asks questions a security or privacy review can miss: could this system treat some groups unfairly, can people understand and challenge its outcomes, and who is accountable when it is wrong? It should be done before launch and repeated when things change. Crucially, a new rule should also reach AI that is already live. Older systems often carry the most weight in people’s lives precisely because they have been running longest, so “it was built before the policy” is a reason to assess them, not an exemption.

The auditor found part of the problem sat with the policy itself. The state’s risk assessment tool gives no guidance on when or how often to use it, or whether it applies retrospectively to systems introduced before the policy (QAO, 2025, Ch. 4). It recommended fixing that.

The time it went wrong, outside the camera

The camera program has been wrong before, though not in the way most people would guess. In September 2023, Queensland’s then transport minister told parliament that a “design fault” in the mobile phone and seatbelt camera system had caused double demerit points to be wrongly applied for some seatbelt offences, affecting 1,842 licence holders between November 2021 and August 2023. Some had their licences suspended. The minister announced an independent review (Bailey, 2023).

That statement does not say the image recognition was at fault. The error lay in how penalties were applied once offences were recorded. It is a useful reminder that the risks of an AI system include the rules, data and processes wrapped around it. The 2025 audit did not examine that episode, but it is exactly the kind of thing a full, system-wide risk assessment is meant to catch.

The chatbot nobody was watching

The audit’s second case study got far less attention and found weaker controls. QChat is a generative AI assistant the Queensland Government built for its own staff, used to summarise documents, draft communications and analyse policy. Between February 2024 and June 2025 it recorded 383,000 conversations from 18,950 users across government, including 72,000 conversations from 3,020 transport department staff (QAO, 2025, Fig. 5A).

At the transport department, the auditor found no governance arrangements for QChat at all. A usage dashboard was available but not used to monitor how staff were using it. The department had not configured the tool’s department-specific instructions, which could reduce wrong or out-of-date answers. Controls recommended in its own August 2024 security assessment, including an AI policy, staff training and processes for privacy breaches, had not all been implemented, and uptake of its 2024 AI training course was low (Ch. 5). The auditor also noted that experts have shown chatbot safeguards like QChat’s can be bypassed with carefully crafted prompts (Ch. 5).

The complication: 2.7 million human checks

Here is the best argument against me. This program has more human checking than many AI systems, the auditor said its governance was “well positioned,” and a missing risk assessment can look like a paperwork gap rather than a danger. Fair. But the program’s strongest feature, its human review, also raises the question the audit did not answer. Reviewers at the vendor looked at 2.7 million assessments in a year. Volume like that is where human checking can quietly become a formality. Operational AI Readiness calls this “automation laundering”: a process that is technically human-approved but structured so that people mostly ratify what the machine suggests, because of volume, poor tools or pressure (Bakos, 2026, Ch. 9). We are not saying that happened here. The audit records weekly checks of AI results against human assessments, and two reviewers at the Revenue Office. But the report does not publish how often reviewers disagreed with the AI, how many fines were withdrawn or overturned, or how review quality is measured. Those numbers are what would show whether the human layers are doing real work.

The verdict

The evidence supports the proposition, with a caveat. At the point of use, Queensland built a genuinely careful pipeline. Above it, the department could not say what AI it ran, had not done the assessment its own policy required, and left a chatbot used in 72,000 staff conversations with no governance at all. The one public failure we found, the demerit point fault, came from exactly the part of the system that a full risk assessment exists to look at. The caveat: the audit examined governance, not whether individual fines were right. Good safeguards at the camera are not the same as knowing the whole system is safe, and the auditor was right to say so.

What happens next

The auditor made seven recommendations. The transport department should update its governance to match the state AI policy, add assurance checks, build visibility of its AI systems, and put monitoring and structured training around QChat. The central digital department should clarify when risk assessments apply, including to existing systems, evaluate its own policy, monitor AI use and risks across government, and share chatbot safety data with agencies. And every public sector body should carry out ethical risk assessments for AI in use or under development. All recommendations were agreed (QAO, 2025, Ch. 2).

What we still don’t know

The audit does not report the AI’s error rates, how many people were wrongly flagged, how many fines were withdrawn or successfully contested, or any finding that a specific fine was wrong. It examined governance and risk management, not the accuracy of individual decisions. The 2023 demerit point fault comes from a ministerial statement, not from the audit, and we have not seen the independent review that followed. We read the report on the Audit Office’s website; we did not open its appendix of formal agency responses. We found no public follow-up, as of October 2026, showing whether the full ethical risk assessments have now been done.

Which is the question we would most like answered: when a reviewer looks at a photo the AI has flagged, how often do they say no?

Questions you can ask

  • If you receive a camera-detected fine: look closely at the evidence it is based on. The audit confirms you can contest a notice in court if you believe it was issued incorrectly.
  • If you are a public servant using a government chatbot: ask whether your agency has rules for what you can put into it, and whether anyone reviews how it is used.
  • For agencies running long-standing AI: has each system had an ethical risk assessment since your current AI policy took effect? If not, schedule it.
  • For anyone relying on human review: measure it. Track how often reviewers overturn the AI, and treat a rate near zero as a question, not a reassurance.

Go further. Operational AI Readiness is FairByDesign’s free field manual for AI that is already live. Chapter 5, Human Oversight That Actually Works, explains how to tell real review from a rubber stamp, and Chapter 9, Approval Gates for High-Impact Actions, covers how to design reviews that people can genuinely say no in. Read it free, or test one of your own systems with the free readiness assessment, starting with domain 3, human oversight that works.

Next step: audit offices around the world are now examining government AI, and we read their reports so you do not have to. For that analysis as it lands, join Mercury Intelligence for $10 a month (or $100 a year; see what is included).


FairByDesign is reader-funded. If this helped, get the free Mercury Brief or see ways to support. Support never buys editorial control.

References

Bailey, M. (2023, September 12). Ministerial statement: Camera Detected Offence Program. Record of Proceedings, Queensland Legislative Assembly. https://documents.parliament.qld.gov.au/speeches/spk2023/Mark_Bailey-Miller-20230912-479961449437.pdf

Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/

Queensland Audit Office. (2025, September 24). Managing the ethical risks of artificial intelligence (Report 2: 2025–26). Queensland Audit Office. https://www.qao.qld.gov.au/reports-resources/reports-parliament/managing-ethical-risks-artificial-intelligence

Leave a Reply