Book a call

Staff under an AI ban break the rules more often, not less

In the largest study of AI at work, risky use was most common where the employer had banned generative AI. A ban removes the approved route and leaves the work. What holds is a short rule people have read and a tool they are allowed to use.

Manish Kumar, founder · · 4 minute read

When a business gets nervous about ChatGPT, the first instinct is to ban it. It feels like the careful choice. The evidence says it is the one that works least.

The University of Melbourne and KPMG surveyed 48,340 people in 47 countries about AI, including how they use it at work. They asked employees whether they had used AI in ways that break their employer's rules, such as uploading sensitive company information into a public tool. Then they split the answers by what the employer's rule was.

Risky AI use is most common where generative AI is banned

Employees reporting AI use that contravenes policy, by their organisation's rule on generative AI, 47 countries

  • Organisation has banned generative AI67 %
  • Organisation has a policy guiding its use56 %
  • Not sure whether there is a policy38 %
  • No policy33 %

Source: University of Melbourne and KPMG, Trust, attitudes and use of artificial intelligence: a global study 2025

Some of that gap is arithmetic. You cannot break a rule that does not exist, so the firms with no policy were always going to score lowest. That does not explain why a ban does worse than a policy. The authors' own reading is that outright bans may be ineffective, and that simply having a policy does not guarantee compliance.

Why the ban loses

A ban removes the approved route. It does not remove the reason people went looking for one.

In the same study, 70 % of employees who use AI at work said they use free, publicly available tools, and only two in five said they use tools their employer provides. More than half said they had used AI at work without knowing whether it was allowed. The quote that is due at four o'clock is still due at four o'clock, the free tool is one browser tab away, and now nobody will mention they used it.

That last part is the real cost. Under a ban, use goes quiet. Over half of employees in the study said they had hidden their use of AI or presented its output as their own. A manager who has banned AI and hears nothing about it has not solved the problem. They have lost sight of it.

What goes into the box

48 %

of employees say they have uploaded company information, such as financial, sales or customer information, into a public AI tool.

Source: University of Melbourne and KPMG, global study 2025, 47 countries

That is the global figure, and it counts anyone who has done it at all, including rarely. There is no Australian breakdown for uploading in the published snapshot. What the Australian snapshot does give is not reassuring.

Australian employees report the same habits

Share of Australian employees who say they have done each, 2025

  • Made mistakes in their work because of AI59 %
  • Relied on AI output without checking its accuracy57 %
  • Presented AI-generated content as their own51 %
  • Used AI in ways that contravene policies and guidelines44 %

Source: University of Melbourne and KPMG, global study 2025, Australia snapshot

Newer Australian surveys agree on direction. In a YouGov poll of 1,293 Australians for Salesforce in May 2026, 56 % of workers using AI at work were using tools their employer had not provided or approved. That one is commissioned by a vendor, so read it as a second opinion and not as the main evidence.

What it looks like when it goes wrong

The clearest Australian case is public, because a regulator wrote it up.

In December 2023 Victoria's Department of Families, Fairness and Housing reported that a child protection worker had used ChatGPT while drafting a report for the Children's Court. The Victorian Information Commissioner investigated and published the findings in September 2024.

The report the worker submitted described a child's doll, which had been reported to Child Protection as used by the father for sexual purposes, as a strength of the parents' efforts to provide age-appropriate toys. The Commissioner found, on the balance of probabilities, that personal and sensitive information about the family had been entered into ChatGPT and so disclosed to OpenAI. The errors did not change the outcome for the child.

Two details matter for an ordinary business. Nearly 900 of the department's staff, almost 13 % of its workforce, had visited the ChatGPT website in the six months to December 2023. And the Commissioner found the department had no controls aimed at the specific privacy risks of generative AI tools. This was not one careless person. It was a tool in general use with no rule about what could go into it.

What the regulators ask for instead

Neither of the two Australian bodies that matter here tells businesses to ban AI.

The Australian Signals Directorate's guidance on engaging with AI says organisations should be cautious about what information they and their staff give to generative AI systems, because it may be incorporated into the system's training data. Its instruction is training: staff should be taught what data can and cannot be input, such as personally identifiable information or the organisation's intellectual property, and how far the outputs can be relied on. It also suggests using private versions of a tool where they exist.

The Office of the Australian Information Commissioner's guidance on commercially available AI products, as the law firms that summarised it report, recommends as best practice that personal information, and particularly sensitive information, is not entered into publicly available generative AI tools.

Put together, the instruction is a rule about data, a tool that is allowed, and people who know both.

Four things that hold

  1. 01Find out what is in useAsk, without blame, and check what is being paid for and what is being visited. You cannot write a rule for tools you do not know about, and people only answer honestly when there is no penalty for it.
  2. 02Approve one tool for each common jobA business-grade assistant on terms that keep your inputs out of model training. Once there is an approved route, the unapproved one loses most of its appeal.
  3. 03Write the data rule on one pageWhat never goes into any AI tool: client records, staff personal information, financials, contracts, credentials. What can. Who to ask when it is not clear.
  4. 04Brief people, then check again in a quarterA rule nobody has read is the situation the survey describes. Twenty minutes with each team, and a repeat of step one after three months to see whether use has moved to the approved tool.

The Victorian regulator did order a block on generative AI tools for child protection staff. That was one of six actions in a compliance notice, and a response to the most sensitive information a government holds. For a firm of 40 people writing quotes and replying to customers, the lesson from the data is the opposite. Give people a safe way to do what they are already doing, and tell them plainly where the line is.

Sources

If you would rather hand this over: AI Readiness Check. Fixed scope · 5 working days · AU $990 ex GST. See the scope

Think you know the numbers? Guess eight published AI figures, then see each one with its source.

Keep reading

One in four firms has named who is responsible for AI

The Australian Government’s Guidance for AI Adoption sets out six practices, and the first is deciding who is accountable. Most organisations have not done it. Here is what the guidance asks a firm of 10 to 200 people to put in writing.

Read more about One in four firms has named who is responsible for AI

Not sure where to start? Book a call.

Twenty minutes. We confirm fit, propose a fixed scope and a date, and tell you if you do not need us.

Book a call

+61 490 083 850 · Wollongong, NSW