In Australia, the UK and the EU, an AI model of a named person is personal information about that person. That includes every guess it makes about them, right or wrong. It’s true even when every input came from the public web. So the question isn’t whether it’s legal in the abstract. It’s whether you can do what the law asks of anyone holding someone’s personal information: tell them, use it only for the purpose you stated, and delete it afterwards. For a firm working from its own records about the client, the answer is usually no. The safer route is to model the role, not the person.
Why ‘it’s all public’ isn’t the answer
The Australian Privacy Act defines personal information as ‘information or an opinion about an identified individual’, whether it is true or not (Privacy Act 1988, s 6(1)).
The privacy regulator, the OAIC, has applied that definition to AI directly. Its guidance on commercially available AI products says that information an AI generates or infers about a person ‘constitutes personal information’, including when it’s wrong. Generating that information counts as collecting it, so the collection rules apply.
Those rules apply to public sources too. The OAIC’s guidelines say that information being on the internet ‘does not allow it to be collected and used in whatever way’ an organisation chooses. They also say it ‘would usually be unfair to collect personal information covertly’. And Australia has no general ‘legitimate interests’ ground to fall back on.
The UK and the EU reach the same place by a different road. Personal data covers opinions and assessments about a person, not only facts (the EU Court of Justice in Nowak). In the UK, a person whose data you collect from other sources is normally owed a notice by the first time you communicate with them, and the exception for ‘disproportionate effort’ is read narrowly. For one named, contactable general counsel, it’s hard to argue the effort is disproportionate.
The bigger problem is usually confidentiality
Firms tend to build a model of a client’s GC from what they hold: emails, call notes, CRM records, past proposals, the fee dispute from last year. That material is usually the client’s confidential information.
Confidential information can’t go to a third party without the client’s authority. Solicitors’ conduct rules in Australia and the UK say so, and so does the accounting profession’s code. An AI vendor is a third party. Using that information to gain an advantage over the client is a separate problem again. For a corporate client, the person whose authority you’d need is often the GC: the same person you’re modelling.
What the law doesn’t stop: modelling a role
None of this stops a firm testing a pitch on the role a general counsel plays. A model of ‘the GC of a large infrastructure operator, under pressure to cut external spend’ is not a model of a named person, so long as nothing in it points to one.
Modelling the role, not the person, is the line The Gauntlet is built on. The Gauntlet is Taleist’s service for testing a document before the client sees it:
- Roles, never named people. It never models a named person, or a role at a named organisation.
- Nothing that belongs to one organisation. A detail that would only fit one client is rewritten to describe the sector before it goes in.
- No client records. No emails, call notes, CRM records or fee disputes. The firm anonymises its document before sending it.
- Text only. No voice, no images, no video of anyone.
What you give up is the illusion that you’re hearing from the actual person. What you keep is the part that was useful: the reasons someone in that role could say no, in the order the roles meet the document, while you can still change it.
Can I build a digital twin of my client?
‘Digital twin’ usually means a model of a customer built from your data about them. Customer-data vendors sell them, and for anonymous, aggregated customers the privacy picture is different.
For one named client contact, the same rules apply as above. The twin is personal information about them, it’s built from records that are usually confidential, and they haven’t been told. If you’d be uncomfortable telling the GC you’d built one of them, that discomfort is the answer.
The tool that fits the job, testing what you’re about to send, works from roles. See How The Gauntlet works.
We are not lawyers, and even if we were, we are not your lawyers. You should get advice on this from your own lawyer before doing or not doing anything.
Sources: Privacy Act 1988 (Cth) s 6(1); OAIC, Guidance on privacy and the use of commercially available AI products; OAIC, APP Guidelines chapter 3 (APP 3), paras 3.7 and 3.87; OAIC, APP Guidelines chapter 5; Nowak v Data Protection Commissioner (C-434/16), para 34; UK GDPR art 14; California Civil Code 1798.140 and 1798.145. Read 25 October 2026.