[ Legal / AI Provenance ]
AI, Authorship and Provenance
Last updated: June 2026
Effective communication comes first — how we think about tools, authorship, provenance, and human responsibility.
Effective communication comes first
We recognise that contemporary writing and communication are increasingly produced through collaboration between people, software and AI systems. This may include spelling and grammar tools, translation systems, generative AI, large language model (LLM) reasoners, research tools, editorial systems and other forms of computational assistance. We do not consider the mere use of such tools to determine authorship, ownership, responsibility or the value of a communication. A tool is a means of production. Its use does not, by itself, establish that the resulting work was authored by the tool, nor does it diminish the contribution of the person who directed, evaluated, edited or accepted the result.
Authorship is not a binary property
We distinguish between authorship, contribution and provenance. A person may originate an idea, supply source material, direct an analysis, challenge a conclusion, make editorial decisions, verify factual claims or accept responsibility for a final communication while using computational tools extensively in its production. Conversely, a person may make only nominal changes to material substantially produced by an AI system. For this reason, statements such as 'AI-generated' or 'human-generated' can be insufficient to describe how a significant communication was actually produced. Where provenance is relevant, we prefer to describe the transformation history of an artefact: its sources, material contributions, tools used, human interventions, editorial decisions and verification.
Provenance does not mean tool surveillance
We do not regard the detection of AI-like characteristics in a text as proof of AI authorship. Statistical detection can provide evidence that a text resembles material produced by a particular class of systems. It cannot, by itself, establish who produced the text, quantify the contribution of a particular tool, or determine whether the use of that tool was legitimate. Accordingly, AI-detection results should be treated as signals requiring appropriate context rather than as definitive evidence of authorship or misconduct. The same principle applies in reverse: the absence of an AI-detection signal does not establish that a text was produced entirely by a human.
Responsibility remains human where responsibility is human
Where a person publishes, submits, approves or otherwise presents a communication as their work, that person remains responsible for the claims and representations they make, subject to the applicable contractual, professional and legal framework. Using an AI or LLM system does not transfer responsibility to the system. Human review should therefore be proportionate to the significance and potential consequences of the communication. Particular care should be taken with factual claims, legal or regulatory assertions, confidential information, personal data and other material where errors may cause harm.
Communication is the objective
The purpose of writing is not to demonstrate that a particular tool was absent from the production process. The purpose is to communicate something accurately, appropriately and effectively to its intended audience. We therefore favour assessment and governance based on the qualities that matter to the communication itself: accuracy, meaning, clarity, context, provenance where relevant, accountability and fitness for purpose. Where a particular context requires unaided human performance — for example, an examination intended specifically to establish an individual's independent capability — that requirement should be stated explicitly and assessed directly. Where tool-assisted performance reflects the real-world capability being assessed or exercised, appropriate use of those tools should not automatically be treated as compromising authorship.
A practical principle
A practical principle:
The fact that a tool participated in producing an artefact is less important than understanding what the tool contributed, what the human contributed, what has been verified, and whether the resulting communication achieves its intended purpose.
We favour provenance over presumption, context over binary classification, and effective communication over artificial distinctions between human and tool contribution.