Good AI practice is not just the absence of bad practice.

A lot of institutional discussion about generative AI has become very good at describing the red lines.

Do not upload confidential material. Do not paste student work into a public chatbot. Do not ask AI to mark an essay and pretend the feedback is yours. Do not cite things you have not checked. Do not claim authorship over work you did not substantially shape. Do not treat a fluent answer as evidence.

All of that matters. The red lines are real. Universities need them, especially in areas like assessment, unpublished research, personal data and student intellectual property.

But there is a problem with a responsible AI conversation that is built almost entirely around what not to do. It can leave people with a strangely thin account of what good practice looks like. Bad practice becomes visible, but good practice is defined only by its absence. The responsible user is the person who has not broken the rules.

That is not enough.

If generative AI is now part of ordinary academic, administrative and professional life, then institutions need a richer language for competent use. People do not only need to know where the hazards are. They need to know what skilled, ethical, useful work with AI actually looks like in practice.

This is the shift I think matters: from avoiding bad AI use to cultivating good AI practice.

Good practice starts with purpose.

A weak use of AI often begins with the tool: here is a chatbot, what can I make it do? A stronger use begins with the work: what am I trying to understand, decide, communicate or improve? The same system can be used to generate a superficial shortcut or to support serious thinking. The difference is not only technical. It is intentional.

For an academic, good use might mean asking AI to summarise a dense document before reading it closely, to generate alternative structures for a lecture, to test whether a grant argument is clear, to turn rough notes into a meeting brief, to identify inconsistencies across versions of a paper or to play the role of a sceptical reviewer. These are not all the same activity. They carry different risks, require different levels of checking and sit in different relationships to authorship and accountability.

That is why purpose has to come first. A responsible user should be able to answer a simple question: what role is the AI playing here?

Is it a search assistant? A drafter? A critic? A translator? A summariser? A formatting helper? A second reader? A coach? A source of examples? A way of reducing administrative friction so that human attention can go back to the substantive work?

Naming the role matters because it changes the standard of scrutiny. If AI is helping brainstorm titles, the risk is low. If it is summarising a complex document, the output needs checking. If it is touching confidential, sensitive or student-owned material, the question may not be how to check the output but whether the material should be there at all.

Good practice is also human-led.

This phrase can sound bland, but it is doing important work. Human-led does not mean the human types every sentence unaided. It means the human remains responsible for the question, the context, the judgment and the final use of the output.

That responsibility is not a ceremonial flourish at the end. It has to shape the whole interaction. The user decides what the task is, what evidence matters, what constraints apply, what counts as a good answer and where the output is not good enough. The AI can help with drafting, organising, challenging and refining. It should not quietly become the agent deciding what the work is.

This is especially important in research. AI can make a paragraph sound more fundable, more polished and more aligned with a call. That can be helpful. It can also pull ideas towards the centre of what has already been written. Good practice means using AI to clarify scientific thinking, not to outsource it. The researcher still has to decide what is worth knowing, what the mechanism is, why the design is appropriate and where the contribution sits.

Good practice is transparent in a meaningful way.

AI declarations need nuance. If someone uses AI to tidy a calendar invite, generate a checklist or suggest three possible headings, that is not the same as using it to draft a published argument, analyse research material or shape assessment feedback. Treating all uses as equivalent makes transparency feel bureaucratic and unhelpful.

A better standard is materiality. Did AI substantially shape the analysis, argument, wording, interpretation, decision or feedback? Would a reader, student, collaborator, funder or colleague reasonably expect to know that AI was involved? If yes, transparency is part of good practice. If no, the useful behaviour may simply be careful checking and appropriate data handling.

This is where universities need more examples, not only more warnings. Staff and students need to see what a sensible acknowledgement looks like. They need examples of acceptable use in planning, teaching preparation, coding, proofreading, translation, grant development, literature mapping and administrative drafting. They also need examples of unacceptable use, but the positive examples are what help people build judgment.

Good practice is context-sensitive.

The same action can be harmless in one context and inappropriate in another. Summarising your own rough notes is different from summarising identifiable interview transcripts. Asking for feedback on a paragraph you wrote is different from asking for comments on a student submission. Using AI to generate practice questions is different from using it to produce answers for a summative assessment. Asking AI to critique your own draft is different from uploading an unpublished manuscript shared in confidence by a collaborator.

This is why generic rules only get us so far. Responsible use has to be interpreted within real work settings: teaching, assessment, supervision, research, management, public engagement, clinical work, commercial work and administration. People need the skill of asking: what kind of material is this, whose rights are involved, what promises of confidentiality apply and what harm could follow if this output is wrong or this input is mishandled?

Good practice is critically checked.

This is the most obvious point, but it is often the least practised. AI outputs can be fluent, useful and wrong. They can flatten uncertainty, invent references, miss relevant literature, overstate evidence or produce advice that sounds plausible because it resembles advice seen elsewhere.

Checking is not a separate chore after the real work. It is part of the method. Good AI use builds in verification from the start: ask for sources, compare against the original document, test the argument against counterexamples, check key facts independently, keep track of what has been inferred and make sure the final wording still matches the evidence.

In academic work, this matters because the polished version can become less true than the messy one. A draft can become clearer while quietly losing a caveat. A manuscript can become more elegant while overstating causal inference. A grant can become more persuasive while drifting away from the actual design. Good practice is not simply producing better-looking text. It is preserving truthfulness while making the work clearer.

Good practice is also equitable and educative.

This part is easy to miss if the conversation stays at the level of compliance. AI literacy is becoming a form of professional capital. People who already know how to brief, question, challenge and evaluate an assistant will get more value from these systems than people who only know that some uses are forbidden.

If universities only teach avoidance, they risk widening that gap. The confident users will keep learning through experimentation. The cautious users will hang back, afraid of doing something wrong. The institution will have technically reduced risk while failing to build capability.

A better approach is to make good practice teachable. Show people how to use AI as a thinking partner rather than a shortcut. Show them how to ask for critique rather than deference. Show them how to preserve authorship, check evidence, protect data and use AI to make work more accessible without making it less rigorous.

This is where I think the frame needs to change.

Responsible AI should not be presented only as a fence around bad behaviour. It should be a craft: a set of habits, judgments and routines that help people use powerful tools well.

For a university, that means moving from prohibition to practice. Not abandoning the red lines, but embedding them in a richer account of good work. What does good AI-supported teaching preparation look like? What does good AI-supported supervision look like? What does good AI-supported grant development look like? What does good AI-supported admin look like? What does good AI-supported writing look like when the human is still thinking, deciding and accountable?

These are much more useful questions than simply asking whether AI is allowed.

They also create space for nuance. Good practice will not look identical across disciplines. A philosopher, a statistician, a clinician, a historian, a public health researcher and a university administrator may all use AI well, but not in the same way. The ethical issues are shared, but the work is situated.

This is why Colleague should not position itself as a tool for getting around the difficult parts of academic work. The better claim is almost the opposite. Colleague should help make good practice easier: clearer boundaries, better checking, more visible provenance, stronger continuity, more thoughtful drafting and more time for the human judgment that cannot be automated away.

The future of responsible AI in universities will not be won by telling people only what not to do. It will depend on helping them recognise, practise and expect better forms of use.

Good AI practice is not the absence of misconduct.

It is purposeful, human-led, transparent where it matters, context-sensitive, critically checked, equitable and educative.

That is a much more demanding standard than simply avoiding the obvious mistakes. It is also a much more useful one.