If We Keep Saying “Use AI Responsibly,” We Should Explain What That Means

I have used the phrase “responsible use of AI” myself. The more I hear it, the more I think we have turned it into a safe sentence. It sounds sensible. Nobody really argues with it. Then the meeting ends, the policy is published, and students are still left asking the same basic question: what am I actually allowed to do?

That is the part I think academia needs to get better at.

I understand the reluctance around generative AI. The concern is not irrational. These systems can produce wrong answers with confidence, invent references, flatten nuance, and make weak work look polished. They can also let a student submit something they did not really think through. Early academic discussions around ChatGPT raised exactly these issues, alongside the possibilities for feedback, accessibility and learning support [1][2].

But pretending AI is optional to the wider world is becoming harder to defend.

Student use started moving quickly almost as soon as these tools became widely available. In a 2024 University of Liverpool study involving 2,555 students, 50.9% said they had used or considered using generative AI technologies for academic purposes, while 35% said they had never used or considered them. Students were much more comfortable with support such as grammar help than with an AI writing an entire essay, and the most popular policy preference was clear university-wide guidance about when AI was and was not appropriate [3]. A separate U.S. study of 1,001 college students also found substantial use, but a lot of uncertainty around institutional rules [4].

Those studies are not a measure of AI use in 2026. They are useful for a different reason: they show how early the shift had already started.

The evidence base has grown since then. A 2026 systematic review of 88 empirical studies reported educational benefits in areas such as performance, engagement and accessibility, but also recurring concerns about over-reliance, fairness, privacy and technical reliability [5]. That mix matters. AI is not a magic solution, and it is not simply a cheating machine either.

“Responsible use” needs a second sentence

A lot of universities now use the language of responsible AI. The principle is right, but the phrase can become vague very quickly.

Responsible for what?

Responsible to whom?

And what does a student actually do differently on Tuesday night when they are working on an assessment?

Research on university policies shows that institutions have generally moved towards an open but cautious position, with recurring concerns around ethics, accuracy and privacy [6]. In Australia, that shift is becoming more concrete. The University of Sydney announced a two-lane assessment model: secure assessments where students demonstrate learning without outside assistance, and open assessments where contemporary tools, including AI, can be used appropriately [7]. TEQSA’s 2026 guidance similarly focuses on assuring learning while developing capabilities such as evaluative judgement, critical thinking and ethical reasoning in an AI-integrated environment [8].

A UNSW piece published in August 2026 described a move from asking only “Can students use AI?” towards asking what kind of learning an assessment is trying to assure. Its model distinguishes between assessments where students can use AI, cannot use AI, and must use AI because AI capability is itself part of the learning outcome [9].

I think that is a much more useful conversation.

Not every assessment should allow AI. Not every assessment should ban it either. The decision should come from the purpose of the task.

If I need to know whether a student can independently explain a concept, write code, perform a calculation, interpret a result or make a professional judgement, there are good reasons to create moments where that ability is demonstrated without AI.

If the task is meant to resemble professional practice, then refusing students access to tools that are already part of that practice may make the assessment less authentic.

And if AI literacy itself is one of the learning outcomes, then students may need to use AI so that we can assess whether they can question it, verify it and know when not to trust it.

That is more complicated than a ban. It is also more educationally honest.

What I would call responsible AI use

I do not think we need a fifty-page policy for a student to make a sensible decision. I would start with a few plain questions.

Is AI allowed for this task, and why?

The student should not have to guess. Assessment instructions should say what is permitted, what is not, and what the assessment is trying to measure.

Is AI helping me think, or quietly doing the thinking for me?

There is a difference between asking for another explanation of a difficult idea and asking a model to produce the argument that you are meant to develop. There is also a difference between getting feedback on code and submitting code you cannot explain.

Can I defend what I am submitting?

This is probably my simplest test. If a student cannot explain a claim, a method, a line of code, a reference or a decision in their own work, then the problem is bigger than whether AI was used.

Have I checked the things that matter?

A fluent answer is not evidence. References should exist. Calculations should be checked. Code should be tested. Claims should go back to credible sources. AI output can be useful, but usefulness and correctness are not the same thing.

Have I been honest about material AI assistance?

Disclosure does not need to become a ritual. If AI materially shaped the work, say how. Monash, for example, asks students to document the tool used and how it contributed when AI use is allowed [10]. A short statement can often do the job: “I used generative AI to test alternative structures for this section and to comment on clarity. I checked the sources and wrote the final argument myself.”

Am I putting anything into the system that I should not?

Responsible use also means thinking about privacy, confidential material, unpublished research, personal information and intellectual property. “I did not plagiarise” is not a complete AI ethics policy.

There is one more question that matters: Would I be better off not using AI here?

AI literacy should include the confidence to say no. A student should be able to decide that a task is better done slowly, independently, or through a conversation with a person. Understanding a tool does not mean using it every time.

We should be careful about what we punish

I also think the current conversation sometimes becomes too focused on catching AI use.

That worries me.

If the centre of our strategy is detection, students learn that the main problem is being caught. That is not the same thing as learning integrity.

Research testing AI-text detectors has found reliability and error problems, including false positives and false negatives [11]. TEQSA-hosted guidance on assessment design makes a related point: criteria should focus on the learning the task is meant to demonstrate, and suspicion should not be treated as evidence [12].

I would rather ask a student to explain how they arrived at a decision than treat a percentage from a detector as a verdict.

That does not mean misconduct should be ignored. If a student submits work they did not produce, cannot explain and was not permitted to generate with AI, that is an integrity issue. But the response should still be based on evidence and academic judgement, not technological anxiety.

There is also an equity issue here. The Liverpool study found that students themselves raised the value of AI for disabled students, international students and students who lacked confidence in academic writing [3]. Baek and colleagues found that non-native English speakers in their U.S. sample were more likely to use ChatGPT for writing [4]. A blanket policy can therefore land differently on different students.

That does not mean “anything goes”. It means policy needs enough nuance to distinguish support from substitution.

AI is also forcing us to look at assessment itself

This is the uncomfortable part.

Sometimes AI is exposing a weakness that was already there.

If an assessment can be completed convincingly by pasting the task into a chatbot and making a few cosmetic edits, I am not saying the assessment is automatically bad. But I do think it is worth asking what learning the task is actually capturing.

Can we ask students to justify decisions rather than only present answers?

Can we see part of the process, not just the final product?

Can students critique a generated response, compare alternatives, defend a method, work with local data, explain where a model failed, or connect an answer to something they actually did?

Can we build a program where students sometimes demonstrate independent capability, sometimes work with AI as they may in professional life, and sometimes have to show that they understand AI itself?

That is harder than adding “AI use is prohibited” to the end of an assessment brief. It also has a better chance of surviving the next generation of tools.

In 2024, an analysis of policies and resources across leading U.S. universities found that many institutions were already providing workshops, templates, teaching resources and local guidance rather than relying on prohibition alone [6]. The direction in Australia is now even clearer. TEQSA is explicitly framing assessment reform around the capabilities graduates need in a world where generative AI is embedded [8].

The question is slowly changing from “How do we stop AI entering assessment?” to “How do we still know what the student can do?”

I think that is the right question.

The human part becomes more important, not less

There is a strange assumption in some AI conversations that accepting the technology means lowering the value of human work.

I see it differently.

The more capable AI becomes, the more important judgement becomes.

Someone still has to decide whether the answer makes sense. Someone has to recognise when the model has misunderstood the context. Someone has to decide which evidence is credible, which trade-off is acceptable, and who is accountable if the decision causes harm.

In research, a recent university framework built around responsible generative AI use centres on honesty, transparency and accountability [13]. Those words are useful because they put responsibility back on people and institutions rather than pretending the tool itself can carry it.

The same principle applies to students.

I do not want graduates who can only work when an AI system is beside them.

I also do not want graduates who were kept away from AI throughout university and are then expected to use it intelligently the first week they enter a workplace.

Both outcomes would worry me.

The balance is not between being “pro-AI” and “anti-AI”. That framing is already too small.

Universities should protect moments of independent thinking. We should protect the value of reading slowly, struggling with a problem, writing a first draft, making mistakes and learning from feedback. There are parts of education that should not be made frictionless.

At the same time, we should teach students how to work with tools that are becoming part of research, software development, business, healthcare, engineering, education and many other professions.

That means teaching verification, disclosure, privacy, bias, limitations and judgement alongside prompting.

It also means academics need support. “Students must use AI responsibly” is not a serious institutional strategy if staff are not given time, tools and guidance to redesign assessment and teach those practices.

Where I land

I am not interested in defending every use of AI in academia.

Some uses are lazy. Some are dishonest. Some weaken learning. Some create privacy or intellectual-property risks. Some simply add AI to a task that was better without it.

But I am equally unconvinced by the idea that keeping AI outside the classroom protects education.

AI is already part of the present. The tools will change, and today's platforms may look primitive quite soon. The durable skill is not knowing how to use one chatbot.

It is knowing how to remain the thinking, questioning and accountable person while using increasingly capable tools.

So when we tell students to “use AI responsibly”, I think we owe them the next sentence.

We should tell them what that means in this task.

We should tell them why.

And then we should design the learning so that responsibility is something they actually practise, not just a phrase in the policy.