It is Tuesday night. The assignment is due Friday. You have twelve readings on the unit reading list, three of them are book chapters, and you have maybe three usable hours before the week swallows them. You cannot read everything. You already know that. The question that actually keeps you stuck is quieter: which of these twelve do I skip, and how do I skip them without accidentally missing the one paper the marker wanted me to engage with?

Most of the advice out there answers a different question. The popular roundups of AI reading tools, like the ones from Scholarcy (2024), are written for researchers doing a literature review, where the goal is broad coverage of a field. That is not your problem. Your problem is triage under a deadline, tied to one specific question and one specific rubric. This article is about that, and only that.

I want to be upfront about the line I am walking here. Using AI to help you decide what to read is a study aid. Using AI to read for you, so you can hand in its summary as your own analysis, is not. The gap between those two is the whole article, so I will keep coming back to it.

Start with the question, not the readings

Here is the mistake I made for my first year of study, and the one I see most people make. They open the first PDF, paste it into an AI tool, and ask “summarise this.” The tool spits out a competent, generic summary. It is useless, because the tool had no idea what you were looking for.

AI can only judge relevance if you tell it what “relevant” means for you. That means giving it two things before you give it a single reading: the exact assignment question, and the marking criteria or rubric if you have one.

Think about it from the tool’s side. “Summarise this chapter” has one answer. “Which parts of this chapter would help me argue that community-based aged care improves patient autonomy, given that my rubric rewards critical evaluation of evidence, not description” has a completely different, far more useful answer. The second prompt turns a summariser into a filter.

So before you open anything, write your assignment question at the top of a document in plain language. If the rubric says things like “demonstrates critical analysis” or “synthesises multiple perspectives,” note that too. Australian university skills pages consistently push students toward reading with a purpose rather than reading cover to cover (University of Sydney, n.d.), and this is the same principle, just with a machine doing the first pass.

A step-by-step filtering workflow

Here is the process I would use with twelve readings and three hours. It is deliberately boring, because boring is what works at 9pm on a Tuesday.

Step one: gather the cheap signals. For each reading, you already have the title, the abstract (for journal articles), and often the introduction and the headings. That is usually enough to filter on. You do not need to feed the full PDF of all twelve into anything.

Step two: give the tool your question, then feed it abstracts. Open a tool like Google’s NotebookLM, Claude, or ChatGPT. Paste your assignment question and rubric notes first. Then paste an abstract and ask it to rate how relevant this reading looks to your specific question, on a simple scale, and to explain why. The “explain why” is the important half. A number with no reasoning is noise; the reasoning is what lets you sanity-check it.

NotebookLM is worth a mention here specifically because it grounds its answers in the sources you upload rather than in whatever it happened to learn during training, which reduces the odds of it confidently inventing something (Scholarcy, 2024 covers the broader tool landscape if you want to compare options).

Step three: sort into three piles. Clearly relevant, clearly irrelevant, and unsure. Be ruthless with the clearly irrelevant pile. That is where the time saving lives. If a reading is about a different cohort, a different country’s policy setting, or a methodology your assignment does not touch, it goes.

Step four: read the survivors properly. This is the step people want to skip and cannot. If you have filtered twelve readings down to four, you now read those four. You can ask the tool to point you at the sections most likely to bear on your argument, then you go and read those sections yourself. The AI flags where to look. It does not do the looking.

That last point is not me being precious about effort. It is because the summary is genuinely not good enough to write from, and I will explain why.

What AI genuinely cannot judge

The filter is fast, but it is judging surface features: keyword overlap, topic, how closely the abstract’s language matches your question. There is a whole category of things it cannot see.

Disciplinary weight. A paragraph that looks thin can be load-bearing in your field. A single sentence establishing a theoretical framework might be the exact thing your marker wants you to engage with, and an AI skimming for topical relevance will walk straight past it because it does not know your discipline’s furniture.

Author credibility and source quality. The tool cannot reliably tell you whether an author is a respected voice or whether a journal is reputable. It also cannot tell whether a source meets your lecturer’s requirements. Plenty of units specify “peer-reviewed sources from the last ten years” or exclude certain types of material, and checking that is on you.

Methodological quality. Whether a study’s sample size supports its conclusions, whether the method matches the research question, whether the findings have been contradicted since: this is critical-appraisal work, and it is very often the actual thing being assessed. The Tertiary Education Quality and Standards Agency has flagged that generative AI raises real risks precisely where students outsource the thinking that assessment is meant to measure (TEQSA, 2024).

So the honest framing is this: AI is good at eliminating the obviously irrelevant. It is poor at judging the merit of what survives. Those are different jobs, and only the first one is safe to delegate.

Your academic integrity obligations in Australia

Let me be clear about the policy landscape, because this is where anxiety lives and where the real risk sits.

Every Australian university I am aware of permits AI use for learning and study. The line is drawn at submission: using AI to understand, triage, brainstorm or revise is generally fine, while submitting AI-generated text or AI-generated analysis as your own work is a breach. The University of Sydney’s guidance frames AI as something you can use to support learning but not to produce the assessable work itself (University of Sydney, n.d.), and the ANU’s best-practice guide for students takes a similar line (Australian National University, n.d.).

Using AI to filter which readings to prioritise sits comfortably on the study-aid side of that line. You are still reading the sources. You are still forming your own view. The AI just helped you decide where to point your attention.

Two obligations follow. First, check your specific unit’s rules, because they vary and the unit outline overrides the general university page. UQ, for example, publishes explicit rules on what is and is not permitted in assessment (The University of Queensland, n.d.). Second, declare your use where required. Many universities now expect an acknowledgement of AI tools even when the use was legitimate, and I have written a fuller walkthrough on how to declare AI use in your assignments if you are unsure what that looks like.

Integrity-safe prompting: what to ask, and what not to

The wording of your prompt is where you either stay on the right side of the line or step over it. Here is the difference in practice.

Ask this: “Which sections of this reading are most likely to support an argument about X?” That keeps the AI in a signposting role and keeps the reading and the thinking with you.

Do not ask this: “Summarise this reading so I don’t have to read it.” That prompt is asking the tool to replace your engagement with the source, and if that summary ends up shaping your analysis, you have crossed from studying into outsourcing.

A few more safe patterns worth keeping:

Each of those keeps you as the reader and the analyst. The tool is a research assistant handing you things to look at, not a ghostwriter.

What to do once you have filtered

Filtering is only half the job. Once you are down to your four readings, the goal is to come out the other side with notes in your own words, not a folder of AI paraphrases.

As you read each relevant section, write a couple of sentences in your own language: what this source claims, and how it connects to your argument. That small act of translation is what lets you synthesise multiple sources later instead of just stacking summaries end to end. If your notes are your own words, your essay will be too, and no integrity question ever arises.

This is also where the time reality check bites. Filtering twelve readings to four does not mean you have finished the reading. It means you have four readings to do properly instead of twelve to do badly. The win is the elimination, not a shortcut through the close reading that still has to happen. Australian skills resources make the same point about reading actively rather than passively skimming for a false sense of completion (University of Wollongong, n.d.).

Where this fits with planning your time

The reason I keep banging on about the three-hours-twelve-readings scenario is that it is a planning problem dressed as a reading problem. You get stuck because you do not know how long the real reading will take, so you cannot commit to a plan, so you stall.

That is the exact gap GradeMap is being built to close. I am building it because I needed it as a mature-age student myself: a coach that helps you see how long readings and tasks realistically take, so you can decide what to prioritise before you open a single PDF. The filtering workflow in this article is designed to feed straight into that kind of planning, turning “I don’t know where to start” into a short, honest list of what to read and roughly how long it will cost you.

Filter with the machine. Read with your own eyes. Write from your own notes. Do those three in order and you save real hours without ever going near the line.

References

Australian National University. (n.d.). Guide for students: Best practice when using generative AI. https://www.anu.edu.au/students/academic-skills/referencing-and-academic-integrity/academic-integrity-best-practice/guide

Scholarcy. (2024). The 5 best AI tools for reading research papers. https://www.scholarcy.com/blog/the-5-best-ai-tools-for-reading-research-papers

Tertiary Education Quality and Standards Agency. (2024). The evolving risk to academic integrity posed by generative artificial intelligence. https://www.teqsa.gov.au/sites/default/files/2024-08/evolving-risk-to-academic-integrity-posed-by-generative-artificial-intelligence.pdf

The University of Queensland. (n.d.). Rules for using AI in assessment. https://itali.uq.edu.au/teaching-guidance/generative-ai-teaching-learning-and-assessment/rules-using-ai-assessment

University of Sydney. (n.d.). Artificial intelligence. https://www.sydney.edu.au/students/academic-integrity/artificial-intelligence.html

University of Sydney. (n.d.). Reading and note taking. https://www.sydney.edu.au/students/study-skills/reading.html

University of Wollongong. (n.d.). Effective reading. https://www.uow.edu.au/student/support-services/academic-skills/online-resources/effective-studying/effective-reading/

FAQ

Is it cheating to use AI to decide which readings to skip?

No, on the condition that you still read the ones you keep and form your own analysis. Deciding what to prioritise is a study decision, the same as choosing which chapters to focus on before an exam. It becomes a problem only if the AI’s output starts substituting for your own reading and thinking. Every Australian university I am aware of permits AI as a study aid while prohibiting the submission of AI-generated work as your own. Check your unit outline for the specific rule.

Do I have to declare that I used AI to filter my readings?

It depends on your university and unit. Some institutions ask you to acknowledge any use of AI tools during your work, even when that use was legitimate and did not touch the writing. Others only require declaration when AI contributed to the submitted content. When in doubt, declare it briefly, because a transparent acknowledgement never counts against you, whereas an undisclosed use can. See my separate guide on declaring AI use for templates.

Can I just trust the AI’s relevance rating?

Treat it as a first pass, not a verdict. AI is reliable at spotting obviously off-topic material, which is where most of your time saving comes from. It is unreliable at judging source quality, author credibility, methodological soundness and disciplinary weight, which are often the exact things your marker is assessing. Use the rating to bin the clearly irrelevant, then apply your own judgement to everything that survives.

Which AI tool is best for this?

Any capable general assistant works, but source-grounded tools like NotebookLM have an edge for this task because they answer from the documents you upload rather than from general training, which lowers the risk of the tool inventing content. The tool matters less than the prompt. Feeding it your assignment question and rubric first is what turns any of these tools from a generic summariser into a genuine filter.

How much time does this actually save?

Realistically, it saves you the reading you would have wasted on irrelevant sources, not the reading you still need to do. If you filter twelve readings down to four, you have removed eight sources worth of wasted effort, but you still owe those four a proper read. The honest gain is direction and confidence: you start reading the right things sooner instead of stalling over where to begin.