TEQSA and AI assessment
What Australia’s regulator now expects from AI-era assessment.
The short answer
TEQSA has not banned AI and has not told providers to buy anything. Since 2023 it has said that reliably detecting AI use is nearly impossible, and it has asked providers to redesign assessment so that it gathers evidence of learning, including evidence of the process of learning. In 2024 every registered higher education provider filed an AI action plan. TEQSA calls its material guidance, not new rules. This is guidance, not legal advice.
A guide from Edudojo, an AI that won’t think for students: its coach asks questions instead. The advice works with or without it.
On this page
Disclosure: I founded Edudojo, which makes an AI coach for students. Edudojo has no customers in Australia and is not endorsed by TEQSA. This page reports public documents and other people’s research.
One week in a first-year unit
Dr Menon is an illustrative example, not a real person.
Dr Menon coordinates a first-year unit of 400 students. On Monday a detector flags an essay from a student who moved from Pune in February. The score is 82% “likely AI”. On Tuesday the student writes back, upset: she wrote it herself, and English is her third language.
Dr Menon has a choice, and so does her university.
| Aspect | Trust the detector score | Go back to invigilated exams | Look at the process, and add secure moments |
|---|---|---|---|
| What the teacher sees | A number | One sitting | Drafts, questions asked, and a short conversation |
| Who it can hurt | Honest students who write plainly, or in a second language | Students who do their best thinking over days | Needs more teacher time at first |
| TEQSA’s view | Detecting AI use with certainty is “all but impossible” | Warns against defaulting to invigilated, time-limited exams | Matches the process and security propositions |
| Cost to learning | Trust | Narrower tasks | Setup work |
On Wednesday Dr Menon invites the student to explain her argument for five minutes. It is obvious that she understands it. The case closes without an accusation.
The case closes without an accusation.
What TEQSA has said, in order
| When | Document | What it says |
|---|---|---|
| Sep 2023 | Discussion paper | Reliable detection “appears almost impossible”. It proposes “evidencing the process of learning over time and in context” |
| 23 Nov 2023 | Final principles launched | The principles are “not intended to be prescriptive”. TEQSA says it intends to seek assurance that providers have plans |
| 5 Mar to 3 Jul 2024 | Request for information | All registered providers asked for a “credible institutional action plan, oversighted by appropriate governance mechanisms” |
| 28 Nov 2024 | Emerging practice toolkit | Built from the plans. TEQSA reports a 100% response rate |
| 24 Sep 2025 | Enacting assessment reform | Five propositions and three pathways. Detection with certainty is “all but impossible”. Redesign assessment rather than “investing primarily in detection mechanisms” |
| Mid 2026 | Assuring quality learning in a gen AI-integrated future | Third in the series, on evaluative judgement and ethical reasoning. The web page shows no date |
The five propositions, in plain words
These are our paraphrases of the propositions in the 2025 resource. Read TEQSA’s own text for the exact wording.
| Proposition (paraphrased) | What it could look like in a unit |
|---|---|
| Engage with gen AI in appropriate, authentic ways | Tasks say what AI use is allowed, and why |
| Plan assessment across the whole program, not unit by unit | Each year builds on the last |
| Pay attention to the process of learning | Drafts, reflections or questions are part of the evidence |
| Let students work with peers and AI appropriately | Some tasks are open, with rules |
| Add security at meaningful points | A supervised oral or in-class task at key moments |
“Secure” in TEQSA’s resource means tasks with mechanisms, generally simple to implement, that verify the person, such as an oral presentation, an in-class task or a supervised practical demonstration. The resource also warns that securing every unit “will not, in and of itself, assure program learning outcomes if a degree program is poorly aligned”.
Why a detector score is a weak foundation
A detector works like a smoke alarm that sometimes rings for toast and sometimes sleeps through a fire. Two studies show how.
| Study | What they tested | What they found |
|---|---|---|
| Weber-Wulff et al., 2023 (International Journal for Educational Integrity) | 14 detection tools, including Turnitin | The tools were “neither accurate nor reliable”. They leaned towards calling text human-written, and disguising the text made them worse |
| Liang et al., 2023 (preprint; a published version is linked from arXiv) | Seven detectors on 91 essays by non-native English writers and 88 essays by US eighth-graders | On average 61.22% of the non-native essays were called AI-generated, and at least one detector flagged 89 of the 91. The eighth-grade essays were classified almost perfectly |
Australia hosted 955,317 international student enrolments to August 2025, and 17% were from India. A tool that misfires on second-language writing misfires on a large group of your students. More on this in how accurate AI detectors are and what to do instead of an AI detector.
A checklist for a unit coordinator
- Say what AI is allowed to do, task by task. Use the AI-use policy template.
- Ask for evidence of the process. Drafts, notes, or the questions a student asked themselves. See process-based assessment.
- Keep one conversation in reserve. A five-minute oral about the argument settles most doubts, and it is the kind of secure task TEQSA describes.
- Never treat a score as proof. Treat it as a reason to talk.
- Choose coaching tools over answering tools. In a field experiment with nearly a thousand high school maths students, a plain GPT-4 interface raised practice grades by 48%, but students then did 17% worse on an exam without AI than students who never had it. A tutor version with learning safeguards largely removed that drop (Bastani et al., PNAS 2025; setting: Knowledge at Wharton). It was school maths in Turkey, not an Australian university.
- Let students think in their home language. See assess the reasoning, not the language.
- Ask where student data is stored. Before any tool touches student work, ask where data is stored, who can read it, and whether it trains models.
- Write down what you decided and why. Your governance body will ask, as TEQSA’s 2024 request did.
Questions coordinators ask
Does TEQSA require AI detection software? We found nothing that requires it. The 2025 resource says detecting AI use with certainty is “all but impossible” and advises redesigning assessment rather than investing primarily in detection.
Does TEQSA approve or certify tools? We found no TEQSA scheme that approves or certifies tools. Its 2025 resource names no vendor. Treat any “TEQSA-approved” claim with suspicion.
Can a student’s chat with an AI tutor count as a secure assessment? Not by itself. In TEQSA’s resource a secure task has mechanisms that verify the person. Process evidence from home work is useful, but it is a different kind of evidence.
Do the privacy laws differ by university? Yes. One law firm’s 2024 summary says the federal Privacy Act covers private universities and the Australian National University, and that other universities follow state and territory laws (MinterEllison, November 2024; check the current position with your privacy officer). The OAIC says an entity that discloses personal information to an overseas recipient must take reasonable steps and stays accountable for it (APP 8 guidelines).
Will TEQSA enforce AI rules in 2026? We found no TEQSA document announcing AI-specific enforcement. TEQSA calls its resources guidance, and the 2026 changes we found were in other areas, such as governance and offshore delivery (MinterEllison).
What should a tool show a teacher? How the thinking was built, so the teacher can make the judgement. A tool should say plainly what it cannot prove.
Where Edudojo fits
Edudojo is an AI coach that asks questions instead of writing for students. It is built for colleges and training providers. It is in beta, and its first pilot, with Parishkar College in Jaipur, is planned (pilot plan). It has no customers in Australia.
How Edudojo answers the checklist today. This is guidance, not legal advice.
| Question | Edudojo today |
|---|---|
| Where is student data stored? | In Mumbai, India. AI model providers process text where they run (FAQ). For an Australian provider we would create a separate instance in Australia; none exists yet |
| Does student work train AI models? | No, and it is not used for research |
| Who can read student work? | The assigned teacher. Admins see records without text. Staff reads are logged (privacy) |
| Is it approved by TEQSA or certified as compliant? | No. Our privacy page says it “does not establish compliance or certification” |
| Does it grade or detect AI? | No. The teacher judges the work, and the record does not prove who wrote it |
Its effect on learning has not been tested. See our methodology. If you coordinate a unit and want to try this with a college, see the pilot page.
Sources (retrieved 2026-10-21)
- TEQSA, Assessment reform for the age of artificial intelligence: discussion paper (September 2023). Regulator guidance.
- TEQSA, Assessment reform for the age of artificial intelligence (launched 23 November 2023). Regulator news.
- TEQSA, Request for information: addressing the risk of artificial intelligence (2024). Regulator page.
- TEQSA, Gen AI strategies for Australian higher education: Emerging practice (28 November 2024). Regulator toolkit.
- TEQSA, Enacting assessment reform in a time of artificial intelligence (24 September 2025). Regulator guidance (PDF).
- TEQSA, Assuring quality learning in a gen AI-integrated future: the role of adaptive capabilities (2026; the web page is undated). Regulator guidance.
- Department of Education, International student monthly summary (year to date August 2025). Government statistics.
- MinterEllison, Privacy reforms in Australia’s higher education sector (4 November 2024) and A year of reform: regulatory changes relating to TEQSA in 2026. Law firm summaries.
- OAIC, APP Guidelines Chapter 8: Cross-border disclosure of personal information (October 2025). Regulator guidance.
- Weber-Wulff et al., Testing of detection tools for AI-generated text, International Journal for Educational Integrity 19(1), article 26 (2023). Study.
- Liang et al., GPT detectors are biased against non-native English writers (2023). Preprint; a published version is linked from the arXiv page. Study.
- Bastani et al., Generative AI without guardrails can harm learning: Evidence from high school mathematics, PNAS 122(26), e2422633122 (2025). Study.
- Knowledge at Wharton, Without guardrails, generative AI can harm education. Setting of the Bastani study (Turkey, nearly 1,000 students). Press article.
What we do not claim.
- Not legal advice
- This page reports public documents and other people’s research. It is guidance, not legal advice.
Try it this week
Pick one task in your unit and write down what AI use is allowed for it. Keep a five-minute conversation about the argument in reserve for any case where you doubt the work.
Trying this with a college? See the pilot plan