Positive Persuasive LLMs

Dr Siying Hu at City University of Hong Kong talked on "Persuasive Simulations in the LLM Era" at the ANU AI, ML and Friends seminar. They suggested AI could be used in a positive way to counter the negative effects of social media and scammers. The Australian government is planning more support for the long term unemployed. Perhaps this technology could help build their confidence, as well as improving job seeking skills.  

Custom Questions for Each Student Based on Their Submitted Work?

Generative AI has the ability to create questions for students from the syllabus. This is very easy to do: you point the AI at the course documentation and tell it how many and what sort of questions you want. If multiple choice it also provides the answers. Could this also be done for each student, with custom questions based on their individual project work? 

Much has been written about oral examinations as the solution to students cheating. This is traditionally done for project based students, such as those undertaking a PhD. I had to go through a similar process at the end of my MEd, asked questions by a panel, based on what I submitted. However, this would be infeasible expensive for a program with hundreds, or thousands, of students. One option would be to have students question each other (this as been used in the ANU Techlaucher program). Also group projects reduce the workload, while introducing their own problems. 

Recently when attending training for Instructure's Canvas Learning Management System, being introduced at the Australian National University, I asked if each student could be given a separate file during an online test. The idea would be to provide them with a copy of their assignment, then ask them questions specifically about it. At the time I assumed the same set questions would be asked of every student. But it occurs to me now, why not ask them each a custom set of questions based on their work, and not provide a copy of the work to them?

Chat GPT Produces a Disrespectful Cartoon by Accident

I asked Chat GPT to:

Chat GPT generated cartoon,
from a prompt by Tom Worthington,
CC-BY 23 September 2025

'Draw a cartoon of the Australian prime minister as a school child, pointing to palestine on a map of the world, with the foreign minister, Penny Wong as the teacher, saying "Well done Albo!". The caption should be "Australia Recognises Palestine".' 

Curiously, Chat GPT said "I’ll avoid anything offensive or disrespectful", but got the location of Palestine wrong, giving the cartoon more of a sting. This is an aspect of Generative AI we needed to keep in mind and teach students.

Quantum AI for Defence at ANU

Dr. Nils Herrmann, 
 Quantum Brilliance

Greetings from the Australian National University where Dr. Nils Herrmann, from Quantum Brilliance is talking on Quantum Machine Learning on Diamond-based Quantum Computers. Quantum Brilliance is an Australia/German startup building quantum computers using diamond microchips. Their Canberra research center ios at the ANU. What got my attention, and probably did of governments and funders, is using the quantum effect for machine learning. Nils envisions a large quantum computer using training data and then a quantum computer on a plug in board in a desktop computer using the model created. The aim is to have a board rugged enough to be used in an aircraft, with obvious military applications. 

"This talk introduces the concept of Quantum Utility - Quantum Brilliance’s vision for practical quantum advantage - and presents early gate-based QML results achieved on an on-site, room-temperature quantum computer. It then highlights the work of the newly formed Quantum Utility and Exploration (QUTE) team, whose mission is to uncover utility through control modes that operate closer to the physical diamond-based hardware. The talk concludes with an introduction to quantum reservoir computing, showcasing QUTE’s latest simulations on future single- and multi-NV systems that establish a compelling platform for near-term, practical quantum machine learning."

AI Has Some Way to Go Generating Images

Image by Open AI,
19 December 2024

AI generated images have started appearing to illustrate social media posts by individuals, corporations and media companies. These can be a bit odd. An example is an AI generated graphic for Australian Army Innovation Day. This shows a soldier in the foreground apparently looking at a handheld screen, but wearing a completely opaque VR headset over their eyes. Despite all their high tech equipment, none of the soldiers have noticed a drone dropping ordnance on them. The soldier in the foreground has high tech gear, but is wearing old camouflage. This was probably produced with a request like 

Generative AI in Health Education

Greetings from the ANU Generative AI in Health Education Symposium in the ANU Moot Court in Canberra. This reports on work to produce and trail guidelines for teaching medical students. This felt like an extension of ASCILITE 2024 in Melbourne last week, as it covered some of the same topics and some of the same people were running the event.

Professor Martyn Kirk, Associate Dean Education, College of Health and Medicine, suggested using generative AI for formative assessment, bit not summative. The idea being the AI could be used for helping students learn, but then they be tested at the end to check they really know what they need to know. I suggest blending the assessments, not having this split between formative and summative. In equity terms not having a large test at the end will allow inclusion of students (such as myself) who can't cope with large exams. It will remove the unnecessary anxiety it causes for many other students. It will also result in more authentic testing, under conditions like a workplace.

The guidelines were trailed in second semester 2024. Students were comfortable using the tool (Microsoft Copilot) and found it helpful for learning. Most students did not receive any training on the tool, but did receive guidance on its use. The approach taken was to issue guidance and leave it to course conveners to tell students about it. Staff and students were provided with assistance in the use of such tools. This is similar to the way in the ANU Techlauncher program we ran a workshop to run students through what Copilot could do, or not do, to help them with an assignment. 

One question which came up was the energy used by Generative AI. Students worry about the effect on the environment. In 2008 I was commissioned to design a course on Green Computing. In this students looked at energy consumption by data centers generally. AI is a more energy intensive form of data center. There are ways to reduce the energy use and carbon emissions resulting. This is something which perhaps should be raised with students generally, not just technical specialists.

One of the courses using generative AI was on climate change and health. In this course tutors demonstrated the use of AI to students in their work. This included using AI to simulate a person for the student to interact with. This was done with face to face and online tutorials. In this case students were given prompts to use with the AI. This extended the student's understanding, where previously they just asked one question and pasted the answer. The approach of simulations could be applied, I suggest, in other disciplines. 

In a course on immunology students could use AI for preparing presentations, but not for writing up their laboratory notebook. How you stop students using AI for the notebook is a challenge. One way would be to require the students to compose directly into an online tool. As with the climate change course, students were given sample prompts to help them. 

The symposium then switched from AI for teaching medical students to teaching students to use AI in medicine. Dr Andrew Tagg, Senior Clinical Lecturer at the University of Melbourne and Emergency Consultant at Western Health, Melbourne, pointed out that AI is already used routinely in diagnosis of cancer from x-rays. He argued that AI should be used more widely to deliver medical care. As someone who recently had to wait 11 hours for medical treatment, this could be a good idea. One area where AI might be useful would be in helping patients filling in the forms they are given. 

The issue of the environmental effects of AI came up again in discussion. It was suggested generating one image took as much energy as charging a mobile phone. This sounded an over estimate to me by three orders of magnitude. The Jevons Paradox came up in discussion: rather than saving effort will AI just result in more resources being used, rather than oit being used sparingly.

ps: Greetings from the ANU College of Health & Medicine located between the Canberra public and private hospitals.  Professor Kirk is giving a workshop for the staff and students at the hospital and those of us at the morning symposium were invited along. One tip is if you are having difficulty with your medical student understanding something ask Generative AI explain it to a ten year old. In groups we were tasked with coming up with something. I was teamed with an anesthetist, so we asked Microsoft Copilot to come up with aspects of a particular condition, then produce multiple choice questions. This worked very well. We were then asked to have AI summarise a paper, so I got it to do one of mine, which worked well. I then asked it to make a ten slide presentation in Powerpoint, then add notes and graphics. 

E-waste not a major challenge with generativeAI

Wang, Chen, Zhang, Tzachor (2024) suggest that the demand for generative AI will create an e-waste problem of 1.2 to 5.0 million tons  for 2020 to 2030. They also suggest this could be made worse by political restriction on access to more advanced  efficient chips and rapid replacement of old hardware. I suggest the problem is not that large and energy use will remain a larger problem than e-waste.

One aspect the authors do not mention is the lack of price signals between server providers and the end user with current generative AI services. This is likely to be self correcting. Currently demand for generative AI is being generated by offering of free services to the public. As the user is not paying for the service and there is therefore no built in fee for responsible disposal of created e-waste, there may be a later problem. Some speculative AI ventures are likely to become bankrupt leaving a toxic legacy (similar to the mountains of scraped e-bikes left by failed startups). However, as users come to rely on Generative AI services, vendors will introduce charges, which can cover e-waste costs.

Currently generative AI server farms are using generic Graphic Processor Unit chips. These are the same chips used for cryptocurrency server farms. The environmental issues are similar with both. There have been articles about AI consuming as much power as small countries (just as there were for crypto). However, there is more of a mainstream use for AI, which will allow for better long term regulation of environmental effects. With its abundance of renewable energy sources and a stable regulatory environment, Australia could provide a popular location for AI centers. This would allow a small query to be sent across the world and answer sent back, effectively embedding the renewable energy in the answer. 

There may also be scope for reuse of older, slower, more energy using AI chips in locations with abundant renewable energy. As more efficient chips were installed close to the user in high energy cost countries, the old hips would be installed further away. Rather than store energy in batteries to run these chips, it may be cheaper to shut them down when the sun isn't shining & the wind isn't blowing. How to do this is something engineers and computer professionals can learn to optimize with specialist training (Worthington, 2012). 

Reference

Wang, P., Zhang, LY., Tzachor, A. et al. E-waste challenges of generative artificial intelligence. Nat Comput Sci (2024). https://doi.org/10.1038/s43588-024-00712-6

Worthington, T. (2012, July). A Green computing professional education course online: Designing and delivering a course in ICT sustainability using Internet and eBooks. In 2012 7th International Conference on Computer Science & Education (ICCSE) (pp. 263-266). IEEE. https://doi.org/10.1109/ICCSE.2012.6295070

Thinking assessment in the large as an answer to AI

Greetings from the CRADLE Seminar "Assessment beyond the individual unit/module and AI". Associate Professor Jason Lodge is talking about how to see how students progress over time, rather than assessing in small packets of instruction. While he did not explicitly say it, I assume he envisions this makes it harder for the student to cheat, using AI, or otherwise, as they would not be showing consistent progress. 

Mentioned by someone was "Assessment reform for the age of artificial intelligence" (TEQSA, 2023). 

Margaret Bearman took us through the logic of current unit based assessment and asks about "big picture" outcomes. My reaction was "Inst that what capstones are for?". You have the student do a big project at the end of their study, where they have to demonstrate the skills needed. 

Surprisingly, there was little mention of AI, which is refreshing. The approach is to get the assessment right and cheating will be harder, however it is done.

I asked the panel:

'Will technology help? Could we give the AI each student's CV and have it suggest what degree requirements they have already met? I help out with applications for course credit and there is a lot of stuff students have done they really don't have to do again. More than once I have thought we should have the student teach the course. ;-)

The next seminar is: "Second Handbook of Academic Integrity (2024) launch".

Grammarly’s Generative AI Writing Assistant Creates Questions for Teachers

Grammarly have introduced a generative AI Writing Assistant as an option for its popular grammar checking program. I have been a fan of grammar correcting tools for decades. Without them, or a human editor, I can't get work published. I recommend them to my students, but Grammarly's AI tool needs to be treated with caution, as casual use could be a career ending.

The tool is available even with the free version of Grammarly. You need to opt in to use it. By default it is I set it to formal, direct, Australian English, by an IT professional. 

Here is the text I gave Grammerly:

'Dr Ryan Young at the ANU National Security College Futures Hub, has prepared "Future Disruptions for Australian Universities" for the Universities Accord Review Panel. The Futures Hub's logo is, appropriately enough a black swan. The accord panel has been advising the Australian Government on the future of higher education...'

Here is how it made the text "sound academic":

'Dr. Ryan Young, from the ANU National Security College Futures Hub, has prepared authored a document titled "Future Disruptions for Australian Universities" for the Universities Accord Review Panel. The Futures Hub's logo emblem, appropriately enough a black swan, seems fitting for their work. The accord panel, has been advising responsible for guiding the Australian Government on the future of higher education's future, ...'

Grammerly has mostly just improved my sentence structure and used more academic sounding words. But the meaning has been changed in a few places for example I wrote "I find older students are easier to teach" to "teaching older students is reported to be more straightforward", thus attributing my comment to the authors of the report. Such mis-attributions could be serious in a student assignment, or a published paper. 

I have become so comfortable with grammar checkers that I accept their recommendations without checking closely. This could be a career ending move with the Generative AI option in Garmmerly. I suggest warning your students, and colleagues, to treat this function with the caution they would a standalone generative AI tool: something to be used to generate ideas, with every detail checked to be right, before being used.

Catastrophic Risks in Artificial Intelligence

Greetings from the Australian National University Colloquium on Artificial Intelligence and Catastrophic Risk. Normally I attend weekly AI, ML and Friends Seminars in the ANU School of Computing, but today I am in the social sciences building, with philosophers. The colloquium is by the ANU Machine Intelligence and Normative Theory Lab. In the first presentation by Professor Seth Lazar points out that "transformer based" generative AI is less brittle: much harder to get it to produce weird results. He also claimed Google was barely able to match Chat GPT's performance. Professor Lazar argued catastrophic and current risks of AI could be addressed together, including through regulation. 

On a positive note Professor Lazar argued AI could be used to efficiently improve human welfare, more than inefficient manual systems. However, I suggest this presents a rosy view of human nature. Resources are not inequitably distributed today due to inefficiency, but because those who have the resources have made a conscious decision to deprive others of them. With an efficient AI system they could implement this deliberate inequity much more effectively. 

Some feasible catastrophic risks Professor Lazar mentioned were discovering new chemical and biological weapons, cyber attacks, and safety critical attacks. A current worry he mentioned is targeting conventional weapons using complex computer systems, as is being used in Gaza now. 

At question time I asked Professor Lazar what advice would give the federal government, which has announced a trial of Microsoft Copilot in 50 government agencies. He suggested a Chief AI Officer in an AI Agency to oversee this. Also he suggested funding an AI Safety Institute. He hoped that Copilot would just be used for wording letters.

Professor Lazar  used computer generated images to illustrate his talk. These were based on the poem 'The Second Coming' by William Butler Yeats. This theosophical work has echoes where I am sitting today. The location of the ANU was decided by two theosophists Walter and Marion Mahony Griffin.

Professor Cameron Domenico,
Rutgers University–Newark

Professor Lazar, will be followed by Professor Cameron Domenico, Rutgers University–Newark, and Professor David Thorstad, Vanderbilt University.

ps: If all this catastrophic risks of AI sounds excessively alarmist, consider that Australia is going to build six optionally crewed ships. Each armed with 32 missiles, these ships will be able to sail thousands of kilometers with no one on board. Given the possibility of an enemy jamming the link to the ship, it will be tempting to build in an autonomous mode. 

Australian Government Trial of Generative AI for Law, Education, Health, and Aged Care

The Digital Transformation Agency has announced 50 Australian Public Service (APS) agencies are conducting a 6-month trial of Copilot (Microsoft's implementation of Chat GPT), ending June 2024. Staff first take learning module. Agencies have also been given some general guidance, on Accountability, Transparency, Explainability, Privacy, Fairness, and Wellbeing. Agencies where AI could be controversial include Attorney-General's, Education, Health and Aged Care, Home Affairs, and the National Disability Insurance Agency.

A previous Australian Government got into difficulties with relatively simple technology in what was known as the "Robodebt Scheme", which resulted in several deaths. Careless application of AI has the potential to cause human misery and casualties, on a much larger scale. It is to be hoped the APS applies the technology with thought, so there is no need for a Robo 2.0 Royal Commission.

ps: As it happens I will be running ANU computer project students though some Generative AI exercises using Copilot, over the next few weeks. Some of the students are studying programming AI in depth, but this is more general on how to use it.

Australian Universities Accord Final Report: According to Generative AI

For the fun of it I asked Microsoft Copilot to summarize and critique the Australian Universities Accord Final Report. Here is what it came up with. ;-)

Summary:

The Australian Universities Accord Final Report, released by the Hon Jason Clare MP, Minister for Education, contains 47 recommendations aimed at reforming higher education to meet Australia’s future skills needs. Here are the key points:

1. Reform Goals: The report sets a vision for the next decade and beyond. It aims to enhance the higher education sector’s effectiveness and responsiveness.

2. Workforce Qualifications: The report recommends that at least 80% of the workforce should have either a vocational education and training (VET) or university qualification by 2050. Currently, this figure stands at 60%1.

3. Long-Term Plan: The Accord Final Report provides a long-term reform plan to address the evolving demands of Australia’s workforce and economy.

For more detailed insights, you can refer to the Summary Report prepared by the Department of Education2. It outlines the themes and issues identified in the full report.

Remember, this is just a concise overview, and I recommend exploring the full report for comprehensive information.

Critique:

Let’s delve into a concise critique of the Australian Universities Accord Final Report. Here are some key points to consider:

1. Ambitious Goals: The report sets ambitious targets, such as aiming for 80% of the workforce to have either a vocational education and training (VET) or university qualification by 2050. While this goal is commendable, achieving it will require substantial effort and investment.

2. Implementation Challenges: The report outlines several recommendations, but translating them into effective policies and practices may face challenges. Implementation across diverse universities and regions could be complex.

3. Equity and Access: While the report emphasizes workforce qualifications, it’s essential to ensure equitable access to higher education. Addressing barriers faced by disadvantaged groups, including Indigenous students and those from low socioeconomic backgrounds, is crucial.

4. Funding and Resources: The report doesn’t delve deeply into funding mechanisms. Sustainable funding models are essential to support quality education, research, and infrastructure.

5. Industry Collaboration: The report highlights the need for stronger collaboration between universities and industries. However, striking the right balance without compromising academic autonomy is critical.

6. Digital Transformation: The report acknowledges the importance of digital skills. Universities must adapt swiftly to technological advancements and prepare graduates for a rapidly changing job market.

7. Research and Innovation: While workforce qualifications are vital, universities’ role in research, innovation, and knowledge creation should not be overshadowed. Balancing practical skills with intellectual growth is essential.

In summary, the Accord Final Report provides a roadmap, but its success hinges on effective execution, stakeholder engagement, and ongoing evaluation. Policymakers, educators, and industry leaders must collaborate to turn these recommendations into meaningful outcomes for Australia’s higher education sector.

Weaponizing ChatGPT

Professor Sarah Kreps,
Cornell University

Since late last year I have been part of presenting, and attended, many seminars on ChatGPT, but this is a first: "Weaponizing ChatGPT?  National Security and the Perils of AI-Generated Texts in Democratic Societies". In this case "weaponising" is not a metaphor, or hyperbole, it really is about using AI to wage war. Greetings from the Australian National University, where a group of defence and AI experts from around the world are meeting, sponsored by the Australian Department of Defence. . This is the sort of seminar where not only have the speakers just published a new books, but so have half the audience. It is a little intimidating.

Professor Sarah Kreps, Cornell University, explained she was in the US Air Force engineering, working on the predator UAV, and AI was a short step from there. She suggested AI started in the civilian sector, and will then be adopted by the military. I am not so sure that is the case. What became web search engines, came out of research sponsored by US DoD. Some of the AI research is similarly sponsored. 

From the Professor's description, the US approach seems to be limited to what the threat might be from an adversary's use of AI. What I suggest western military's need to do is consider how they will use AI. As an example, if AI is being used defensively to create plausible fake news to undermine your nation, how can it be use to create instant factual responses, or offensively to create a largely factual narrative to undermine the enemy. 

Professor Kreps characterized western countries as an open system which could be exploited by misinformation, and AI can be used to enhance this. The result would be customized fake news to appeal to specific groups. 

It would be interesting to see what the ADF's Information Warfare Division is doing with generative AI. Just as Australia's military cyber-security experts have had an increasing role protecting government and civilian systems, with its name changed from Defence Signals Directorate, to Australian Signals Directorate. Is there a similar role for IWD?

Professor Kreps suggested the vocabulary used could indicate when generative AI is used. I suggest that can't be relied on. Professional speechwriters know to use the language and cadence of their customers. It will not be difficult for AI to write using the language of a particular individual, or group.

This would appear to be an area where DARPA, its UK & US equivalents, could provide funding for universities. This could produce free open access tools to counter misinformation.