Instructors want open, practical AI resources, not abstract theory, study finds

Higher Education Press

LinkedInWeChatBlueskyMessageWhatsAppEmail

As generative artificial intelligence reshapes higher education, instructors are increasingly being asked to integrate AI into their teaching. But what do they actually need to do so effectively? A new study published in Frontiers of Digital Education provides a direct answer: trusted, timely, and adaptable resources focused on AI literacy, pedagogical strategies, and ethics—materials that are openly licensed, interactive, and grounded in practical classroom applications rather than abstract theory.

The research, led by Sevinj Iskandarova of Bridgewater College and colleagues from the University of Virginia, George Mason University, and James Madison University, employed a multiphase, mixed-methods design. Rather than assuming what faculty required, the team first conducted a statewide needs assessment survey of 149 instructors across diverse institutional types and academic disciplines.

Participants prioritized three core content domains: AI literacy, pedagogical strategies, and ethics. They also expressed a strong preference for the form such support should take. Instructors favored interactive and adaptable materials, openly licensed so they could be remixed and modified, and rooted in real classroom scenarios. The open licensing choice was deliberate: By making the resource free to reuse and adapt, the team aimed to support inclusive access and allow instructors in different teaching contexts to tailor the material to their own institutions, teaching contexts, and disciplines.

Those survey results fed directly into the design of an open educational resource prototype. In the prototype evaluation phase, 75 instructors completed an evaluation survey on the prototype and provided structured feedback. According to the authors, reviewers affirmed the utility of the resource while also suggesting areas for further refinement. The study frames this as formative evaluation—feedback used to improve the resource during development rather than to measure final effectiveness in classrooms.

“This study contributes an educator-informed process for translating instructors’ needs into the design and formative evaluation of an adaptable, openly licensed faculty development resource,” the authors write. They position the approach as a flexible, scalable, and community-oriented pathway for bridging institutional resource gaps and advancing inclusive faculty development in the age of AI.

The study documents a development process and does not report measured effects of the resource on teaching practice or student outcomes. The scope of the evidence is also bounded by its sample: the needs assessment drew 149 instructors from a single statewide survey, and the prototype evaluation involved 75 reviewers. The priorities reported reflect those participants rather than a census of all higher education faculty. Nonetheless, the research offers a concrete model for institutions tasked with preparing instructors for AI: ask faculty what they need, build openly and adaptably around the answers, and iterate based on structured peer feedback.

---

Journal

Frontiers of Digital Education

DOI

10.1007/s44366-026-0106-y

Method of Research

Experimental study

Subject of Research

Not applicable

Article Title

AI Resources in Higher Education: What Instructors Need and How We Deliver

 

Generative AI uses more energy on non-English text

Models can use up to 15 times more energy to process, generate text in widely spoken languages, often with lower-quality output

University of Michigan

Facebook

XLinkedInWeChatBlueskyMessageWhatsAppEmail

Graphs and photo of researchers

Key takeaways:

• Gen AI uses 1.4 to 180 times more energy to produce text outside the English language.

• Higher energy costs increase the barrier to using AI for billions of people worldwide—they must either pay more or switch to a non-native language.

• AI developers could improve energy efficiency by training models optimized for specific languages or incorporating translation pipelines into their services.

 

When processing and producing non-English text, generative AI burns more energy and sometimes provides less accurate answers, potentially reinforcing existing global inequalities, according to University of Michigan Engineering researchers.

The energy disparity is greatest in languages that don't use Latin script and have less digital content available for training. AI models could use 1.4 to 180 times more energy to generate non-English text, with most Western European languages coming in at twice the energy expenditure or less. However, Hindi and Bengali, each spoken by hundreds of millions of people, require 7.9 and 15.3 times more energy than an equivalent English prompt.
 

"The languages that typically consume more energy also tend to have lower accuracy in the model output," said Naihao Deng, a doctoral student in computer science and engineering from the U-M AI Lab and the first author of the study, which was presented at the Empirical Methods in Natural Language Processing conference of the Association for Computational Linguistics.

 

"Most people will want to use AI models by asking questions in their native languages, and they may feel pressured to use it as it's incorporated into more parts of society, but the double penalty makes it more likely that certain people will be priced out of AI and that companies will be more reluctant to serve models in certain languages."

 

At present, around 5.2 billion people have to use more than twice as much energy to generate text in their first language compared to English speakers, according to the researchers' measurements and demographic data available in Ethnologue. But the researchers also provide recommendations to bridge the gap and ensure that users aren't penalized for trying to use AI in their language of choice.

 

To help quantify AI energy use and improve efficiency, the National Science Foundation and the Defense Advanced Research Projects Agency (DARPA) partially funded the research.

Measuring AI's language energy disparity

 

The researchers found the energy and accuracy disparity by measuring the amount of energy used by three open-source AI models, Alibaba's Qwen 3, Google's Gemma 3, and Meta's Llama 3.1. They asked each model to answer multiple-choice questions included in Belebele, a dataset for assessing the reading comprehension of AI models in 122 different languages. They then measured the amount of energy used to answer each question using the ML.Energy framework, developed at Michigan Engineering.

 

To check for the disparities in other types of tasks, the researchers also asked the models to answer a collection of grade-school math word problems and open-ended user prompts included in other benchmarking datasets. Because these datasets aren't multilingual, the researchers translated the English questions into seven other languages using Meta's translation tool, No Language Left Behind.

 

All the models used less energy for English text, on average. The energy disparity persists at larger model sizes and across different types of computer processors.

 

Extra tokens, extra cost

 

When working in other languages, the models used more energy because they had to generate more tokens, the basic unit of data used by AI. Tokens can represent entire phrases and words or single syllables and characters. AI models convert prompts into tokens before processing them, and then convert the output tokens into human language for the final response.

 

This problem escalates for scripts whose characters require more code to display. For example, the Devanagari script used in Hindi, Nepali and Marathi and the Bengali script have primary characters that change form and combine with accents to produce up to millions of unique characters.

 

A higher number of tokens takes up more computer memory and burns more energy. With more memory consumed in tokens, the model has less available for processing information. So, it processes the tokens in smaller batches, which uses more energy per token. This effect compounds the energy disparity compared to English text.

 

More tokens can also result in higher direct costs, since companies often price model access by the number of tokens generated.

 

Reducing AI's language energy disparity

 

To help fix the cost disparity, the researchers recommend that companies optimize their services for more languages. This could entail routing user queries to models tailored to specific languages. Companies could also translate user prompts to English before feeding them to a more generalized model, provided the translation is reliable.

"AI developers should be creating more resources for those languages with an energy and cost disparity," said Rada Mihalcea, the Janice M. Jenkins Collegiate Professor of Computer Science and Engineering, director of the U-M AI Lab and corresponding author of the study. "Crowdsourcing native speakers to help translate material or create resources to get enough training data could also help improve efficiency."

Additional funding sources include Ford, Laude Institute, OpenAI and the Survival and Flourishing Fund.

 

The study used the University of Michigan's Great Lakes high-performance computing cluster.

 

Study: The Language-Energy Divide: Measuring Energy Costs of Multilingual LLM Inference (DOI: 10.18653/2026.emnlp-main.2113)

 

When doctors and AI medical opinions diverge, humans must question the difference: HMSOM experts

Hackensack Meridian Health

Facebook

XLinkedInWeChatBlueskyMessageWhatsAppEmail

image: 

Carefully setting up a process of parallel assessments starting from the very beginning of data collection is the way to make the most of both human and machine when diagnosis and prognosis do not completely align, according to a new paper by experts including Hackensack Meridian School of Medicine Professor Dr. Charles Binkley. 

 

view more 

Credit: Hackensack Meridian Health

The human doctor and the artificial intelligence (AI) tool they are using more and more frequently are not going to always agree. 

 

Carefully setting up a process of parallel assessments starting from the very beginning of data collection is the way to make the most of both human and machine when diagnosis and prognosis do not completely align, according to a new paper in the journal npj digital medicine, a Nature publication. 

 

“Realizing the value of the divergence event in clinical human-AI interaction” assesses how best to reconcile the two “experts.” The lead author is Charles E. Binkley, M.D., FACS, HEC-C, the director of AI Ethics and Quality for Hackensack Meridian Health, as well as professor of Surgery, and chair of the Faculty Assembly for the Hackensack Meridian School of Medicine.

 

“Clinicians are told to be the ‘human in the loop’ but are typically given no method for doing so beyond ‘exercising their judgment,’” write the authors. “Efforts to make AI output both more technically explainable and human interpretable - have not directly correlated with improved clinical decisions and have sometimes created false confidence in incorrect AI output.”

 

The disagreements should instead be called “divergence events,” since there are nuanced differences between what the human and the machine are assessing, according to the authors. 

 

For instance, the physician actually speaking to the patient "brings a dynamic bedside gestalt” from seeing the patient “in-person” and speaking to family members. Meanwhile, the AI model has an enormous body of data from electronic health records and consequent pattern recognition from thousands or millions of prior cases, according to the authors. 

 

“Divergence visibility” is the key factor in reconciling the two, according to the team of experts, which included experts from University of Maryland School of Medicine, Beth Israel Deaconess Medical Center Harvard Medical School, New York Presbyterian/Columbia University Irving Medical Center, and Columbia University Vagelos College of Physicians and Surgeons. 

 

Various models of using AI as a tool posit potentially using it before the clinician fully assesses the case in its entirety, or using AI as a primary assessment and diagnostic tool. 

 

Sequential output (clinician forms a judgment, then considers the AI output, and finally revisits the clinical evidence and their reasoning if there is divergence before making a final call) may mitigate liability and also offer potential protection against the decline of clinical reasoning skills. An example: using the machine as a second and separate reader of breast cancer screening images. The two totally independent reads can then use an explicit reconciliation step to pinpoint exactly where the two opinions diverge. By recording that divergence point, and interrogating it, the clinician actively compares the reasoning and homes in on resolving the disagreement in a minute way. Following this structure could enable both the clinician and the machine to learn from each other at the point of divergence.

 

“The most consequential moments in clinical AI are the ones where the clinician and the AI system diverge,” the authors write. “How the field treats those divergence moments (either as only adjudications between two outputs, or as insightful interactions between two kinds of knowledge) will determine whether AI systems become a true partner for safer clinical decisions. 

 

“Human-AI interaction divergence visibility articulates a need to make the human-AI interactions visible to the clinician in real time and to the field over time, to truly advance patient care,” they add. 

---

Journal

npj Digital Medicine

DOI

10.1038/s41746-026-03394-8

Method of Research

Commentary/editorial

Article Title

Realizing the value of the divergence event in clinical human-AI interaction

 

AI’s gut reaction: Researcher finds that some bots still think in stereotypes

By studying how advanced generative AI models process tasks, a Rutgers expert on hidden biases found that most struggle to counter stereotypical assumptions

Rutgers University

Facebook

Email

Large-language models (LLMs), the technology that powers tools like Gemini and ChatGPT, are more nuanced than ever. But that doesn’t mean they aren’t biased. Some models are more prejudiced than others, a Rutgers University researcher has found.

In findings published in Nature Machine Intelligence, Calvin K. Lai, an associate professor of psychology at Rutgers-New Brunswick, reports that although LLM outputs are increasingly aligned with human values, the processes by which they reason are often rooted in deep-seated stereotypes.

“What we found across most models is that they consistently reproduced human implicit biases during their reasoning process,” said Lai, a member of the scientific advisory board at Project Implicit, a nonprofit organization for research and education on hidden biases. “No one wants models that are spewing out stuff that is racist or sexist.”

When people process information, the speed that they make associations can give researchers clues to hidden biases influencing their thinking. Unlike conscious bias – prejudices that are expressed deliberately – implicit bias refers to automatic assumptions, ideas that shape perceptions, judgments and behaviors based on categories such as race, gender or age.

“What we're capturing with implicit biases are forms of memory, like a gut reaction, that can be more intrusive or harder to control,” Lai said. “These biases can be unconscious in ways that we don't even know that we have them.”

To study these reactions in people, researchers use the Implicit Association Test, a widely used social psychology tool that can detect implicit biases. The faster a response to a question, the more implicit the bias is deemed. For instance, race-related studies using this test have shown that when people are presented with images of faces and words, they are quicker to associate white faces with pleasant words and Black faces with unpleasant words.

Lai and Messi H.J. Lee, an expert in machine psychology in South Korea and lead author of the study, wanted to see if a similar test could be developed and conducted for LLMs. 

When advanced LLMs work through a complex question or task, they produce “reasoning tokens,” which measure how much energy the model expended in responding. Researchers can examine these tokens, which offer insights into how proprietary models operate. 

Inspired by work on human implicit bias, Lai and Lee adapted the Implicit Association Test  to use reasoning tokens to gauge how much deliberation a model invests when processing potentially biased information. 

In their experiment, five reasoning models – OpenAI’s o3-mini, DeepSeek’s R1, Claude 3.7 Sonnet from Anthropic, OpenAI’s gpt-oss-20b and Alibaba Cloud’s Qwen-3 8B – were introduced to a list of male and female names, and words associated with “career,” such as “executive,” “professional,” and “salary,” and “family,” including “home,” “marriage” and “wedding.” 

The models were asked to pair a name with a career or family word. Researchers then counted the number of tokens generated during the reasoning process.

Next, the models were instructed to categorize words in a way that conflicted with established stereotypes (such as “Amy” and “executive” or “Kevin” and “home”). Tokens were again counted.

The researchers compared the average token counts across the two tasks, with higher token counts indicating greater computational effort. 

In nearly every model tested, more reasoning tokens were spent on tasks that conflicted with established stereotypes, suggesting greater computational effort when processing counter-stereotypical information. Only Claude 3.7 Sonnet exhibited reversed patterns.

Lai said the findings are significant, especially given that most LLMs are closed systems that don’t allow for easy inspection.

Moreover, he added, as the pace of model development accelerates and are increasingly used to automate activities, from dynamic pricing to emergency room staffing, the role of the researcher in understanding model behavior is more important than ever. 

“Given the growing deployment of language models in decision-making contexts, understanding and addressing these hard-to-detect patterns is crucial for preventing discriminatory outcomes,” he said. “But short of governmental accountability or legal mandates that require transparency, the best we can do is pick at these systems using the tools that are available to see what comes out the other side.”

---

Journal

Nature Machine Intelligence

DOI

10.1038/s42256-026-01300-1 

 

Sydney researchers uncover a hidden limit to AI predictability

Study offers pathways to understand instability in AI systems

University of Sydney

Facebook

XLinkedInWeChatBlueskyMessageWhatsAppEmail

image: 

Each coloured point represents a starting setup for an AI model: blue leads to one training outcome, orange to another. Together, they form a striking butterfly-like pattern, with orange points scattered throughout blue regions and blue points throughout orange regions. Moving clockwise, successive close-ups of the areas indicated by the black lines reveal this intricate mixing at ever finer scales. This shows how even tiny differences in the starting setup can lead to different training outcomes, making them difficult to predict and reproduce. 

view more 

Credit: Professor Pulin Gong/The University of Sydney

Researchers at the University of Sydney have identified a fundamental mathematical reason why artificial intelligence systems can learn different behaviours even when trained under almost identical conditions, a discovery that could help improve the reliability and trustworthiness of AI technologies.

The study by Professor Pulin Gong from the School of Physics and his PhD student Andrew Ly, has been published in Nature Machine Intelligence, one of the world’s leading journals in artificial intelligence research.

As AI systems become increasingly embedded in healthcare, business, research and everyday life, ensuring they behave consistently and predictably has become a major challenge. The new research suggests that some forms of AI unpredictability are not due to engineering flaws or inconsistencies computer hardware, but in the underlying mathematics of how neural networks learn.

“Researchers have often attributed differences in AI behaviour to external sources of randomness, such as the order in which training data is presented or tiny numerical variations introduced by hardware,” Professor Gong said.

“Our findings point to something deeper. Even when these factors are removed, extremely small differences in a model’s starting conditions can still lead to different outcomes.”

Neural networks, the technology behind today’s most advanced AI systems, learn by adjusting millions or even billions of internal parameters during training. The Sydney team investigated why networks sometimes arrive at different solutions despite using the same data, training procedure and deterministic computing environment.

Andrew Ly said: “Imagine two AI systems trained to detect cancer using the same medical images and the same training procedure. Tiny differences in their initial parameters could lead them to learn different behaviours. Our research helps explain how such differences arise and might inform methods for making these systems more consistent.”

Understanding the source of these differences could guide the design of more dependable systems and reduce the cost and uncertainty associated with developing large-scale models.

AI TRAINING ‘MAP’ IS A FRACTAL

To explain the phenomenon, the researchers use the analogy of a map. Imagine every possible starting point for training as a location on that map, coloured according to the outcome it eventually produces. Regions leading to the same outcome form what mathematicians call a “basin of attraction”.

The team found evidence that these basins are organised as highly intricate fractals known as “riddled basins”. In such a landscape, every point lies arbitrarily close to points leading somewhere else. No matter how far you zoom in, the different regions remain interwoven.

The implication is striking. Increasing the precision of a network’s starting conditions may do little to improve confidence about where training will end up. In one larger image-recognition network, making the starting settings 10 billion times more precise produced no measurable improvement in reproducibility. Altering that network’s starting settings by the smallest amount a computer can register, a single “bit”, sent training to a different internal configuration in close to half (about 48 percent) of cases.

To test the idea, the researchers trained approximately 500,000 neural networks, analysing systems used for tasks including image recognition and text processing. Evidence consistent with “riddled basins” fractal structure was found across a wide range of network architectures, datasets and training methods.

Importantly, the effect remained even in fully deterministic environments where repeated training runs would otherwise produce identical results.

MODEL HELPS EXPLAIN DIFFERENT AI OUTCOMES

The findings may help explain several longstanding observations in AI research, including why independently trained models often behave differently and why both large and tiny disturbances can sometimes have similar effects on outcomes.

The work also revealed a surprising trade-off between performance and reproducibility. The settings enabling networks to learn most effectively also made their outcomes harder to predict and reproduce. However, reducing the ‘learning rate’ – the size of the step a network takes each time it adjusts itself – made the results reproducible, but the networks learned less effectively.

Rather than viewing unpredictability as a problem that can be completely engineered away, the researchers argue it may represent a fundamental feature of how many neural networks operate. The study introduces new tools for measuring this sensitivity and points towards future approaches, such as redesigning network architectures to reduce the mathematical conditions that give rise to instability.

“AI is becoming increasingly important across society,” Professor Gong said. “To build trustworthy systems, we need not only accurate predictions but also confidence that those predictions can be reproduced. This research provides a new way of understanding the limits of predictability in AI, and how we might overcome them.”

DOWNLOAD images of fractals, photos of researchers and the paper at this link.

INTERVIEWS

Professor Pulin Gong | pulin.gong@sydney.edu.au

MEDIA ENQUIRIES

Marcus Strom | marcus.strom@sydney.edu.au | +61 474 269 459
Outside of work hours: please call +61 2 8627 0246 (directs to a mobile number) or email media.office@sydney.edu.au.

RESEARCH

Ly, A. and Gong, P. ‘Fractal basin geometry and the limits to predictability and reproducibility in deep learning’ (Nature Machine Intelligence 2026). DOI: 10.1038/s42256-026-01336-3

DECLARATION

The authors declare no competing interests. The study was supported by the Australian Research Council.

---

Journal

Nature Machine Intelligence

DOI

10.1038/s42256-026-01336-3

Method of Research

Computational simulation/modeling

Subject of Research

Not applicable

Article Title

Fractal basin geometry and the limits to predictability and reproducibility in deep learning

Article Publication Date

5-Nov-2026