TOKYO — Research papers from 14 academic institutions in eight countries — including Japan, South Korea and China — contained hidden prompts directing artificial intelligence tools to give them good reviews, Nikkei has found.
Nikkei looked at English-language preprints — manuscripts that have yet to undergo formal peer review — on the academic research platform arXiv.
University of Hawaiʻi at Mānoa researchers have made a significant advance in studying how artificial intelligence (AI) understands endangered languages. This research could help communities document and maintain their languages, support language learning and make technology more accessible to speakers of minority languages.
The paper by Kaiying Lin, a PhD graduate in linguistics from UH Mānoa, and Department of Information and Computer Sciences Assistant Professor Haopeng Zhang, introduces the first benchmark for evaluating large language models (AI systems that process and generate text) on low-resource Austronesian languages. The study focuses on three Formosan (Indigenous peoples and languages of Taiwan) languages spoken in Taiwan—Atayal, Amis and Paiwan—that are at risk of disappearing.
Using a new benchmark called FORMOSANBENCH, Lin and Zhang tested AI systems on tasks such as machine translation, automatic speech recognition and text summarization. The findings revealed a large gap between AI performance in widely spoken languages such as English, and these smaller, endangered languages. Even when AI models were given examples or fine-tuned with extra data, they struggled to perform well.
“These results show that current AI systems are not yet capable of supporting low-resource languages,” Lin said.
Zhang added, “By highlighting these gaps, we hope to guide future development toward more inclusive technology that can help preserve endangered languages.”
The research team has made all datasets and code publicly available to encourage further work in this area. The preprint of the study is available online, and the study has been accepted into the 2025 Conference on Empirical Methods in Natural Language Processing in Suzhou, China, an internationally recognized premier AI conference.
The Department of Information and Computer Sciences is housed in UH Mānoa’s College of Natural Sciences, and the Department of Linguistics is housed in UH Mānoa’s College of Arts, Languages & Letters.
OpenAI is reorganizing its Model Behavior team, a small but influential group of researchers who shape how the company’s AI models interact with people, TechCrunch has learned.
In an August memo to staff seen by TechCrunch, OpenAI’s chief research officer Mark Chen said the Model Behavior team — which consists of roughly 14 researchers — would be joining the Post Training team, a larger research group responsible for improving the company’s AI models after their initial pre-training.
As part of the changes, the Model Behavior team will now report to OpenAI’s Post Training lead Max Schwarzer. An OpenAI spokesperson confirmed these changes to TechCrunch.
The Model Behavior team’s founding leader, Joanne Jang, is also moving on to start a new project at the company. In an interview with TechCrunch, Jang says she’s building out a new research team called OAI Labs, which will be responsible for “inventing and prototyping new interfaces for how people collaborate with AI.”
The Model Behavior team has become one of OpenAI’s key research groups, responsible for shaping the personality of the company’s AI models and for reducing sycophancy — which occurs when AI models simply agree with and reinforce user beliefs, even unhealthy ones, rather than offering balanced responses. The team has also worked on navigating political bias in model responses and helped OpenAI define its stance on AI consciousness.
In the memo to staff, Chen said that now is the time to bring the work of OpenAI’s Model Behavior team closer to core model development. By doing so, the company is signaling that the “personality” of its AI is now considered a critical factor in how the technology evolves.
In recent months, OpenAI has faced increased scrutiny over the behavior of its AI models. Users strongly objected to personality changes made to GPT-5, which the company said exhibited lower rates of sycophancy but seemed colder to some users. This led OpenAI to restore access to some of its legacy models, such as GPT-4o, and to release an update to make the newer GPT-5 responses feel “warmer and friendlier” without increasing sycophancy.
Techcrunch event
San Francisco
|
October 27-29, 2025
OpenAI and all AI model developers have to walk a fine line to make their AI chatbots friendly to talk to but not sycophantic. In August, the parents of a 16-year-old boy sued OpenAI over ChatGPT’s alleged role in their son’s suicide. The boy, Adam Raine, confided some of his suicidal thoughts and plans to ChatGPT (specifically a version powered by GPT-4o), according to court documents, in the months leading up to his death. The lawsuit alleges that GPT-4o failed to push back on his suicidal ideations.
The Model Behavior team has worked on every OpenAI model since GPT-4, including GPT-4o, GPT-4.5, and GPT-5. Before starting the unit, Jang previously worked on projects such as Dall-E 2, OpenAI’s early image-generation tool.
Jang announced in a post on X last week that she’s leaving the team to “begin something new at OpenAI.” The former head of Model Behavior has been with OpenAI for nearly four years.
Jang told TechCrunch she will serve as the general manager of OAI Labs, which will report to Chen for now. However, it’s early days, and it’s not clear yet what those novel interfaces will be, she said.
“I’m really excited to explore patterns that move us beyond the chat paradigm, which is currently associated more with companionship, or even agents, where there’s an emphasis on autonomy,” said Jang. “I’ve been thinking of [AI systems] as instruments for thinking, making, playing, doing, learning, and connecting.”
When asked whether OAI Labs will collaborate on these novel interfaces with former Apple design chief Jony Ive — who’s now working with OpenAI on a family of AI hardware devices — Jang said she’s open to lots of ideas. However, she said she’ll likely start with research areas she’s more familiar with.
This story was updated to include a link to Jang’s post announcing her new position, which was released after this story published. We also clarify the models that OpenAI’s Model Behavior team worked on.
At the Hebrew University of Jerusalem, scientists created a new way to tell someone’s age using just a bit of DNA. This method uses a blood sample and a small part of your genetic code to give highly accurate results. It doesn’t rely on external features or medical history like other age tests often do. Even better, it stays accurate no matter your sex, weight, or smoking status.
Bracha Ochana and Daniel Nudelman led the team, guided by Professors Kaplan, Dor, and Shemer. They developed a tool called MAgeNet that uses artificial intelligence to study DNA methylation patterns. DNA methylation is a process that adds chemical tags to DNA as the body ages. By training deep learning networks on these patterns, they predicted age with just a 1.36-year error in people under 50.
Time leaves invisible fingerprints on your cells. One of the most telling signs of age in your body is DNA methylation—the addition of methyl groups (CH₃) to your DNA. These chemical tags don’t change your genetic code, but they do affect how your genes behave. And over time, these tags build up in ways that mirror the passage of years.
What makes the new method so effective is its focus. Instead of analyzing thousands of areas in the genome, MAgeNet zeroes in on just two short genomic regions. This tight focus, combined with high-resolution scanning at the single-molecule level, allows the AI to read the methylation patterns like a molecular clock. Professor Kaplan explains it simply: “The passage of time leaves measurable marks on our DNA. Our model decodes those marks with astonishing precision.”
The study, recently published in Cell Reports, used blood samples from more than 300 healthy individuals. It also included data from a 10-year follow-up of the Jerusalem Perinatal Study, which tracks health information across lifetimes. That long-term data, led by Professor Hagit Hochner from the Faculty of Medicine, helped the team confirm that MAgeNet works not just in the short term but also across decades.
Importantly, the model’s accuracy held up no matter the person’s sex, body mass index, or smoking history—factors that often throw off similar tests. That consistency means the tool could be widely used in both clinical and non-clinical settings.
The medical uses are easy to imagine. Knowing someone’s true biological age can help doctors make better decisions about care, especially when signs of aging don’t match the number of candles on a birthday cake. Personalized treatment plans could become more effective if based on what’s happening at the cellular level, not just what appears on a chart.
But this breakthrough also has major potential in the world of forensic science. Law enforcement teams could one day use this method to estimate the age of a suspect based solely on a few cells left behind. That’s a big step forward from current forensic DNA tools, which are good at identifying a person but struggle with age.
“This gives us a new window into how aging works at the cellular level,” says Professor Dor. “It’s a powerful example of what happens when biology meets AI.”
As they worked with the data, the researchers noticed something else: DNA doesn’t just age randomly. Some changes happen in bursts. Others follow slow, steady patterns—almost like ticking clocks inside each cell. These new observations may help explain why people age differently, even when they’re the same age chronologically.
“It’s not just about knowing your age,” adds Professor Shemer. “It’s about understanding how your cells keep track of time, molecule by molecule.”
This could also impact the growing field of longevity research. Scientists are increasingly interested in how biological aging differs from the simple count of years lived. The ability to measure age so precisely from such a small DNA sample may become a key tool in developing future anti-aging therapies or drugs that slow down cellular wear and tear.
The method created by the Hebrew University team marks a turning point in how we think about aging, identity, and health. In the past, DNA told us who we are. Now it can tell us how old we truly are—and possibly how long we’ll stay healthy. The implications stretch from hospital rooms to courtrooms.
As the world faces rising healthcare demands from aging populations, tools like MAgeNet offer a smarter, faster way to assess risk, track longevity, and understand what aging really means. It’s no longer just a number on your ID.
Thanks to AI and a deep dive into the chemistry of life, age has become something you can measure with stunning accuracy, from the inside out.
The Guardian view on Trump and the Fed: independence is no substitute for accountability | Editorial
Building Trust in Military AI Starts with Opening the Black Box – War on the Rocks
SDAIA Supports Saudi Arabia’s Leadership in Shaping Global AI Ethics, Policy, and Research – وكالة الأنباء السعودية
Journey to 1000 models: Scaling Instagram’s recommendation system
Mumbai-based Perplexity Alternative Has 60k+ Users Without Funding
VEX Robotics launches AI-powered classroom robotics system
Kayak and Expedia race to build AI travel agents that turn social posts into itineraries
Happy 4th of July! 🎆 Made with Veo 3 in Gemini
Macron says UK and France have duty to tackle illegal migration ‘with humanity, solidarity and firmness’ – UK politics live | Politics
OpenAI 🤝 @teamganassi
