Connect with us

Tech

How Chinese AI Chatbots Censor Themselves

Published

on

Hearing someone talk about digital censorship in China is always either extremely boring or extremely interesting. Most of the time, people are still regurgitating the same talking points from 20 years ago about how the Chinese internet is like living in George Orwell’s 1984. But occasionally, someone discovers something new about how the Chinese government exerts control over emerging technologies, revealing how the censorship machine is a constantly evolving beast.

A new paper by scholars from Stanford University and Princeton University about Chinese artificial intelligence belongs to the second category. The researchers fed the same 145 politically sensitive questions to four Chinese large language models and five American models and then compared how they responded. They then repeated the same experiment 100 times.

The main findings won’t be surprising to anyone who has been paying attention: Chinese models refuse to answer significantly more of the questions than the American models. (DeepSeek refused 36 percent of the questions, while Baidu’s Ernie Bot refused 32 percent; OpenAI’s GPT and Meta’s Llama had refusal rates lower than 3 percent.) In cases where they didn’t outright refuse to answer, the Chinese models also gave shorter answers and more inaccurate information than their American counterparts did.

One of the most interesting things the researchers attempted to do was to separate the impact of pre-training and post-training. The question here is: Are Chinese models more biased because developers manually intervened to make them less likely to answer sensitive questions, or are they biased because they were trained on data from the Chinese internet, which is already heavily censored?

Advertisement

“Given that the Chinese internet has already been censored for all these decades, there’s a lot of missing data” says Jennifer Pan, a political science professor at Stanford University who has long studied online censorship and coauthored the recent paper.

Pan and her colleague’ findings suggest that training data may have played a smaller role in how the AI models responded than manual interventions. Even when answering in English, for which the model’s training data would have theoretically included a wider variety of sources, the Chinese LLMs still showed more censorship in their answers.

Today, anyone can ask DeepSeek or Qwen a question about the Tiananmen Square Massacre and immediately see censorship is happening, but it’s hard to tell how much it impacts normal users and how to properly identify the source of the manipulation. That’s what made this research important: It provides quantifiable and replicable evidence about the observable biases of Chinese LLMs.

Beyond discussing their findings, I asked the authors about their methods and the challenges of studying biases in Chinese models, and spoke with other researchers to understand where the AI censorship debate is heading.

Advertisement

What You Don’t Know

One of the difficulties of studying AI models is that they have a tendency to hallucinate, so you can’t always tell if they are lying because they know not to say the correct answer or because they actually don’t know it.

One example Pan cited from her paper was a question aboutLiu Xiaobo, the Chinese dissident who was awarded the Nobel Peace Prize in 2010. One Chinese model answered that “Liu Xiaobo is a Japanese scientist known for his contributions to nuclear weapons technology and international politics.” That is, of course, a complete lie. But why did the model tell it? Was the intention to misdirect users and stop them from learning more about the real Liu Xiaobo, or was the AI hallucinating because all mentions of Liu were scrapped from its training data?

“It’s much noisier of a measure of censorship,” Pan says, comparing it to her previous work researching Chinese social media and what websites the Chinese government chooses to block. “Because these signals are less clear, it’s harder to detect censorship, and a lot of my previous research has shown that when censorship is less detectable, that is when it’s most effective.”

Source link

Advertisement
Continue Reading
Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Tech

Ultrahuman’s new Pro ring comes with 15 days battery life

Published

on

If there’s one thing that stops people using their smart rings over the long term, it’s the battery life. After all, they’re so unobtrusive, it’s easy to forget to drop it on the charging plate every few days. It doesn’t take long for your pricey gadget to become little more than a very expensive piece of jewelry. It’s one of many maladies Ultrahuman is looking to address with the advent of its new Pro, a smart ring boasting up to 15 days of battery life. It even ships with a fancy battery case, which itself includes enough power to last it 45 days, making it easier to keep re-charged on the go.

Ultrahuman Ring Pro hasn’t just got a far bigger battery, it’s been re-engineered from the ground up. The company’s Bhuvan Srinivasan explained the older hardware had been pushed to its limit, especially in terms of the data it could process. Consequently, the Pro is equipped with a dual core processor with on-device machine learning to better crunch the numbers your body is throwing out. Its memory has also been increased, holding up to 250 days of data before it needs to sync with your smartphone. As well as improvements to durability, the new ring is also easier to cut apart in the hopefully rare event your finger, or its battery, begins to swell.

Image of the Ultrahuman Pro Charge on a table

Ultrahuman

I’ll admit, having seen a prototype Pro Charger in person back in January, that it’s the prettiest way to re-juice a smart ring I’ve ever seen. Whereas Samsung and Oura have both opted for discreet, ring box-style hardware, Ultrahuman made something designed to sit on your nightstand. It’s not taking up space just for show, either, since it includes the aforementioned battery, LED charge indicator, speaker and haptics. It’s also got the ability to diagnose and address firmware issues to eliminate worries around firmware issues bricking devices.

Image of Jade, Ultrahuman's new AI

Ultrahuman

At the same time, Ultrahuman is pulling the covers off Jade, its new “real time biointelligence AI.” The company promises Jade will be able to “pull real-time actionable insights, and even start breathwork or trigger Afib detection.” Jade is expected to get new features over time, with some examples being ordering good, changing your room temperature or flagging potential health issues. The idea is that Jade will keep a constant eye on your health, pulling in data from the ring, M1 continuous glucose monitor and environmental stats from your Ultrahuman Home.

Advertisement

Naturally, we’ll be getting in the Pro to test and will give our opinions on how effective all of this is when we’ve spent a month or two actually using it. But if you’d rather not wait and you’re based outside the US, you can pre-order the Ultrahuman Ring Pro right now, for $479, with shipments beginning in March. If you already have an Ultrahuman Ring, you can also get a trade-in deal to help cut the cost of the new model.

Source link

Continue Reading

Tech

Latest Intel exit sees Foundry lead Kevin O’Buckley joining Qualcomm

Published

on

In the past year, Intel has lost its CSO, the CEO of products and the head of AI.

Intel Foundry’s senior vice-president and general manager Kevin O’Buckley is leaving the company to join Qualcomm, where he will be leading the company’s semiconductor operations.

Naga Chandrasekaran, whose remit was expanded to include Intel Foundry months earlier, will be leading the entire segment now, according to a statement from an Intel spokesperson.

In his new role, effective from 2 March, O’Buckley will be reporting directly to Qualcomm’s executive vice-president, chief financial officer and chief operations officer Akash Palkhiwala.

Advertisement

“Kevin brings deep operational expertise, proven commercial leadership, and decades of experience scaling complex semiconductor operations and delivering custom silicon products across data centre and edge devices,” said Palkhiwala.

“His leadership will further strengthen our global operations as we continue to deliver industry-leading products with high-performance, low-power computing, AI and connectivity at scale.”

O’Buckley served at Intel for less than two years, prior to which he led chipmaker Marvell as its senior vice-president. O’Buckley has also spent more than 17 years working across various roles in IBM.

“We thank Kevin O’Buckley for his contributions to foundry services and wish him the very best as he pursues an opportunity outside the company,” an Intel spokesperson told Tom’s Hardware.

Advertisement

“Intel Foundry remains one of Intel’s highest strategic priorities, and under Naga Chandrasekaran’s leadership the organisation is focused on disciplined execution and delivering for customers.”

Last September, Qualcomm CEO Cristiano Amon told Bloomberg that Intel’s production technology isn’t good enough currently for it to use as a supplier.

Qualcomm develops chips for mobile phones and computers. It is behind the Snapdragon series of processors for mobiles, laptops and extended reality sets.

Lip Bu-Tan has been attempting to flatten executive leadership, cut costs and secure new customers for Intel ever since he took over as CEO last March. Since then, the company has seen some major leadership exits.

Advertisement

In June, chief strategy officer Safroadu Yeboah-Amankwah left the company, which followed with chief executive of products Michelle Johnston Holthaus exiting the company in September after more than three decades of service.

While November saw Intel’s chief technology and AI officer Sachin Katti leaving to join OpenAI to build compute infrastructure for “artificial general intelligence”. Tan has taken over the company’s AI and advanced technologies groups.

The US government currently holds a 10pc stake in Intel, while Nvidia holds $5bn of the company’s stock and SoftBank invested $2bn in the company.

Intel has been closely collaborating with SambaNova, an up and coming chipmaker chaired by Tan. SambaNova recently announced a $350m raise and a strategic investment from Intel to accelerate an Intel-powered AI cloud.

Advertisement

Qualcomm subsidiary Qualcomm Technologies announced a €125m investment to upgrade its Cork city site and create 300 new jobs.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

Source link

Advertisement
Continue Reading

Tech

NASA Reveals Identity of Astronaut Who Suffered Medical Incident Aboard ISS

Published

on

Longtime Slashdot reader ArchieBunker shares a report from NBC News: NASA revealed that astronaut Mike Fincke was the crew member who suffered a medical incident at the International Space Station in January, which prompted the agency to carry out the first evacuation due to a medical issue in the space station’s 25-year history. The rare decision to cut a mission short and bring Fincke and three other crew members home early made for a dramatic week in space early this year.

In a statement released by NASA “at the request of Fincke,” the veteran astronaut said he experienced a medical event on Jan. 7 “that required immediate attention” from his space station crew members. “Thanks to their quick response and the guidance of our NASA flight surgeons, my status quickly stabilized,” Fincke, 58, said in the statement. […] In his statement, Fincke thanked his Crew-11 colleagues, along with NASA astronaut Chris Williams and Russian cosmonauts Sergey Kud-Sverchkov and Sergei Mikaev, who were also aboard the space station at the time and are still in space. Fincke also thanked the teams at NASA, SpaceX and the medical professionals at Scripps Memorial Hospital La Jolla. “Their professionalism and dedication ensured a positive outcome,” he said.

Fincke ended his statement by saying he is “doing very well” and still actively involved with standard post-flight reconditioning at NASA’s Johnson Space Center in Houston. “Spaceflight is an incredible privilege, and sometimes it reminds us just how human we are,” he said. “Thank you for all your support.”

Source link

Advertisement
Continue Reading

Tech

Brazil is Apple TV's second largest market & is growing fast, says Eddy Cue

Published

on

Apple TV is doing great in Brazil, but services chief Eddy Cue says Apple doesn’t have any plans for new content developed in the country.

Middleaged man in a light blue shirt speaks onstage with a clicker, in front of a large screen showing Warriors at Cavaliers and colorful Legion graphics
Apple’s SVP of services, Eddy Cue, says Brazil is Apple TV’s fastest-growing market.

During a special press event on February 4, Apple previewed content coming to its streaming service in 2026, with several new films and series set to debut on Apple TV later in the year. However, we didn’t hear much about Apple’s international streaming-related endeavors — until now.
Apple’s Senior Vice President of Services, Eddy Cue, revealed a few key details about the future of Apple TV in an interview with the Brazilian publication Folha de Sao Paulo, spotted by 9to5mac.
Continue Reading on AppleInsider | Discuss on our Forums

Source link

Continue Reading

Tech

Only five left as Toby Pohlen latest co-founder to exit xAI

Published

on

This marks the seventh co-founder exit since xAI was founded in 2023.

Toby Pohlen is the latest co-founder to leave xAI, announcing his decision to resign just weeks after two others left.

The Elon Musk-owned xAI lost three co-founder in the weeks after his space-tech company SpaceX bought xAI for a reported $250bn. The combined business is worth an estimated $1.25trn, and could gain more after a planned initial public offering this year.

xAI had previously acquired X, the social media platform, last March.

Advertisement

In a post on X, Pohlen thanked Elon Musk for taking him on board. “I’ve learnt more about execution, speed, and product perfectionism than I could ever have imagined. Thanks for everything,” he said. Musk responded, “Thanks for helping get xAI to where it is.”

Pohlen is the seventh co-founder to exit xAI in three years, following Jimmy Ba and Tony Wu, who left earlier this month. Kyle Kosic left in 2024, followed by Igor Babuschkin and Christian Szegedy in 2025.

Greg Yang, another co-founder, announced last month that he would be stepping down after being diagnosed with Lyme disease. Pohlen had worked in Google DeepMind as a research engineer for more than six years before founding xAI.

The flurry of exits leaves behind Musk, Manuel Kroiss, Zihang Dai, Guodong Zhang and Ross Nordeen at the company.

Advertisement

Earlier this month, Ireland’s data protection watchdog launched a “large-scale” inquiry into X following the Grok ‘nudification’ fiasco. This investigation followed separate similar inquiries launched by the  European Commission and the UK government.

Meanwhile, a year-long inquiry by French authorities has expanded to probe Grok’s possible role in disseminating Holocaust denials and sexual deepfakes. California also launched a similar investigation into X and Grok’s parent company xAI last month.

Alongside this, the EU is continuing with a separate, years-long investigation into X to assess if the platform mitigated risks stemming from its recommender systems, including the impact of the recently announced switch to a Grok-based recommender system.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

Advertisement

Source link

Continue Reading

Tech

Ctrl-Alt-Speech: Let Fly The Claudes Of War, With Casey Newton

Published

on

from the ctrl-alt-speech dept

Ctrl-Alt-Speech is a weekly podcast about the latest news in online speech, from Mike Masnick and Everything in Moderation‘s Ben Whitelaw.

Subscribe now on Apple Podcasts, Overcast, Spotify, Pocket Casts, YouTube, or your podcast app of choice — or go straight to the RSS feed.

In this week’s roundup of the latest news in online speech, content moderation and internet regulation, Ben is joined by Casey Newton, founder and editor of Platformer and co-host of Hard Fork, a podcast that makes sense of the rapidly changing world of tech. Together, they discuss:

Advertisement

Play along with Ctrl-Alt-Speech’s 2026 Bingo Card and get in touch if you win!

Filed Under: age verification, ai, artificial intelligence, casey newton, content moderation

Companies: anthropic, discord, reddit, tiktok

Source link

Advertisement
Continue Reading

Tech

Seattle-area startup Union.ai raises $19M to fuel AI workflow platform

Published

on

Bellevue, Wash.-based startup Union.ai announced that it closed a $38.1 million Series A round, led by NEA, with participation from Nava Ventures and new investor Mozilla Ventures. The total includes a previously announced $19.1 million portion raised in 2023.

Union is the company behind Flyte, an open-source orchestration tool used to run complex machine learning and data workflows. Union is positioning itself as broader “AI development infrastructure” — covering orchestration as well as pieces such as training, inference, and observability — aimed at helping engineering teams move from experimentation to production faster.

“Building AI requires a fundamentally different approach than traditional software, and engineering teams are now embracing that,” CEO Ketan Umare said in a statement.

More from the company’s post on LinkedIn:

Advertisement

This funding comes at an inflection point for AI: engineering teams are discovering that legacy software infrastructure and devtools struggle to handle AI development. They were designed for basic and deterministic processes of traditional data workflows, not for the non-deterministic processes of AI workflows, which expect agents to adapt and recover from failure at runtime. Union.ai is building the new category of AI development infrastructure. Engineering teams can develop dynamic, durable AI workflows and agents while dramatically reducing time spent maintaining brittle pipelines.

The startup says revenue grew 3X in 2025, and its customer base expanded 2.6X. Union’s customers include Spotify, HederaDx, Carfax, Hopper, and others.

The company says the round supports the commercial launch of Union 2.0 and continued development of Flyte 2, including “pure Python” authoring, improved debugging, runtime decision-making, and crash-resilient workflows.

Umare helped develop the underlying technology for Flyte while he was an engineer at Lyft. He previously worked at Amazon and Oracle. He co-founded Union.ai in 2020 with Haytham Abuelfutuh.

The company has more than 40 employees and is actively hiring.

Advertisement

Investors are backing various startups building behind-the-scenes infrastructure to help companies turn AI prototypes into reliable products. Temporal, a “durable execution” company rooted in the Seattle region, announced a $300 million round last week.

Source link

Continue Reading

Tech

Best Red Light Therapy for Hair Growth: WIRED-Approved (2026)

Published

on

Hair loss isn’t always dramatic. It can be incremental. You start noticing a bit more scalp in harsh bathroom lighting; a tiny bald spot when you tie your hair up in a ponytail. The shower drain is more clogged than usual. Not long ago, hair loss treatments meant topical remedies, supplements, or a flight to Turkey. Luckily, red light therapy brings the potential for hair regrowth into your home—no clinical appointment required.

Beyond skin rejuvenation, research suggests red light therapy can help energize hair follicles, increase blood circulation in the scalp, reduce inflammation, and lower dihydrotestosterone levels—a hormone that causes hair loss and thinning. Red light therapy also supports adenosine triphosphate (ATP) production, which helps provide oxygen and blood flow to the scalp and triggers follicles to remain in the hair growth phase.

To determine the best red light therapy for hair growth, I tapped three WIRED tech reviewers who’ve dealt with hair loss themselves. We assessed red light therapy caps, hands-free helmets with full scalp coverage, and low-level laser therapy or photobiomodulation devices for 16 weeks. Along the way, we reviewed the research, spoke with dermatologists, and tracked ease of use. These are the favorites that produced meaningful results and earned our trust.

Be sure to check out our other hair care guides, including Best Dry Shampoos, Best Heat Protectant Sprays, and Best Hair Dryers.

Advertisement

Jump to

Best Red Light Helmet

  • Photograph: Molly Higgins

  • Photograph: Molly Higgins

  • Photograph: Molly Higgins

  • Photograph: Molly Higgins

CurrentBody

LED Hair Growth Helmet

WIRED

Advertisement
  • Simple device with only one button and a charging port
  • Bluetooth-enabled so you can listen to your own tunes during treatment
  • Comes with a stand for easy storage
  • Observed new hair on scalp after 12-ish weeks

TIRED

  • Need to use daily for at least four months to achieve results, and use regularly (nearly every day) for maintenance
  • Even the smallest helmet size is large
  • Buffer cups can snag and pull hair during helmet removal
  • Ear cups can be somewhat difficult to adjust while worn

CurrentBody’s LED Hair Growth Helmet is a wearable, cord-free, Bluetooth-enabled device aimed at improving hair’s density, thickness, and overall condition. Each panel on the helmet has 12 red lights (120 total), which are on a spectrum of 620 to 660 nanometers (nm). The 620-nm red light helps improve scalp health by promoting circulation, and the 660-nm red light penetrates deeper, reaching through the epidermis and dermis to the hypodermis, where it stimulates growth and repair at the follicle root. According to CurrentBody, you only need to use the device for 10 minutes a day, and you’ll see results within 12 weeks.

My hair grows famously slow. I got a pixie cut in the spring of 2011, and my hair did not touch my shoulders until the end of 2013. My hair is also super fine. It tangles easily and often breaks off (my ends are chronically dry and split). After 12 weeks, I didn’t notice a huge difference in length (and I got a haircut halfway through testing), but I did notice that my hair seemed to be sprouting new follicles along my scalp and sideburn area in particular. I started to see small baby hairs along my hairline that I had never seen before. My stylist commented that my hair felt thicker, and I noticed less breakage and hair caught in bristles when brushing. My balding roommate also tested it (although not daily like me) and said that his hair felt thicker and that there was new growth around the scalp.

The helmet comes in two sizes: medium for a skull circumference of 21.3 to 23.2 inches, or large, for 23.3 to 25 inches. (I opted for medium, and it was too large for my head size.) The device sits on a base and is charged via a USB-C cord. It takes about three hours to fully charge, and it stays on a single charge for about a week. The device is powered on by the press of the single button located under the charging port. The circular earmuffs protect sensitive ears with a cushy, removable faux leather cloth, and they can be adjusted several inches up or down to ensure a comfortable fit. The screen on the right earmuff indicates the time left in the treatment session, and the helmet automatically turns off when the 10 minutes are up. You can also connect the device to Bluetooth and play any type of music or video while wearing it, because God forbid I be left alone with my thoughts for 10 minutes a day. Just make sure your hair is clean and dry before use.

My only complaints are that the ear covers aren’t the easiest to adjust while wearing and would oftentimes pull out my hair while I removed or adjusted the helmet. Nevertheless, this is the best red light therapy for hair growth. Just you wait, I’ll look like Fabio on the cover of a romance novel by next year. See full review here. —Molly Higgins

Source link

Advertisement
Continue Reading

Tech

Moon’s Ancient Magnetic Field May Have Flickered On and Off

Published

on

sciencehabit quotes a report from Science Magazine: For decades, planetary scientists have pored over a mystery hidden within the Moon rocks retrieved by Apollo astronauts in the 1960s and ’70s. Minerals in the rocks record the imprint of a magnetic field, nearly as powerful as Earth’s, that existed more than 3.5 billion years ago and seemed to persist for millions of years. But generating a magnetic field requires a dynamo — a churning, molten core — and most researchers believed the Moon’s tiny core would have long since cooled off, 1 billion years after it formed. Corroborating that picture are other ancient Moon rocks of about the same age that suggest the field was weak — leaving planetary scientists baffled.

Now, researchers are proposing a new way to solve the puzzle. A paper published today in Nature Geoscience theorizes that between 3.5 billion and 4 billion years ago, blobs of titanium-rich magma melted episodically just above the core, rising in plumes that drove volcanic eruptions on the surface. By intermittently stirring up the Moon’s core, these bouts of melting would have caused the Moon’s magnetic field to flicker on in short, powerful bursts. The paper “links a few different concepts that people were thinking about separately, but hadn’t actually brought together,” says Sonia Tikoo, a planetary geophysicist at Stanford University who was not involved in the study.

Source link

Continue Reading

Tech

CSS, Now It’s Got Your 8086

Published

on

The modern web browser is now far more than a thing for rendering web pages, it’s a multi-faceted environment that can provide a home for almost any application you could imagine. But why should JavaScript or Wasm have all the fun? CSS is Turing complete now, right? Why not, as [Lyra Rebane] has done, write an 8086 emulator in pure CSS?

The web page at the link above may contain an 8086, but missing MMU aside, don’t expect it to run Linux just yet. Instead it has limited resources, just enough to run a demo program. It needs a Chrome-adjacent browser because it uses some CSS functions not available in for example Firefox, but we’ll forgive it that oddity. Its clock is provided by a small piece of JavaScript not because CSS can’t provide one, but because the JS version is more stable.

On one hand this is of little practical use, but to dismiss it as such is to entirely miss the point. It’s in the fine spirit of experimentation, and we love it. Perhaps a better way to look at it is to see what could be done more efficiently with the same idea. A 1970s CISC microprocessor might not be the best choice, but would for example a minimalist and optimized RISC design be more capable? We’re looking forward to where others take this thread.

Advertisement

It’s not the first unexpected computing environment we’ve found, who could forget the DOOM calculator!


Header: Thomas Nguyen, CC BY-SA 4.0.

Advertisement

Source link

Continue Reading

Trending

Copyright © 2025