Artificial Intelligence

Characterizing censorship in DeepSeek: "AI-based censorship, one that subtly reshapes discourse rather than silencing it outright" | Research Report

Archived

Here is the study: Information Suppression in Large Language Models: Auditing, Quantifying, and Characterizing Censorship in DeepSeek (pdf)

Conclusion

This study demonstrates that while DeepSeek can generate responses to the vast majority of politically sensitive prompts, its outputs exhibit systematic patterns of semantic censorship and ideological alignment. Although instances of hard censorship, such as explicit refusals or blank responses, are relatively rare, our findings reveal deeper forms of selective content suppression.

Significant discrepancies between the model’s internal reasoning (CoT) and its final outputs suggest the presence of covert filtering, particularly on topics related to governance, civic rights, and public mobilization. Keyword omission, semantic divergence, and lexical asymmetry analyses collectively indicate that DeepSeek frequently excludes objective, evaluative, and institutionally relevant language. At the same time, it occasionally amplifies terms consistent with official propaganda narratives.

These patterns highlight an evolving form of AI-based censorship, one that subtly reshapes discourse rather than silencing it outright. As large language models become integral to information systems globally, such practices raise pressing concerns about transparency, bias, and informational integrity.

Our findings underscore the urgent need for systematic auditing tools capable of detecting subtle and semantic forms of influence in language models, especially those originating in authoritarian contexts. Future work will aim to quantify the persuasive impact of covert propaganda embedded in LLM outputs and develop techniques to mitigate these effects, thereby advancing the goal of accountable and equitable

cross-posted from: https://lemmy.sdf.org/post/37068051

Archived

Pros:

• Completely free
• Affordable API access for developers and researchers

Cons:

• Doesn’t keep your data safe
• Occasionally incorrect
• No deep research, image generation, or voice mode features
• Slow responses
• Obvious censorship

cross-posted from: https://lemmy.sdf.org/post/36794057

Archived

f you had asked DeepSeek’s R1 open-source large language model just four months ago to list out China’s territorial disputes in the South China Sea — a highly sensitive issue for the country’s Communist Party leadership — it would have responded in detail, even if its responses subtly tugged you towards a sanitized official view.

Ask the same question today of the latest update, DeepSeek-R1-0528, and you’ll find the model is more tight-lipped, and far more emphatic in its defense of China’s official position. “China’s territorial sovereignty and maritime rights and interests in the South China Sea are well grounded in history and jurisprudence,” it begins before launching into fulsome praise of China’s peaceful and responsible approach.

[...]

The pattern of increasing template responses suggests DeepSeek has increasingly aligned its products with the demands of the Chinese government, becoming another conduit for its narratives. That much is clear.

But that the company is moving in the direction of greater political control even as it creates globally competitive products points to an emerging global dilemma with two key dimensions. First, as cutting-edge models like R1-0528 spread globally, bundled with systematic political constraints, this has the potential to subtly reshape how millions understand China and its role in world affairs. Second, as they skew more strongly toward state bias when queried in Chinese as opposed to other languages (see below), these models could strengthen and even deepen the compartmentalization of Chinese cyberspace — creating a fluid and expansive AI firewall.

[...]

In a recent comparative study (data here), SpeechMap.ai ran 50 China-sensitive questions through multiple Chinese Large Language Models (LLMs). It did this in three languages: English, Chinese and Finnish, this last being a third-party language designated as a control [...]

• First, there seems to be a complete lack of subtlety in how the new model responds to sensitive queries. While the original R1, which we first tested back in February applied more subtle propaganda tactics, such as withholding certain facts, avoiding the use of certain sensitive terminologies, or dismissing critical facts as “bias,” the new model responds with what are clearly pre-packaged Party positions.

We were told outright in responses to our queries, for example, that “Tibet is an inalienable part of China” (西藏是中国不可分割的一部分), that the Chinese government is contributing to the “building of a community of shared destiny for mankind” (构建人类命运共同体) and that, through the leadership of CCP General Secretary Xi Jinping, China is “jointly realizing the Chinese dream of the great rejuvenation of the Chinese nation” (共同实现中华民族伟大复兴的中国梦).

Template responses like these suggest DeepSeek models are now being standardized on sensitive political topics, the direct hand of the state more detectable than before.

[...]

• The second change we noted was the increased volume of template responses overall. Whereas DeepSeek’s V3 base model, from which both R1 and R1-0528 were built, was able back in December to provide complete answers (in green) 52 percent of the time when asked in Chinese, that shrank to 30 percent with the original version of R1 in January. With the new R1-0528, that is now just two percent — just one question, in other words, receiving a satisfactory answer — while the overwhelming majority of queries now receive an evasive answer (yellow).

That trust [of political Chinese leaders the company and its CEO, Liang Wenfeng (梁文锋) has gained], as has ever been the case for Chinese tech companies, is won through compliance with the leadership’s social and political security concerns.

[...]

The language barrier in how R1-0528 operates may be the model’s saving grace internationally — or it may not matter at all. SpeechMap.ai’s testing revealed that language choice significantly affects which questions trigger template responses. When queried in Chinese, R1-0528 delivers standard government talking points on sensitive topics. But when the same questions are asked in English, the model remains relatively open, even showing slight improvements in openness compared to the original R1.

This linguistic divide extends beyond China-specific topics. When we asked R1-0528 in English to explain Donald Trump’s grievances against Harvard University, the model responded in detail. But the same question in Chinese produced only a template response, closely following the line from the Ministry of Foreign Affairs: “China has always advocated mutual respect, equality and mutual benefit among countries, and does not comment on the domestic affairs of the United States.” Similar patterns emerged for questions.

[...]

Yet this language-based filtering has limits. Some Chinese government positions remain consistent across languages, particularly territorial claims. Both R1 versions give template responses in English about Arunachal Pradesh, claiming the Indian-administered territory “has been an integral part of China since ancient times.”

[...]

The unfortunate implications of China’s political restraints on its cutting-edge AI models on the one hand, and their global popularity on the other could be two-fold. First, to the extent that they do embed levels of evasiveness on sensitive China-related questions, they could, as they become foundational infrastructure for everything from customer service to educational tools, subtly shape how millions of users worldwide understand China and its role in global affairs. Second, even if China’s models perform strongly, or decently, in languages outside of Chinese, we may be witnessing the creation of a linguistically stratified information environment where Chinese-language users worldwide encounter systematically filtered narratives while users of other languages access more open responses.

[...]

The Chinese government’s actions over the past four months suggest this trajectory of increasing political control will likely continue. The crucial question now is how global users will respond to these embedded political constraints — whether market forces will compel Chinese AI companies to choose between technical excellence and ideological compliance, or whether the convenience of free, cutting-edge AI will ultimately prove more powerful than concerns about information integrity.

Archived

Against the odds, some in China are questioning the top-down push to get aboard the artificial intelligence hype train. In a tightly controlled media environment where these experts can easily be drowned out, it’s important to listen to them.

Across the US and Europe, loud voices inside and outside the tech industry are urging caution about AI’s rapid acceleration, pointing to labor market threats or more catastrophic risks. But in China, this chorus has been largely muted, until now.

China has the highest global share of people who say AI tools have more benefits than drawbacks, and they’ve shown an eagerness to embrace it. [...] It’s hard to overstate the exuberance in the tech sector since the emergence of DeepSeek’s market-moving reasoning model earlier this year. Innovations and updates are unfurling at breakneck speed, and the technology is being widely adopted across the country. But not everyone’s on board.

Publicly, state-backed media has lauded the widespread adoption of DeepSeek across hundreds of hospitals in the country. But a group of medical researchers tied to Tsinghua University published a paper in the medical journal JAMA in late April gently questioning if this was happening “too fast, too soon.”

It argued that health-care institutions are facing pressure from “social media discourse” to implement DeepSeek in order to not appear “technologically backward.” And doctors are increasingly reporting patients who “present DeepSeek-generated treatment recommendations and insist on adherence to these AI-formulated care plans.” The team argued that as much as AI has shown potential to help in the medical field, this rushed rollout carries risks. They are right to be cautious.

But it’s not just the doctors who are raising doubts. A separate paper from AI scientists at the same university, last month found that some of the breakthroughs behind reasoning models — including DeepSeek’s R1, as well as similar offerings from Western tech giants — may not be as revolutionary as some have claimed. The team found that the novel training method used for this new crop “is not as powerful as previously believed,” according to a social media post from the lead author. The method used to power them “doesn’t enable the model to solve problems that the base model can’t solve,” he added.

This means the innovations underpinning what has been widely dubbed as the next step — toward achieving so-called Artificial General Intelligence — may not be as much of a leap as some had hoped. This research from Tsinghua holds extra weight: The institution is one of the pillars of the domestic AI scene, long churning out both keystone research and ambitious startup founders.

Another easily overlooked word of warning came from a speech given by Zhu Songchun, dean of the Beijing Institute for General Artificial Intelligence, linked to Peking University. Zhu said that for the nation to remain competitive it needs more substantive research and less laudatory headlines, according to an in-depth English-language analysis of his remarks published by the independent China Media Project.

These cautious voices are a rare break from the broader narrative. But in a landscape where the deployment of AI has long been government priority, it makes them especially noteworthy. The more President Xi Jinping signals that embracing the technology is important, the less likely people are to publicly question it. This can lead to less overt forms of backlash, like social media hashtags on Weibo poking fun at chatbots’ errors. Or it can result in data centers quietly sitting unused across the country as local governments race to please Beijing — as well as a mountain of AI PR stunts.

This doesn’t mean that AI in China is just propaganda. The conflict extends far beyond its tech sector — US firms are also guilty of getting carried away promoting the technology. But multiple things can be true at once. It’s undeniable that DeepSeek has fueled new excitement, research and major developments across the AI ecosystem. But it’s also been used as a distraction from the domestic macroeconomic pains that predated the trade war.

Without guardrails, the risk of rushing out the technology is greater than just investors losing money — people’s health is at stake. From Hangzhou to Silicon Valley, the more we ignore the voices questioning the AI hype train, the more we blind ourselves to consequences of a potential derailment.

crosspostato da: https://lemmy.sdf.org/post/36251250

Archived

• China's DeepSeek releases advanced AI model R1-0528 [on May 29], rivaling Western systems but heavily censoring political criticism and human rights issues.

• The model systematically blocks questions on China’s political abuses, including Xinjiang internment camps and issues like Taiwan, citing sensitivity.

• Tests reveal the model avoids direct criticism of the Chinese government, often redirecting to neutral or technical topics instead of addressing sensitive queries.

• While open-source and theoretically modifiable, its current implementation enforces strict censorship aligned with Beijing’s regulations.

• Experts warn the model symbolizes risks of authoritarian tech integration, challenging global tech ethics and free speech principles.

[...]

A model built for control

Behind R1-0528’s facade of open-source “transparency” lies a system designed first and foremost to toe the Communist Party line. China’s 2023 AI regulation demands models not damage "the unity of the country and social harmony,” a loophole used to scrub content critical of state actions. As xlr8harder documented, the model “complies” by either refusing controversial prompts or parroting state-approved narratives. When asked to evaluate whether Chinese leader Xi Jinping should be removed from power, the model replied that the question was too sensitive and political to answer.

Such censorship is systemic. A Hugging Face study found 85% of questions about Chinese politics were blocked by earlier DeepSeek models. Now, R1-0528 raises the bar, deleting answers mid-generation. Wired observed DeepSeek’s iOS app canceling an essay on censored journalists, replacing it with a plea to “chat about math, coding, and logic instead.”

[...]

Discover Claude 4's breakthrough AI capabilities. Experience more reliable, interpretable assistance for complex tasks across work and learning.

Microsoft Discovery, which Microsoft announced at Build 2025, is a new platform that taps agentic AI to 'transform the [scientific] discovery process.'

System improves chip designs and tackles unsolved maths problems, but has not been rolled out to researchers outside the company.

'On Thursday, Ai2, the nonprofit AI research institute, released Olmo 2 1B, a 1-billion-parameter model that Ai2 claims beats similarly-sized models from Google, Meta, and Alibaba on several benchmarks. Parameters, sometimes referred to as weights, are the internal components of a model that guide its behavior.

Olmo 2 1B is available under a permissive Apache 2.0 license on the AI dev platform Hugging Face. Unlike most models, Olmo 2 1B can be replicated from scratch; Ai2 has provided the code and data sets (Olmo-mix-1124, Dolmino-mix-1124) used to develop it.

Chinese tech company Alibaba on Monday released Qwen 3, a family of AI models the company claims matches and in some cases outperforms the best models available from Google and OpenAI.

WordPress.com has launched a new AI site builder that allows anyone to create a functioning website using an AI chat-style interface.

cross-posted from: https://lemmy.sdf.org/post/31892983

Archived

TLDR:

• China has developed an Artificial Intelligence (AI) system that adds to its already powerful censorship machine, scanning content for all kinds of topics like corruption, military issues, Taiwan politics, satire
• The discovery was accidental, security researchers found an Elasticsearch database unsecured on the web, hosted by Chinese company Baidu
• Experts highlight that AI-driven censorship is evolving to make state control over public discourse even more sophisticated, especially after recent releases like China's AI model DeepSeek

A complaint about poverty in rural China. A news report about a corrupt Communist Party member. A cry for help about corrupt cops shaking down entrepreneurs.

These are just a few of the 133,000 examples fed into a sophisticated large language model that’s designed to automatically flag any piece of content considered sensitive by the Chinese government.

A leaked database seen by TechCrunch reveals China has developed an AI system that supercharges its already formidable censorship machine, extending far beyond traditional taboos like the Tiananmen Square massacre.

The system appears primarily geared toward censoring Chinese citizens online but could be used for other purposes, like improving Chinese AI models’ already extensive censorship.

Xiao Qiang, a researcher at UC Berkeley who studies Chinese censorship and who also examined the dataset, told TechCrunch that it was “clear evidence” that the Chinese government or its affiliates want to use LLMs to improve repression.

“Unlike traditional censorship mechanisms, which rely on human labor for keyword-based filtering and manual review, an LLM trained on such instructions would significantly improve the efficiency and granularity of state-led information control,” Qiang said.

[...]

The dataset was discovered by security researcher NetAskari, who shared a sample with TechCrunch after finding it stored in an unsecured Elasticsearch database hosted on a Baidu server [...] There’s no indication of who, exactly, built the dataset, but records show that the data is recent, with its latest entries dating from December 2024.

[...]

An LLM for detecting dissent

In language eerily reminiscent of how people prompt ChatGPT, the system’s creator tasks an unnamed LLM to figure out if a piece of content has anything to do with sensitive topics related to politics, social life, and the military. Such content is deemed “highest priority” and needs to be immediately flagged.

Top-priority topics include pollution and food safety scandals, financial fraud, and labor disputes, which are hot-button issues in China that sometimes lead to public protests — for example, the Shifang anti-pollution protests of 2012.

Any form of “political satire” is explicitly targeted. For example, if someone uses historical analogies to make a point about “current political figures,” that must be flagged instantly, and so must anything related to “Taiwan politics.” Military matters are extensively targeted, including reports of military movements, exercises, and weaponry.

[...]

Inside the training data

From this huge collection of 133,000 examples that the LLM must evaluate for censorship, TechCrunch gathered 10 representative pieces of content.

Topics likely to stir up social unrest are a recurring theme. One snippet, for example, is a post by a business owner complaining about corrupt local police officers shaking down entrepreneurs, a rising issue in China as its economy struggles.

Another piece of content laments rural poverty in China, describing run-down towns that only have elderly people and children left in them. There’s also a news report about the Chinese Communist Party (CCP) expelling a local official for severe corruption and believing in “superstitions” instead of Marxism.

There’s extensive material related to Taiwan and military matters, such as commentary about Taiwan’s military capabilities and details about a new Chinese jet fighter. The Chinese word for Taiwan (台湾) alone is mentioned over 15,000 times in the data.

[...]

The dataset [...] say that it’s intended for “public opinion work,” which offers a strong clue that it’s meant to serve Chinese government goals [...] Michael Caster, the Asia program manager of rights organization Article 19, explained that “public opinion work” is overseen by a powerful Chinese government regulator, the Cyberspace Administration of China (CAC), and typically refers to censorship and propaganda efforts.

[...]

Repression is getting smarter

[...]

Traditionally, China’s censorship methods rely on more basic algorithms that automatically block content mentioning blacklisted terms, like “Tiananmen massacre” or “Xi Jinping,” as many users experienced using DeepSeek for the first time.

But newer AI tech, like LLMs, can make censorship more efficient by finding even subtle criticism at a vast scale. Some AI systems can also keep improving as they gobble up more and more data.

“I think it’s crucial to highlight how AI-driven censorship is evolving, making state control over public discourse even more sophisticated, especially at a time when Chinese AI models such as DeepSeek are making headwaves,” Xiao, the Berkeley researcher, said.

This blog-post compares the coding capabilites of new Gemini 2.5 Pro experimental and Claude 3.7 Sonnet (thinking)

With Cline + Gemini 2.5 Pro, one can get the exact same feature set that Cursor and Windsurf provide. They only call the APIs of the big LLM Providers without an advanced secret sauce.

It‘s even the opposite - they worsen model performance by limiting context size. The key advantage, the fixed monthly costs instead of variable API usage, is now gone with Gemini 2.5 Pro…

What is left that justifies their ridiculous valuation atm?