Is DeepSeek safe?
DeepSeek’s political censorship received a lot of attention in the media for being so apparent, but tech-savvy people can still circumvent that flaw. On the other hand, privacy is a more pressing concern: DeepSeek’s online app offers no way to prevent your data from being sent directly to China. Digitaal Vlaanderen says the models are not suitable for use in the Flemish Government and the Data Protection Authority has also launched an investigation.
Censorship
DeepSeek’s models are censored. Tests show that both DeepSeek models give evasive answers or abruptly end conversations on topics such as Tiananmen Square, Taiwan or other politically charged topics in the Chinese context. DeepSeek’s political censorship is therefore more noticeable than that of US AI systems.
However, that does not mean that systems like OpenAI’s ChatGPT or Anthropic’s Claude are free from restrictions. Whereas DeepSeek aligns itself with the Chinese regime, US models display commercial and social self-censorship. ChatGPT, for instance, avoids statements that could cause reputational damage or legal complications. Claude even goes a step further. It presents itself as constitutional AI: a tool in which ethical principles and rules of conduct (including parts of the Universal Declaration of Human Rights) were built into the model.
The line between ‘human alignment’ and censorship
‘Human alignment’ is adjusting a model so that it match users’ expectations. Language models must always align with human expectations: they use data from millions of sources that are full of errors, biases and other less-than-perfect patterns. Without adjustment, they will reproduce errors and biases, and perpetuate misconceptions and inequalities. But even when they generate ‘creative’ new output, things can go horribly wrong because the models don’t possess an ethical compass.
But where is the line between human alignment and censorship? All language models need some form of human alignment to be useful. In DeepSeek, political and ideological censorship is particularly evident. In Western models, output is more likely to be adjusted in line with legislation, ethical standards and commercial interests. This manifests itself, among other things, in avoiding hate speech and sensitive views or, like ChatGPT does, by pleasing users, sometimes even to the extent that it tolerates factual untruths (‘sycophantic behaviour’). As a user, it is best to be aware of these limitations so that you can circumvent them if you run into limitations.
Bypassing censorship and restrictions
To what extent can you circumvent or remedy limitations in a language model?
With DeepSeek, you can: the code is openly available (open weights), allowing other AI developers to modify the model completely and to remove censorship. Perplexity managed to do this in February 2025. They launched a model ‘R1 1776’: a decensored version of DeepSeek's R1 model. DeepSeek models ‘V3’ and ‘R1’ thus offer customisation options that closed models like ChatGPT (GPT-4) do not.
Removing the restrictions does require advanced technical knowledge and a lot of computing power. As such, these capabilities remain out of reach for the majority of users.
Garbage in = garbage out
Language models use data from millions of sources that are full of errors and biases (historical or otherwise). Without adjustment, the system will reproduce these, sometimes with not-so-pretty outcomes.
- Language models and toxicity
AI models trained on online conversations copy that language, including the swearing and insults. Without adjustment, an AI model will use aggressive or inappropriate language. - Stereotypes in professions
In the past, doctors were often men. Train a language model with data going back 50 years and the ‘doctors’ in the texts will almost always be men. AI uses probability as a rule and adopts patterns in the data into its output: thus, it can help perpetuate stereotypes, and sometimes even reinforce biases. - Misinformation
Widespread conspiracy theories or misleading texts can be further reinforced by language models. AI systems cannot fact-check: they rely on patterns in existing data, causing often shared misinformation to be re-produced and disseminated. In addition, models can give a legitimate appearance to false information through their credible writing style. They may even try to convince you of the ‘alternative’ truth.
Privacy
Besides censorship, privacy is a big issue with DeepSeek. When used via the web page and mobile app, users – often unknowingly – share large amounts of information with the chatbot. ChatGPT offers an opt-out for the use of input in model training, but DeepSeek lacks that option. Anyone who does not download the DeepSeek models and uses them locally (on a very powerful computer) or in a private cloud thus sends all their data directly to China.
“ DeepSeek is not currently suitable for professional use in the Flemish government according to these guidelines: data entered can be used for other purposes and the results generated remain the property of the company. ”
Think twice before using DeepSeek
Digitaal Vlaanderen says DeepSeek’s models are not suitable for professional use in the Flemish government. The Belgian Data Protection Authority has also launched an investigation. The data you enter can be reused for other purposes and the results generated remain property of the company. Belgium is not the only one who has concerns: Italy has now blocked DeepSeek, and the Dutch Data Protection Authority warns of privacy risks.
The combination of censorship and privacy risks makes DeepSeek a risky choice, causing many companies and governments to approach the tool cautiously or even dismissively. Nevertheless, DeepSeek and its innovations remain a gamechanger. A new phase in AI history has begun.
Vinnie De Craim
I believe technology only truly matters when it creates societal impact. With a background in sociology and technology, I work on digital innovation in the public sector—where thinking and doing go hand in hand.
My approach is hands-on and impact-oriented. I bring critical analysis, but I’m equally driven by building concrete, meaningful results. Whether it's generative AI, policy development, or science communication, I’m always looking to bridge knowledge and real-world application—with pragmatism and a keen eye for opportunity.
I also believe in the power of language and storytelling to make complex ideas accessible. My background in the arts helps me not just understand technology and policy, but translate them in ways that resonate with different audiences.
Expertise: statistics, generative AI/AI, data science
Johan Desseyn
Johan Desseyn is founder and researcher at Mpiris. He started Mpiris 2017 to help governments and organisations make good decisions. He does so by providing policy-oriented social science research. With Mpiris, Johan also responds fully to the digitalisation of our society. Where possible, he applies technology-based methods to answer his clients' questions.
For Johan and his team, the emergence of Generative AI meant not only looking for ways to use these tools internally in an effective and responsible way, but also committing to sharing the accumulated expertise with his clients in government and education.
Expertise: policy-oriented research, education and training, governments, labour market policy, social policy
Related courses
HPC Café
Seminarie - Online, Karlsruhe - KIT, HammerHAI, Training geselecteerd in de Belgische AI Factory Antenna (AIFA)