DeepSeek: Potential and challenges in education

In the last few days, we have received quite a few questions about DeepSeek and whether it can be used as part of the education system. Therefore, we are diving into the possibilities of using the model, which has quickly become the most downloaded app on the App Store, and which many praise for being on par with ChatGPT.

DeepSeek R1 is a language model developed by the Chinese AI company DeepSeek, which has proven to be a strong competitor to models such as ChatGPT, Gemini, and Claude. It has been trained at a significantly lower cost than many of its competitors, which has generated a lot of attention in the AI world. Additionally, DeepSeek R1 is open-source under the MIT license, meaning anyone can use it for free, including for commercial purposes, modify it for specific tasks, and review its code. You can access DeepSeek via the website chat.deepseek.com or through their app in the App Store or Google Play. It is also possible to download the models via Hugging Face.

Can we use DeepSeek in education?

If you use DeepSeek via their website or app, you need to be very mindful of the privacy policy regarding the use of the service.

Below, I have asked Claude to review DeepSeek's privacy policy:

• DeepSeek collects many personal data points when you use their services, including profile information, user input, technical information, usage data, and cookies. This is a comprehensive form of data collection.
• DeepSeek shares your information with service providers, advertising partners, their corporate group, and potentially other third parties, such as in business transactions or legal obligations. This means your data can be shared widely.
• Your personal information may be stored on servers outside your home country, possibly in China, where DeepSeek has servers. Data protection laws vary between different countries.
• DeepSeek retains information as long as necessary to provide their services and for other purposes, such as compliance with legal obligations. No specific time limit is mentioned.
• Depending on where you live, you may have certain rights regarding your personal data, such as the right to access, change, delete, etc. However, exercising these rights may affect your ability to use the services.

Overall, DeepSeek collects and shares a significant amount of personal data, and it is unclear how long this data is retained.

As an educational institution, one should always conduct a risk assessment of the AI tools they consider using in education. Here, DeepSeek is unlikely to comply with GDPR, and the company has not publicly disclosed the opportunity to enter into data processing agreements. This is also impossible, as user data is stored on servers in China. These concerns are emphasized by the fact that Italy's data protection authority has now asked DeepSeek to clarify how the company uses and stores personal data – a step that points to the potential GDPR challenges posed by the tool.

Using DeepSeek in Education

When it is still feasible to use DeepSeek in education, it is because one can also download an offline version that can be run on one's computer or a local server.

The video below explains how to perform a local installation of DeepSeek using the program LM Studio. When doing this, everything is executed locally on your computer, and no data is sent to third parties. However, it requires a powerful computer to run the models, and the setup can seem somewhat technical.

We have our own version of DeepSeek running locally at our high school, where we are testing the possibilities in a closed environment. In our tests of DeepSeek at Viden.AI, it appears as a competent language model with both strengths and weaknesses. Therefore, it is about choosing the right language model for the task you want it to assist with.

Bias in DeepSeek R1

Besides data collection via their website, it is also important to be aware that DeepSeek R1 is not without bias. The model tends to avoid answering questions that are controversial with respect to China, such as questions about Taiwan, Arunachal Pradesh, or events like the protests at Tiananmen Square in 1989.

At the same time, there is a risk that DeepSeek could influence users' worldviews by promoting Chinese ideology rather than basing its responses on general, global information. It can also be problematic that the model blocks access to certain information, which can confuse or mislead users. Although some experts argue that these biases are rare and can be mitigated through fine-tuning, and that similar biases are found in models from Western companies, it is emphasized that one should be cautious about using DeepSeek if they are concerned about bias.

Sources:

DeepSeek R1’s bold bet on reinforcement learning: How it outpaced OpenAI at 3% of the cost

DeepSeek R1’s Monday release has sent shockwaves through the AI community, disrupting assumptions about what’s required to achieve cutting-edge AI performance. This story focuses on exactly how DeepSeek managed this feat, and what it means for the vast number of users of AI models. For enterprises developing AI-driven solutions, DeepSeek’s breakthrough challenges assumptions of OpenAI’s dominance -- and offers a blueprint for cost-efficient innovation.

VentureBeatMatt Marshall

DeepSeek V3 and the cost of frontier AI models

The $5M figure for the last training run should not be your basis for how much frontier AI models cost.

InterconnectsNathan Lambert

How DeepSeek Trained AI 30 Times Cheaper?

Learn how DeepSeek defies the norms of AI training by achieving remarkable results at a fraction of the cost. Explore their unique approach.

Analytics VidhyaHimanshi Singh

How DeepSeek R1 Was Trained: A Cost-Efficient Revolution in AI

Open-Source, Affordable, and Matching OpenAI’s Best

MediumAbdullah Afzal

DeepSeek Releases an OpenAI-Comparable Large Language Model Trained at a Fraction of the Cost

DeepSeek-R1 performs on par with OpenAI’s equivalent — but was trained for under $10 million, and is available to download free of charge.

Hackster.ioGareth Halfacree

China’s cheap, open AI model DeepSeek thrills scientists

DeepSeek-R1 performs reasoning tasks at the same level as OpenAI’s o1 — and is open for researchers to examine.

NatureElizabeth Gibney

Here’s everything you need to know about the Chinese wunder-AI 'DeepSeek' that is now sending shockwaves through the AI sector

A surprisingly cheap and effective artificial intelligence from a smaller Chinese tech company is turning conventional ideas in the AI industry upside down. Here’s everything you need to know about DeepSeek.

ComputerworldNiels de Boissezon, Technology Editor

DeepSeek is the hottest new AI chatbot—but it comes with Chinese censorship built in

As with the popular TikTok alternative RedNote, Western users are finding some topics off-limits in DeepSeek R1.

FortuneDavid Meyer

DeepSeek Privacy Policy

DeepSeek is highly biased, don’t use it

DeepSeek has taken the Generative AI arena by storm. Both their models, be it DeepSeek-v3 or DeepSeek-R1 have outperformed SOTA models by a…

Data Science in Your PocketMehul Gupta

DeepSeek collects keystroke data and more, storing it in Chinese servers

You might want to learn about DeepSeek’s privacy policy before you sign up.

MashableCecily Mauran

This article has been machine-translated into English. Therefore, there may be nuances or errors in the content. The Danish version is always up-to-date and accurate.