On Monday, January 27, Chinese AI startup DeepSeek took the tech world by storm, dethroning its rival, OpenAI, for the coveted spot for most downloaded free app in the US on Apple’s App Store (DeepSeek’s own AI Assistant surpassed ChatGPT). The popularity of DeepSeek began surfacing earlier in the month when the startup released R1, its reasoning model which is a type of large language model (LLM) that can perform complex reasoning tasks. The startups’ R1 model generated much buzz due to the alleged claim that it only cost $6m to train which is a fraction of the “over $100m” alluded to OpenAI by boss Sam Altman when discussing the creation of GPT-4. DeepSeek is also providing its R1 models under an open-source license, enabling free use by third parties.
Nevertheless, are there any other factors that set DeepSeek apart from its rivals and are there any limitations upon its functioning given that it is a Chinese app? This article will briefly explore these nuances.
What is DeepSeek?
DeepSeek is an AI development firm that is based in Hangzhou, China, founded by Liang Wenfeng in 2013. Wenfeng is equally the co-founder of High-Flyer, a China-based quantitive hedge fund that is the mother company of DeepSeek. The startup is currently operating as an independent AI research lab which began in 2023 and is currently under the purview of High-Flyer.
DeepSeek offers two main models to its users: DeepSeek-V3 and DeepSeek-R1. The former focuses on general-purpose tasks such as answering questions and generating content while the latter excels in complex reasoning and logical problem solving. The new R1 model sets DeepSeek apart from its competitors as unlike traditional AI models, it uses a Mixture-of-Experts (MoE) Architecture which ensures that only a small portion of the model’s parameters are active at once. This feature reduces computing power requirements and enables faster learning while still maintaining the efficiency of the system. Additionally, unlike ChatGPT’s premium features, all capabilities of DeepSeek are free to use for its users with no restrictions. Users of DeepSeek will equally not encounter shortcomings in terms of the information bank available to them as through the AI, the web can be searched directly to obtain real-time updated information.
Why Are There Concerns About the Operation of This AI?
Even though many of the features offered by DeepSeek are attractive to both users and developers alike, concerns have been raised about data privacy while using AI. According to DeepSeek’s privacy policy, the service collects a trove of user data including chat and search history, the device a user is on, keystroke patterns, IP addresses, internet connection, and activity from other apps. Even though it is not abnormal for AI models to collect such a substantial amount of information, security concerns are raised over the fact that the data that is collected from Americans is sent to servers in China. Samm Sacks, a research scholar in cybersecurity at Yale remarked :“that data, in the aggregate, can be used to glean insights into a population, or user behavior that could be used to create more effective phishing attacks or other nefarious manipulation campaigns.”
Featured image courtesy of: Chris Ellis/Midjourney
When it comes to Europeans, questions are posed as to whether the AI is compliant with the GDPR. Despite being available in Europe and collecting EU personal data like email addresses and user interactions, DeepSeek’s privacy policy doesn’t make a single reference to the GDPR. Additionally, DeepSeek makes no mention of the sources it uses for its data training and does not outline whether EU citizen data was used during the learning process, breaching Articles 13 and 14 of the GDPR.
DeepSeek AI also makes users responsible for the outputs generated by the AI tool, as stated in their Terms of Service. A list is outlined with a range of prohibited outputs that range from generating discriminatory content to damaging society or the economy. Furthermore, DeepSeek is censored when tackling certain sensitive topics, refusing to provide answers.
The US-based AI security and compliance company Enkrypt AI found that DeepSeek’s R1 model was 11 times more likely to generate harmful output compared to OpenAI’s o1 model. Even in testing done by the cyber intelligence platform Kela, DeepSeek could generate a ransomware program alongside step-by-step instructions on how to distribute it and target victims. Through a further study, 78% of cybersecurity tests successfully tricked the R1 model into generating insecure or malicious code. The uncertainties surrounding the system’s functioning have led to some regulators either trying to completely ban the app or restricting its accessibility in certain sectors. Such is the case for Italy, Taiwan, Australia, South Korea, the USA, and India.
What’s next for DeepSeek?
Even though countries are attempting to ban DeepSeek, such a ban would be difficult to enforce in practice considering the model is fully open source and can be utilized by third parties. In terms of competition within the market, perspectives diverge on whether or not DeepSeek will fully replace ChatGPT. Rather, some claim that DeepSeek has started a new movement within the industry that will encourage more companies to enter and provide AI models that are slightly differentiated from each other at a cheaper price. Regardless, within the regulatory space, there are still many unanswered questions pertaining to DeepSeek´s functioning that need to be addressed before the public is fully receptive to its integration.
Featured image courtesy of Nicolas Economou/NurPhoto/Shutterstock. DeepSeek upended the tech world and Wall Street with its high-performing AI model that it claims was trained at a fraction of the cost of major competitors.
