ABSTRACT
Recent events revolving around the financial turmoil created by the launch of the Deepseeker-R1 model and the response by OpenAI with the release of Chat GPT o3-mini raised the bar of a previously ongoing international AI chatbot competition. The continuous release of new and more efficient models by American and Chinese tech companies is key to understanding the active role the European Union (EU) can play by investing in an EU chatbot rather than passively leaving its member states to develop a constellation of these on their own. Therefore, this article aims to recount the recent developments revolving around AI chatbots, emphasizing how AI companies in the U.S. and China are increasing their efforts to release the best possible AI model used for the functioning of the chatbot, as will be shown by the AI model comparison of DeepSeek-R1 with specific top performing models: ChatGPT-o1, ChatGPT o3-mini, and o3-mini-high and with the non-reasoning model ChatGPT 4.5. Furthermore, it is contended that, within this intense AI competition, the absence of an EU-level chatbot constitutes a significant delay in ensuring that EU values and standards regarding AI and data protection can be safeguarded. Finally, this article clarifies some of the main drawbacks of the absence of an EU chatbot, particularly concerning the protection of personal data.
If 2025 was declared one of the golden years of Artificial Intelligence (AI) breakthroughs and breakups, then the major global state actors would already experience their first industrial revolution of the year: low-cost, high-efficiency AI chatbots (O’Donnel 2025). Awaiting the further developments on cheaper and more efficient “AI Agents,” which generally include all those AI models that can function without any form of human intervention, potentially constituting less data protection issues, the current most used form of AI by individuals and businesses remain AI chatbots.
Why chatbots?
In a study of 2022 surveying 11,000 US citizens, researchers found that these people used, at least sporadically, an AI to solve problems or answer questions daily (27% interacted several times a day, 28% claimed to interact with an AI about once per day, and 44% claimed to have occasionally used an AI) (Kennedy, B. et al. 2023). In the same study, 68% of the people interviewed could easily identify an AI chatbot rather than AI applied to music playlist recommendations or other algorithms applied to security or video surveillance measures, thus showing that one of the chatbots constituted the most well-known platform for the survey participants (Ibid.).
The widespread use of AI chatbots is derived from instances of Artificial Narrow Intelligence (ANI) and Artificial General Intelligence (AGI), meaning that the chatbot’s versatility allows it to carry out both specialized and general tasks in a timely and efficient manner compared to humans. Therefore, some of the main positive aspects that made AI chatbots one of the AIs most used by people entail tasks ranging from aerospace engineering problems, shopping, or enhancing a business plan (Holdsworth, J. 2024). In other words, AI chatbots provide immediate feedback and responses when: solving bothersome, time-consuming tasks, improving productivity, or filling up our spare time with a fictitious pal friend (Brandtzaeg, P.B. et al. 2017, 377 and ff.).
The AI chatbot competition
Beginning in early January 2025, a new “rush” to create the best-performing AI chatbot began. This competition marks a slight change in the pre-existing paradigm in which the U.S. AI companies have little to no foreign AI developers who have rivalled them. At the beginning of this year, tech companies in the US faced, for the first time, substantial competition to get the most attention, with their latest AI model improving their respective chatbots. Since late January 2025, the world has watched AI chatbots becoming a means to translate the wider competition between states with a sort of “AI chatbot rush,” from now on, “the competition.” Interestingly, while the Western part of the world appeared satisfied with their AI achievements in 2024, the Chinese government created the conditions leading to the development of an AI model capable of mimicking and, to some extent, outperforming the best AI models created in the U.S., particularly those released by OpenAI, one of the tech giants leading this industry.
With the launch of China’s latest open-source reasoning model Deepseek-R1 in January 2025 by this until-recently unknown Chinese startup Deepseek, the world witnessed for the first time an AI chatbot capable of reaching the performance of some of the then top existing AI models, ChatGPT-o1 and ChatGPT-o3, requiring drastically fewer economic resources for its creation and functioning. This release was a major unprecedented development in the AI scenario as the latest model before Deepseek-R1, namely ChatGPT-o1, was rivalled by a “cheaper clone” of its previous model, equally efficient, and capable of positively stirring the whole Chinese tech stock market upward while downthrowing Western tech stocks (Ng., K., Dreanon, B. et al. 2025).
Is DeepSeek-R1 that special?: a regular chatbot shocking financial markets
From a critical perspective, Deepseek-R1’s greatest breakthrough was perhaps financial. This model allowed to demonstrate how oblivious Western investors could be to new AI models released around the world, as this Chinese model led to a massive U.S. tech stock plummet. In the wake of the Deepseek-R1 release, as Reuters reported, the biggest hit in Europe was the chipmaker and chip seller ASML (-7%), whereas in the U.S., many more companies went drastically down, ranging from Broadcom Inc. (-17.4%) to ChatGPT (-2.1%), and passing through Google Alphabet (-4.2%) (Carew, S. et al. 2025). Particular attention must be paid to the tech colossus Nvidia (-11.8%), which registered its worst days between the 24th and 27th of January 2025 when it went from $3.5tn to $2.9tn) (Saul, D. 2025; Investing.com). This was a huge hit as this release stained Nvidia’s unexpected performance in 2024 when the company reached a staggering +171.2% (Ermey, R. 2025). Additionally, other tech giants, such as Microsoft, Meta, and Broadcom, experienced lighter yet significant drops.
This highlights that DeepSeek was capable of achieving very similar results without necessarily using an extensive budget, a key yet simple equation allowing drastically cheaper technological advancement (Cisternino 2025). However, within this AI chatbot competition, it is essential to analyze the real metrics of DeepSeek-R1 by comparing them to the top-performing, fairly recent AI models of OpenAI. This is key to understanding how, within this competition, AI companies such as OpenAI and DeepSeek aim to release the best possible model, overthrowing those of the competitors.
What is DeepSeek-R1?
DeepSeek-R1 is an AI reasoning model that allows DeepSeek’s chatbot to function. A reasoning model is a type of Large Language Model (LLM) used to produce responses that simulate human reasoning when answering problems or generating ideas (Kerner, S.M. 2025). From a more technical viewpoint, it can be seen that even if DeepSeek-R1 requires less than $ 6 million to develop compared to the billions invested by OpenAI in similarly performing models, this model achieved similar performance metrics to just some of the OpenAI models while being surpassed by others (Ibid.). Immediately after its release, the R1 model proved better than most AI reasoning chatbots previously created by U.S. AI companies. Experts noticed its improved reasoning to solve logical, mathematical, and other science-based problems, showing impressive achievements, reaching the level of the ChatGPT-o1 model, yet not being able to surpass it, for instance, in terms of output tokens (the single units composing a response given by the AI) or in terms of intelligence (Artificial Analysis (a)).
As DeepSeek-R1 did not come from nothing, the model retained many of the features of DeepSeek’s previous LLM, DeepSeek-V3. However, the innovations brought by the DeepSeek-R1 model concerning “reinforcement learning” and “emergent behaviour network” are particularly worth mentioning (Kerner, S.M. 2025; Deepseek 2025, 5-11). The first term pertains to machine-learning-specific terminology and generally refers to a system of trial-and-error cycles through which the AI learns and adapts to the best answer context-specific questions (Sutton, R.S., Barto, A.G. 2015, 16-20). At the same time, the second term highlights that DeepSeek-R1 is based on a network, allowing this model to develop complex reasoning patterns without ad hoc explicit programming (Kerner, S.M. 2025).
Another key aspect of DeepSeek-R1 is that it is entirely open source. Therefore, developers from around the world can use this model to create other models or build on it (Goldmann, D. 2025), as an Italian more recent clone of DeepSeek-R1 was created with just $30, a far more impressive achievement which has not shaken the U.S. tech giants (Barbera, D. 2025). Accessing the model to its full extent constitutes an advantage compared to ChatGPT models that are only partially open, such as ChatGPT-4o mini (limited requests per user) or ChatGPT-4 (available exclusively to ChatGPT Plus and a few other users). Moreover, the open-source nature aligns with the business vision of High-Flyer, which aims to develop and launch exclusively open-source models (Kerner, S.M. 2025). Notwithstanding this trait, the creation of an open-source AI model does not equate to guaranteeing its transparency as many pieces of information on how this model was trained and on the type of data used to train it remain still unknown (Goldmann, D. 2025).
The competition is on!: is DeepSeek-R1 better than its American counterparts?
Despite DeepSeek-R1’s most noticeable aspects of extremely low costs of production and open-source nature, the question concerning the perceived superiority of its metrics compared to the models present on the market in January 2025 of OpenAI remains. The following analysis focuses on OpenAI models as the company most reactive to the introduction of a more competitive model of DeepSeek-R1 in the U.S. markets. The main aspects of comparison are: the model’s intelligence, speed, and cost of the input and output tokens.
DeepSeekR1 and ChatGPT-o1
In the case of the o1 model, it can be seen that both DeepSeek-R1 and ChatGPT-o1 are reasoning models functioning through the “chain of thought” structure to solve reasoning problems (Cisternino, A. 2025). This allows engineers to improve language models on tasks concerning logic or calculation through a prompt structured in the same way a human would do (Ganesha, V., Kavlakoglu, E. 2024).
To cause such a stock market downturn, one should expect DeepSeek-R1 to produce impeccable and exceedingly fast responses. However, when compared to the ChatGPT-o1 model (released in September 2024), DeepSeek is not only less intelligent than the older ChatGPT-o1 model, but also immensely slower (Artificial Analysis (b); Artificial Analysis. 2025. State of AI: China - Q1 2025).
In other words, DeepSeek-R1 works very similarly to the o1 model while underperforming in some aspects, yet costing less in terms of both creation and functioning. Additionally, without considering the yet-to-confirm performance metrics of ChatGPT 4.5, the latest model of OpenAI, ChatGPT-o1, remains the most expensive reasoning model, amounting to $15 per 1 million input tokens and $60 per 1 million output tokens. Therefore, even when creating a new AI model without focusing on delivering a less expensive model, AI engineers are likely to produce a cheaper outcome than the o1 model (Artificial Analysis (a); Artificial Analysis. 2025. State of AI: China - Q1 2025). Moreover, the model architecture of DeepSeek-R1 uses a Mixture of Experts (MoE) type of language models, different from ChatGPT and other GPAI. The MoE language model allows for higher levels of versatility while avoiding undermining performance. Conversely, ChatGPT models employ a transformer-based architecture composed of fixed sets of parameters contributing to higher costs of token output, thus making ChatGPT comparatively more expensive when answering questions (Modular 2025).
DeepSeekR1 and ChatGPT o3-mini and o3-mini-high
Subsequently, the ChatGPT o3-mini and o3-mini-high (included in ChatGPT Plus) variants of ChatGPT-3 were officially released on January 31, a few days after the high tide of the financial shock caused by DeepSeek-R1. One can hypothesize that such a shock might have accelerated the launch of OpenAI o3-mini models to bring back the focus to OpenAI. This intention seems to be confirmed by Sam Altman’s post on X, hinting at two crucial factors: the open-source nature of the model, the same 360° open source announced for DeepSeek-R1, and its declared unmatched logical and STEM reasoning skills (Sam Altman 2025).
By analyzing the technical features of the o3-mini model and comparing them with DeepSeek-R1, it can be seen that both models are reasoning models based on a chain of thought type of reasoning, similar to the ChatGPT-o1 model. ChatGPT o3-mini was introduced as a response to the unexpected release of DeepSeek-R1, developed to answer STEM and coding requests with the possibility of changing the reasoning effort level of the o3-mini model by choosing low, medium, or high (Docsbot.AI (a)). In more general terms, while DeepSeek-R1 features are similar to those of ChatGPT-o1, the o3-mini model shows far more impressive reasoning capabilities while being significantly faster than DeepSeek-R1.
Notwithstanding the improvements of the V3 model, DeepSeek-R1 demonstrates slightly lower intelligence compared to the more advanced ChatGPT o3-mini, while also costing less in terms of output tokens ($ per 1 million tokens produced) and input tokens. Therefore, each o3-mini request sent to the chatbot and the subsequent response cost more than that of the same request to DeepSeek-R1 (Artificial Analysis (c); OpenAI. API Pricing). However, the most surprising metric of the o3-mini model within this comparison is its enormous seven-fold speed (tokens per second produced) compared with the far slower speed of the DeepSeek-R1 model (ibid.).
The o3-mini-high model presents a singular difference in terms of metrics analyzed from the o3-mini, as the o3-mini-high version has an almost equal level of intelligence, while its speed is halved compared to that of the o3-mini (Artificial Analysis (d)).
To introduce a new meter of comparison between DeepSeek-R1 and the OpenAI models, after the release of o3-mini, users reported a substantial improvement of the o3-mini in relation to ChatGPT-o1, which possesses complementary metrics to DeepSeek-R1. By using this similarity of DeepSeel-R1 to the ChatGPT-o1 model, ChatGPT o3-mini made 39% fewer significant mistakes on real-world questions while delivering conclusions 24% faster than the ChatGPT-o1 model. Therefore, comparable percentages potentially hold true even against DeepSeek-R1 (Wiggers, K. 2025).
However, it is also true that in some AI tests, such as AIME 2024 or SWE-bench Verified, o3-mini surpasses DeepSeek-R1 only with high reasoning effort, while, for example, in the GPQA Diamond (test for PhD analyses regarding the STEM fields), the DeepSeek-R1 model beats o3-mini (ibid.). In other words, both the o3-mini and the -high variant can solve PhD-level STEM problems as well as simpler questions; however, the results and efficiency levels are unclear (Mulligan, S.J. 2025). Therefore, DeepSeek-R1 reached higher levels than the o3-mini model for highly technical scientific questions.
OpenAI model diversification: ChatGPT 4.5
From this perspective, it appears that after the release of DeepSeek-R1, OpenAI moved in two opposite directions. While releasing o3-mini met the need for a more powerful AI to solve STEM, logical reasoning, and coding questions, the company updated ChatGPT-4o, the standard ChatGPT model. This second was further developed to provide more contextually accurate responses, particularly for: cultural and social phenomena, and to understand visual inputs (i.e., image uploads), diagrams, and visually driven technical problems (OpenAI 2025. Model Release Notes). A further addition to the training of this model is about improving ChatGPT-4o math and coding reasoning (ibid.). This constitutes a peculiar step, underlining that OpenAI is investing in both “rigid” reasoning models to solve real-world scientific problems and address more human-related aspects.
While focusing on the release of improved models for logical and scientific reasoning since the summer of 2024, after DeepSeek-R1 was presented as a fierce Chinese competitor completely rivalling several U.S. models, OpenAI showed its ability to invest in a wider spectrum of chatbots created for both logical, cultural, social, and non-technical humanlike purposes. This trend is confirmed by the announcement of the ChatGPT 4.5 release on February 27, 2025. Some classical reasoning and non-reasoning performance metrics of the model are different from those of the previous OpenAI models mentioned above. This non-reasoning model aims to offer users a “natural” and “human-like” experience as they are chatting with another human being. In addition, the model can develop and recognize patterns or provide creative responses without logical connections, thus requiring far more training. This partially explains why it is the most expensive model to date (Artificial Analysis (e): OpenAI 2025. Model Release Notes).
This is a surprising element, as many reasoning models tend to be more expensive than their non-reasoning counterparts (Mulligan, S. J. 2025). Moreover, ChatGPT 4.5 is said to cost $75 per million input tokens and $150 per output tokens, which, compared to the DeepSeek-R1 input price of $0.55 and output of $2.19, shows a drastic surge in price for a humanlike interaction only available with the ChatGPT Plus tier. Additionally, in terms of speed, the 4.5 model is just double the speed of the DeepSeek-R1 model and less than half of the total speed of the o3-mini (Artificial Analysis.Comparison of Models). From this analysis, OpenAI will most certainly continue to release reasoning models that improve on its top-tier model, the o3. However, it is clear that Sam Altman is ready to invest more in human-like and non-reasoning models, thus choosing to pursue alternative paths to that recently taken by DeepSeek.
The Competition is still on, but just between the U.S. and China
In the weeks after the release of ChatGPT o3-mini, several OpenAI executives, U.S. diplomats, and officials addressed this intensification of the U.S.-Chinese chatbots releases as a “very real competition”, criticising the authoritarian trait of the Chinese DeepSeek model (Browne, R. 2025). From January to February 2025, DeepSeek-R1 showed that pre-existing competition increased in terms of the quality of AI models and the rapidity of the AI models released, now becoming fully fledged.
What is new is perhaps the more direct tone between U.S. and Chinese AI companies, which have officially entered a new phase of this competition marked not only by the even faster development and release of new models but also by the social media posts of some chief executives (Browne, R. 2025). This confrontation highlighted two previously hidden key aspects: the absence of “rules” for these AI companies in developing greater models than their competitors, and the centrality of AI chatbots in this competition.
The first can be exemplified by the key decision of DeepSeek and High Flyer, the holding company owning the startup DeepSeek, to smuggle hundreds of Nvidia chips, both Nvidia H800 chips and H100 models, from the U.S. (Kam Li Yee, S. 2025). China was prevented from accessing these chips and other vital products to develop competitive AI models (i.e., dozens of different types of equipment for semiconductors), which were not available to Chinese companies appearing in an “entry list” (Iyengar, R., Lu, C. 2024). Additionally, a private consultancy firm, SemiAnalysis showed that DeepSeek might have invested more than $ 500 million in Nvidia chips, including both top-ranked models H800 and H100 (Kam Li Yee, S. 2025). Furthermore, the decision to bypass the ban set by the U.S. government goes against a 2024 statement by the Chinese Congress, which requires firms to refrain from purchasing some of the best-performing Nvidia chips (N.d. 2024).
Interestingly, Nvidia did not reject the possibility of selling its chips to China, thus acting against the intentions of both Biden and the second Trump administration (Kam Li Yee, S. 2025). Despite the clear nature of the U.S. ban on impeding AI and, inter alia, other quantum computing projects of Chinese companies, smuggling constitutes a further step forcing the strategic interests and defences of the U.S. in keeping foreign business from reaching U.S. products for the development of better AI models.
The core of the competition: chatbots
The second hidden aspect is the centrality of chatbots. As previously analyzed, an AI chatbot is one of the main ways people interact with AI, helping AI companies gather more data from users to be later used to train the AI in return. Subsequently, open-access chatbots, which do not require any sort of payment, provide these models with “fresh data” with a hidden cost: the private data of users. Such data, as for other online platforms (i.e., social media), constitute the “currency” of the future (Gates, C. Matthews, P. 2014). Despite OpenAI and DeepSeek’s models, other models of the plethora of competing chatbots currently released in the U.S., including Google, Anthropic Meta, and Elon Musk’s Grok model, function through this mechanism.
Consequently, it is important to recognize that both the U.S. and the EU must compete with several top-performing Chinese chatbots (Artificial Analysis. 2025. State of AI: China - Q1 2025, 3-4). In the last months of 2024, several Chinese AI labs managed to close the gap with the top-performing U.S. AI models, including Alibaba, Tencent, and predicted-to-lead Chinese AI tech leaders MoonShot and Zhipu (ibid. 5). To date, DeepSeek-R1 remains the best model China produced across all other Chinese tech companies (ibid. 9).
A necessary halt to the competition?
This continuing and intensified release of greater models demonstrates that both the U.S. and China are increasing the pace of developing even better AI models without leaving space for ethical-existential considerations of the future of humanity in relation to AI. In other words, this unregulated challenge might lead to irreversible outcomes, particularly in terms of the development of artificial superintelligence exiting the borders set by humans. To this extent, the Chinese ambassador to the U.S., Xie Feng, has recently called for a U.S.-China cooperation agreement on artificial intelligence, lowering the tone of the competition (Ziwen, Z. 2025). Such a possibility is alluded to by the Chinese foreign minister with the expression “Pandora’s box” (Ziwen, Z. 2025).
The issue is not the regulation of particular aspects of the models but rather the need to regulate AI development competition, potentially leading to uncontrollable Orwellian technological outcomes. Human-centric legal frameworks already exist at the EU and UN levels and could prevent such scenarios. However, these frameworks have not been fully contemplated by the main AI technology giants or by the U.S. and Chinese governments. Therefore, it appears that the rapidity and intensity of this competition have drastically shifted the possibility of a negotiation table to lower the tone of the competition because of the DeepSeek-R1 release. However, another question remains to be addressed.
Is the “Old lady” sleeping?: where is the EU positioned in this competition?
The EU developed regulatory tools and parameters with its AI Act to address concerns regarding the limits of AI chatbot development of Chinese officials and, perhaps, those of many other executives of leading AI companies through the AI Act.
However, not too long ago, the EU invested substantially in this AI competition while not creating an EU-level chatbot, leaving space for member states to develop their own. In this regard, the main European chatbot available is “Le Chat”, a chatbot released by the French AI startup “Mistral AI” on 26 February 2025 (Six, N. 2024). Mistral AI has been active since 2023 and has produced several models. This chatbot managed to accelerate output token speed, signaling higher speed in offering responses, compared to ChatGPT and DeepSeek, at a cost of $15 per month compared to the $20 of ChatGPT Plus and $200 of ChatGPT Pro.
Additionally, the AI models “Mistral Large” and “Mistral Small” used by the Le Chat allow this last to produce super-fast answers of a wide range of topics, however, without reaching the same analytical depth of the U.S. and Chinese competitors in solving logical, scientific STEM tasks (Docsbot.AI (c); Docsbot.AI (d)). Moreover, Le Chat is promising in answering complex questions, for instance, regarding European literature or technological inventions, while failing to summarise specific texts (Six, N. 2024). Generally, Mistral AI models are balanced compared to many others, thus offering a competitive solution.
Another recent EU-member chatbot, Vitruvian AI, was released by the Asc27. Vitruvian-114B is an AI model developed to support the government and private enterprises (Crescenzi, C. 2025). Unfortunately, there is insufficient data to compare this model to those previously analyzed for all key metrics.
This fragmentation of AI startups confirms the absence of an EU-level long-term plan for developing AI models and chatbots that meet the data protection needs of EU citizens and businesses ready to invest in the AI sector (Patella, A. 2025; Albizzati, A. et al. 2025: 31). Additionally, the response to the latest events described in this article saw the Von der Leyen Commission answer with a €200 billion fund for AI investments, announced at the AI Action Summit in Paris on February 11, 2025. The initiative named “InvestAI” mobilises €200 billion for general investments in the AI sector and 20 billion for a more specific objective to foster the creation of AI gigafactories aimed at developing AI models through “collaborative development” (Kroet, C. 2025).
InvestAI will constitute the largest public-private partnership in Europe for developing AI, respecting EU standards (ibid.). Notwithstanding the positive mediating impact of this fund, it must be recognized that InvestAI will be created through the economic resources already allocated to other EU programs such as the Digital Europe Programme, Horizon Europe, and InvestEU (European Commission 2025).
Although this fund is warmly accepted across EU member states as a positive message for AI developers ready to use world-class supercomputers and enormous computing infrastructures, such an introduction arrives late to this ever faster AI competition, as the gigafactories cannot be created in a short time (ibid.). Furthermore, the InvestAI fund meets the needs of the EU Competitiveness Compass, given its structure (ibid.). Notwithstanding the fact that the fund is essential for the long-term development of a structured EU AI sector, the EU has not yet released a chatbot.
Why an EU chatbot?
A crucial step not mentioned in the presentation of the InvestAI initiative is the creation of a single AI chatbot at the EU level, following EU values. Although Von der Leyen stated the need for an “AI to be a force for good and for growth,” the focus of the EU, at the moment, is not directed at the production of an AI chatbot to compete with the U.S. and Chinese opponents. An EU chatbot would bring a series of benefits that should not be ignored, especially given the rise of totalitarian and a-democratic tendencies in the international arena and the role of “tech oligarchs”.
An EU chatbot would not only have to abide by EU law and ethical parameters but also constitute a pillar within the AI chatbot competition, ensuring that EU citizens, who already use ChatGPT and other chatbots, have their private data collected, treated, processed, and used in line with the standards of three key EU secondary acts: the AI Act, the GDPR, and the DSA.
It is not by chance that the Draghi’s Report on Competitiveness highlighted in its in-depth analysis and recommendations (Part B) that the EU must keep an eye on U.S. and Asian Ai competitors, ensuring that the EU remains competitive particularly in the telecom sector and in the “digital data sovereignty” (European Commission 2024: 68-72). However, interestingly, the report does not mention a proper EU AI chatbot. Although the Report clearly states that the chatbot could be potentially integrated into businesses, improving their profitability and revenues as U.S. businesses are showing while adding further privacy and business secrets security levels than a non-EU AI chatbot could guarantee, the peril of being dependent on a foreign AI chatbot posed a huge risk for both businesses, especially EU citizens (ibid. 79).
Risks of using a non-EU AI chatbot
The lack of an EU chatbot restricts not only competitiveness for businesses but also the rights of opinion, expression, and knowledge of EU citizens. This is because of the potential direct or indirect censure that may be intentionally presented in the responses of non-EU chatbots. This possibility was found by many users asking the Chinese chatbot DeepSeek, after the release of the R1 model, questions regarding Tiananmen Square, Xi Jin Ping, the independence of Taiwan and the Uyghurs seeing just a blank screen with excuses as these topics might be too troublesome for the “core socialist values” and therefore subject to intentional distortions or to censure (Cisternino 2025).
This censure is related to a consecutive issue derived from not having an EU AI chatbot or narrative manipulation that leverages mis/disinformation. The EU must be aware of strategic narrative control by AI tech giants that own non-EU chatbots. Therefore, an EU chatbot would constitute a more reassuring solution to other chatbots, sparing the national Data Protection Authorities (DPAs) of EU member states, as in the cases of Italy and Belgium, in the need to sanction or block these chatbots by constituting a trustworthy alternative for use by EU citizens and businesses ( GPDP 2025). Perhaps the most important benefit an EU chatbot would bring is the tranquility for both EU businesses and, especially, for EU citizens in posing their questions to an AI while having their private data collected, processed, and stored in a “safe place.” This would guarantee that the GDPR data protection standards are unmatched by other national legislation applicable in the U.S. and China. When data are sent and stored in foreign servers owned by OpenAI, DeepSeek, and other AI companies owning these chatbots, no data protection guarantee reaches the European level.
Furthermore, even if OpenAI complies with the GDPR, there is no certainty of full compliance with the articles of the regulation, as the data of EU citizens are stored in the servers of OpenAI, which also have to abide by U.S. law while trying to guarantee EU data protection standards (OpenAI. Security & privacy). Additionally, ChatGPT conversations containing private data of individuals allow them to disable the training function, essentially requesting that ChatGPT not use their data for training instead of allowing ChatGPT 30 days for using these data to train its models. However, it is not specified if these data may be used for purposes other than those listed on the official OpenAI page, even when respecting the European GDPR, as OpenAI servers cannot be subject to inspections (T., L. 2024).
The same does not hold true for DeepSeek which presents several additional concerns particularly relating to the collection and processing of data, transparency requirements relating data sources to prevent disinformation or training methods of the AI required by the GDPR, DSA and AI Act (Jarovsky, L. 2025). Additionally, DeepSeek models transparency revealed that its models, particularly DeepSeek-R1 are subject to misuse as its lacks strong guardrails allowing more expert users to generate ill-intentioned content posing a threat to human rights of people or to simply leak data collected and stored by DeepSeek as the “deepleak” of personal information case of few days after the release of the DeepSeek-R1 model highlighted (Desmarais, A. 2025; Daniel, L. 2025).
Consequently, given the uncertainties surrounding the use of data stored on U.S. servers and the numerous challenges associated with employing Chinese chatbots without a strong data protection framework that ensures adherence to privacy rights, the EU might consider prioritizing the creation of an EU chatbot on its agenda.
It is also true that, up to this very moment, individuals seem to be unwilling to “opt-out” from using the main non-EU chatbots in the hope that their data are collected, processed, and used for AI training with respect to their rights, which further marks the importance of producing an EU chatbot.
Conclusions
The competition between the U.S. and China entered a new, more intense, and rapid phase since the DeepSeek-R1 release, which showed that, under certain commercial “shortcuts,” China can close the gap with the top AI producers in the U.S., reaching almost-similar levels.
In parallel, the EU has taken its first bold step in the creation of a 200 billion fund under the recent InvestAI initiative. However, the EU does not appear to have considered the production of an EU-level chatbot, allowing it to compete with the U.S. and China while protecting the private data of Europeans. The competitiveness suggestions of the Draghi Report highlight that the EU must act quickly and accelerate the creation of gigafactories while pursuing other initiatives. While not competing with an EU chatbot, the role of the EU in the AI global industry is undermined, as it induces EU citizens to use U.S. or Chinese chatbots, without a safer option despite OpenAI’s, as well as other U.S. models, ostensible compliance with the GDPR as no substantial checks of U.S. servers can be carried out. Furthermore, several data protection issues, particularly of the main Chinese AI developer company DeepSeek, show that EU citizens still sending their data are facing huge risks from simple yet effective disinformation regarding historical facts to the more discreet use of their data.
Finally, incapable of convincing foreign AI companies such as OpenAI or DeepSeek to fully ensure respect for EU legal acts, the EU must consider the production of its own chatbot. Conversely, the EU might never truly enter the competition and be capable of guaranteeing comprehensive protection of Europeans’ personal data by further relying on foreign AI models, which remain appealing to EU citizens. As Alessandro Aresu recalls in his work “Geopolitica dell’Intelligenza Artificiale,” the original intention of AI development companies remains the one outlined in the 2018 OpenAI Charter and is still centered on creating AI that benefits humanity. However, this objective does not exclude the necessity and tempting desirability of economic returns for these AI companies, which are currently prioritized over human-centric perspectives (Alessandro Aresu, 310-312). This competition demonstrates that data-driven algorithms for the functioning of chatbots and the necessity of these last human inputs to improve and understand are located within this economic necessity. To this extent, the EU, by becoming competitive, may gain a crucial role in pushing AI development in another direction, the one grounded on EU values and the broad and comprehensive guidelines of the AI Act.
BIBLIOGRAPHY
Albizzati, A. et al. 2025. The French AI Report 2024 - Key Challenges, Financing Trends, and Emerging Champions. February. pp.1-68. Available: https://thefrenchreport.ai/TheFrenchAI_Report2024.pdf
Aresu, A. Geopolitica dell’intelligenza artificiale. Feltrinelli. 22 October 2024. pp.1-576.
Artificial Analysis. 2025. State of AI: China - Q1 2025. pp.1-25. Available: https://artificialanalysis.ai/downloads/china-report/2025/Artificial-Analysis-State-of-AI-China-Q1-2025.pdf
Artificial Analysis. Comparison of Models: Intelligence, Performance & Price Analysis. Available: https://artificialanalysis.ai/models#intelligence
Artificial Analysis (a). o1: Intelligence, Performance & Price Analysis, Available: https://artificialanalysis.ai/models/o1
Artificial Analysis (c). o3-mini: Intelligence, Performance & Price Analysis, Available: https://artificialanalysis.ai/models/deepseek-r1
Artificial Analysis (b). DeepSeek-R1: Intelligence, Performance & Price Analysis, Available: https://artificialanalysis.ai/models/deepseek-r1
Artificial Analysis (d). o3-mini (high): Intelligence, Performance & Price Analysis, Available: https://artificialanalysis.ai/models/o3-mini-high
Artificial Analysis (e). GPT-4.5 (Preview): Intelligence, Performance & Price Analysis. Available: https://artificialanalysis.ai/models/gpt-4-5
Barbera, D. 2025. Il clone di DeepSeek creato con 30 dollari - L'esperimento dell'Università della California ispirato al controverso modello cinese. Wired, February 3. Available: https://www.wired.it/article/tiny-zero-clone-deepseek-30-dollari/
Brandtzaeg, P.B. et al. 2017. “Why people use chatbots”. In Internet Science, Kompatsiaris I, et al., editors. Cham: Springer, pp.377-392.
Browne, R. 2025. “‘Game on’: Tech execs say DeepSeek ramps up China-U.S. competition but won’t hurt OpenAI”. CNBC. February 17. Available: https://www.cnbc.com/2025/02/17/deepseek-ramps-up-china-us-competition-but-wont-hurt-openai.html
Carew, S. et al. 2025. DeepSeek sparks AI stock selloff; Nvidia posts record market-cap loss. Reuters, January 28. Available: https://www.reuters.com/technology/chinas-deepseek-sets-off-ai-market-rout-2025-01-27/
Cisternino, A. 2025. “DeepSeek vs Chatgpt, la nostra prova: chi vince nel confronto”. Agenda Digitale, January 30. Available: https://www.agendadigitale.eu/mercati-digitali/deepseek-r1-vs-openai-chi-vince-nel-confronto-tra-lunderdog-cinese-e-i-colossi-dellia/
Crescenzi, C. 2025. “Cos'è Vitruvian-1, l'AI italiana che compete con ChatGpt”. Wired, February 18. Available: https://www.wired.it/article/vitruvian-1-intelligenza-artificiale-italiana/
Kam Li Yee, S. 2025. “Containment Can’t Win the U.S.-China Tech Race Alone”. Foreign Policy, March 3. Available: https://foreignpolicy.com/2025/03/03/artificial-intelligence-ai-us-china-competition-deepseek-containment/
Daniel, L. 2025. “DeepSeek Data Leak Exposes 1 Million Sensitive Records”. Forbes, February 1. Available: https://www.forbes.com/sites/larsdaniel/2025/02/01/deepseek-data-leak-exposes--1000000-sensitive-records/
Deepseek. 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, pp.1-22, Available: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf
Desmarais, A. 2025. “What are the data privacy issues plaguing Chinese AI DeepSeek in the EU?”. Euronews, February 6. Available: https://www.euronews.com/next/2025/02/06/what-are-the-data-privacy-issues-plaguing-chinese-ai-deepseek-in-the-eu
Docsbot.AI (a). Compare: o3-mini vs. DeepSeek-R1. Available: https://docsbot.ai/models/compare/o3-mini/deepseek-r1
Docsbot.AI (b). Compare: Claude 3.7 Sonnet - Extended Thinking vs o3-mini. Available: https://docsbot.ai/models/compare/claude-3-7-sonnet-extended-thinking/o3-mini
Docsbot.AI (c). Compare: Mistral Large 2 vs DeepSeek-R1. Available: https://docsbot.ai/models/compare/mistral-large-2/deepseek-r1
Docsbot.AI (d). Compare: Mistral Large 2 vs o1. Available:
https://docsbot.ai/models/compare/mistral-large-2/o1#pricing
Ermey, R. 2025. The top S&P 500 stock of 2024 returned 340.5%—it wasn’t Nvidia. CNBC make it, January 13 Available: https://www.cnbc.com/2025/01/13/top-performing-stocks-of-2024.html#:~:text=Instead%2C%20the%20battle%20for%20highest,171.2%25%20lands%20it%20in%20third
European Commission. 2024. The Draghi report: In-depth analysis and recommendations (Part B). September 17. Pp.1-328. Available: https://commission.europa.eu/topics/eu-competitiveness/draghi-report_en#paragraph_47059
European Commission. 2025. EU launches InvestAI initiative to mobilise €200 billion of investment in artificial intelligence. February 11. Available: https://ec.europa.eu/commission/presscorner/detail/en/ip_25_467
Gates, C. Matthews, P. 2014. Data Is the New Currency. Research Gate, September. pp.105-116. Available: https://www.researchgate.net/publication/301467741_Data_Is_the_New_Currency
Ganesha, V., Kavlakoglu, E. 2024. “What is chain of thoughts (CoT)?”. IBM, August 12. Available: https://www.ibm.com/think/topics/chain-of-thoughts
Garante per la Protezione dei Dati Personali (GPDP). 2025. COMUNICATO STAMPA - Intelligenza artificiale: il Garante privacy blocca DeepSeek. January 30. Available: https://www.garanteprivacy.it/web/guest/home/docweb/-/docweb-display/docweb/10097450#english
Goldmann, D. 2025. What is DeepSeek, the Chinese AI startup that shook the tech world?. CNN, January 27. Available: https://edition.cnn.com/2025/01/27/tech/deepseek-ai-explainer/index.html
Investing.com. “Nvidia Corporation (NVDA)”, Available: https://www.investing.com/equities/nvidia-corp
Iyengar, R., Lu, C. 2024. “With a Final China Chip Battle”. Foreign Policy, December 5. Available: https://foreignpolicy.com/2024/12/05/us-china-trade-semiconductor-chips-gallium-germanium-export-control-ban/
Holdsworth, J. 2024. “Unlocking the power of chatbots: Key benefits for businesses and customers”. IBM Think. January 18 Available: https://www.ibm.com/think/insights/unlocking-the-power-of-chatbots-key-benefits-for-businesses-and-customers
Jarovsky, L. 2025. “DeepSeek’ Legal Pitfalls”. Luiza’s Newsletter, February 2. Available: https://www.luizasnewsletter.com/p/deepseeks-legal-pitfalls
Kennedy, B. et al. 2023. “Public Awareness of Artificial Intelligence in Everyday Activities Limited enthusiasm in U.S. over AI’s growing influence in daily life”. Pew Research Centre, February 15. Available: https://www.pewresearch.org/science/2023/02/15/public-awareness-of-artificial-intelligence-in-everyday-activities/
Kerner, S.M. 2025. DeepSeek explained: Everything you need to know, TechTarget, Available: https://www.techtarget.com/whatis/feature/DeepSeek-explained-Everything-you-need-to-know#:~:text=DeepSeek-R1.,a%20context%20length%20of%20128%2C000
Kroet, C. 2025. EU to mobilise €200 billion for AI investment. Euronews, February 11. Available: https://www.euronews.com/next/2025/02/11/eu-to-mobilise-200-billion-for-ai-investment
Modular. 2025. DeepSeek-R1 vs ChatGPT: A comparative analysis. Available: https://www.modular.com/ai-resources/deepseek-r1-vs-chatgpt-a-comparative-analysis#:~:text=DeepSeek%2DR1%20has%20demonstrated%20strengths,structured%20and%20idea%2Dfocused%20content
Mulligan, S.J. 2025. “OpenAI releases its new o3-mini reasoning model for free”. MIT Technology Review, January 31. Available: https://www.technologyreview.com/2025/01/31/1110757/openai-makes-its-reasoning-model-for-free/
N.g. 2024. “China tells firms to avoid Nvidia chips”. Taipei Times, September 30. p.12 Available: https://www.taipeitimes.com/News/biz/archives/2024/09/30/2003824544#:~:text=The%20US%20government%20banned%20Nvidia,US%20Department%20of%20Commerce%20regulations
N.g., K., Dreanon, B. et al. 2025. DeepSeek: The Chinese AI app that has the world talking. BBC News, February 4, Available: https://www.bbc.com/news/articles/c5yv5976z9po
OpenAI. 2025. Model Release Notes. Available: https://help.openai.com/en/articles/9624314-model-release-notes
O’Donnel, J. 2025. “Surging emissions, battlefield algorithms, Trump's chip war, and other predictions”. MIT Technology Review, January 14. Available: https://www.technologyreview.com/2025/01/14/1109958/whats-next-for-ai-in-2025-2/
OpenAI. API Pricing. Available: https://openai.com/api/pricing/
OpenAI. Security & privacy. Available: https://openai.com/security-and-privacy/
Patella, A. 2025. “Le startup europee di intelligenza artificiale sono sempre più interessanti”. Wired, February 5. Available: https://www.wired.it/article/startup-europa-intelligenza-artificiale-2024-finanziamenti/
Sam Altman (@sama). 2025. “o3-mini is out! smart, fast model. available in ChatGPT and API. it can search the web, and it shows its thinking. [...]”. X (post). Available: https://x.com/sama/status/1885441031656632737
Saul, D. 2025. Biggest Market Loss In History: Nvidia Stock Sheds Nearly $600 Billion As DeepSeek Shakes AI Darling. Forbes, January 27. Available: https://www.forbes.com/sites/dereksaul/2025/01/27/biggest-market-loss-in-history-nvidia-stock-sheds-nearly-600-billion-as-deepseek-shakes-ai-darling/
Six, N. 2024. “We tested Le Chat, Mistral AI's French-style ChatGPT”. Le Monde, March 1. Available: https://www.lemonde.fr/en/pixels/article/2024/03/01/we-tested-le-chat-mistral-ai-s-french-style-chatgpt_6576433_13.html
Sutton, R.S., Barto, A.G.. 2015. “Reinforcement Learning: An Introduction”. The MIT Press, Massachusetts, pp.1-328, Available: https://web.stanford.edu/class/psych209/Readings/SuttonBartoIPRLBook2ndEd.pdf
T., L. 2024. “Where Does My ChatGPT Data Go?”. RedPandas, January 2. Available: https://www.redpandas.com.au/blog/where-does-my-chatgpt-data-go/#:~:text=When%20you%20interact%20with%20ChatGPT,data%20to%20improve%20their%20models
Wiggers, K. 2025. “OpenAI launches o3-mini, its latest ‘reasoning’ model”. TechCrunch, January 31. Available: https://techcrunch.com/2025/01/31/openai-launches-o3-mini-its-latest-reasoning-model/
Ziwen, Z. 2025. “China and US need to cooperate on AI or risk ‘opening Pandora’s box’, ambassador warns”. South China Morning Post, March 2. Available: https://www.scmp.com/news/china/diplomacy/article/3300738/china-and-us-need-cooperate-ai-or-risk-opening-pandoras-box-ambassador-warns
Bình luận