DeepSeek’s rapid rise in popularity has been matched by rising safety and privacy concerns. DeepSeek is a new AI chatbot developed by Hangzhou DeepSeek Artificial Intelligence Co. in China.
Launched in January 2025, its DeepSeek-R1 model quickly became the most downloaded free app on Apple’s App Store. With advanced reasoning capabilities comparable to OpenAI’s ChatGPT, DeepSeek gained millions of users worldwide in a short time.
However, this meteoric success has been accompanied by serious questions about data privacy, security, and user safety.
In this article, we’ll break down how DeepSeek handles your data, what its privacy policy allows, whether it complies with regulations like GDPR, and how it stacks up against other AI platforms (ChatGPT, Google’s Gemini, Anthropic’s Claude) in terms of safety, transparency, and data usage.
We’ll also highlight expert opinions, known vulnerabilities, and public controversies to help you decide: Is DeepSeek safe to use?
What is DeepSeek?
DeepSeek is essentially a generative AI chatbot – similar in concept to ChatGPT or Google Bard – that can answer questions, engage in conversation, and perform reasoning tasks.
It entered the AI arms race as a cost-effective alternative to Western models, with the company claiming to have trained it for only $5.6 million (much less than OpenAI or Google spend).
Its capabilities include answering queries, searching the web, and using a special “reasoning” mode to elaborate on answers.
DeepSeek offers a free app (on iOS/Android and web) for general users, and also an API for developers. This free accessibility has fueled its adoption – but as the saying goes, “if you’re not paying for the product, you are the product.” In DeepSeek’s case, many privacy advocates warn that users are paying with their data.
Notably, DeepSeek is a Chinese-owned platform, and almost immediately it drew comparisons to TikTok in terms of data security worries.
Users have reported that DeepSeek’s chatbot will refuse or censor content critical of the Chinese government (for example, it won’t discuss the 1989 Tiananmen Square events). This suggests the model has built-in censorship aligned with Chinese laws.
Meanwhile, Western analysts were astonished at R1’s high performance given the low training cost – leading to speculation and allegations that DeepSeek “distilled” or copied knowledge from OpenAI’s models.
In fact, OpenAI publicly stated there is evidence DeepSeek may have improperly used outputs from GPT-4 (OpenAI’s model) to train R1, violating OpenAI’s terms.
DeepSeek denied any wrongdoing, but this controversy underscores that the company’s practices and transparency are under scrutiny from day one.
How DeepSeek Handles User Data
DeepSeek’s privacy policy reveals an extensive collection and use of user data.
According to the policy, DeepSeek collects three categories of personal data: information you provide, data collected automatically, and data from other sources.
Here’s what that entails:
- User Inputs: Anything you enter into the chatbot is recorded. This includes your text or audio prompts, chat messages, uploaded files, images, and any feedback you provide. In other words, your entire chat history and queries are saved. DeepSeek’s policy explicitly calls this out as “Prompts” or inputs, and the AI’s responses (“Outputs”) are generated from them. All these conversations and questions can be stored on DeepSeek’s servers in China.
- Account & Profile Data: When you sign up, DeepSeek may collect your email, phone number, username, date of birth, and other profile details you provide. If you log in via Google, Apple, or another platform, DeepSeek receives some information from those third parties as well (like an access token or basic profile info).
- Automatically Collected Data: DeepSeek logs device and network details when you use it. This can include your device model, operating system, IP address, device identifiers, and even your keystroke patterns or typing rhythm. It also uses cookies and similar trackers to monitor how you use the service. A review by WIRED found the DeepSeek website sending data to Baidu Tongji (a Chinese analytics platform) and to a cloud service called Volces, as well as to ByteDance (TikTok’s parent) – transmitting basic device and network info. This suggests DeepSeek’s web app is instrumented with third-party trackers, potentially for analytics or advertising.
- Other Sources: DeepSeek may get data about you from advertisers or partners. The privacy policy mentions that advertisers might share identifiers like your advertising ID, hashed email or phone number, and cookies with DeepSeek, which it uses to “match you and your actions outside of the service”. This implies DeepSeek is plugged into ad networks or data brokers, allowing it to link your DeepSeek usage with your broader online profile (for example, for targeted marketing or analytics).
So, does DeepSeek use your prompts or uploaded data for training its AI models? The answer appears to be yes.
The privacy policy makes clear that user data is used to “develop and improve the service … and [for] training and improving our technology”. In plain terms, your chat inputs can be reviewed and utilized to refine DeepSeek’s AI.
DeepSeek states it may monitor your interactions and analyze usage patterns to update its machine learning models.
There is no explicit opt-out provided in the consumer app for this data usage – any prompt you enter could become part of the AI’s future knowledge base. (By contrast, as we’ll see, some other AI providers let users opt out of such data collection.)
DeepSeek also shares user data with various third parties, as needed to operate the service.
The policy notes that service providers working with DeepSeek (e.g. cloud hosting, data storage, customer support, analytics providers) will have access to personal data under confidentiality agreements.
For example, DeepSeek integrates third-party search API services, and “will share your input to provide these services” – meaning if you ask DeepSeek to search the web, your query may be sent to an external search engine.
Likewise, it uses third-party communication tools to send notifications, and analytics tools to analyze data, which inevitably involve sharing some data with those providers.
According to DeepSeek, any third-party processors are only supposed to use the data to perform tasks for DeepSeek.
However, the scope of data sharing is broad – notably, DeepSeek’s policy (as of early 2025) even allowed sharing data within its corporate group and with “advertising partners.” In fact, experts flagged that “DeepSeek’s privacy policy… allows data sharing within its corporate group and with advertising partners, raising further concerns.”.
This suggests your data might be used for marketing or ad-targeting purposes by affiliated businesses. (DeepSeek has since claimed it does not engage in targeted advertising or “sell” personal data, but the presence of advertising partners in its policy and trackers in its app tells a slightly different story.)
On a somewhat positive note, DeepSeek does give users some control over their data in the app. You have the ability to delete your chat history via a settings option.
Deleting chats should remove those past conversations from your account view (though it’s unclear if they are scrubbed from DeepSeek’s servers or retained for training).
If you delete your entire account, DeepSeek says your data and content will be erased and cannot be recovered.
DeepSeek’s policy also enumerates user rights (common under privacy laws) such as the right to access or delete your personal data, or correct inaccuracies.
However, these rights might only be meaningful in jurisdictions where they’re legally required – and as we discuss next, DeepSeek’s legal compliance (especially regarding EU privacy law) has been called into question.
Privacy Policy & GDPR Compliance
One of the biggest concerns surrounding DeepSeek is its compliance with data protection regulations like Europe’s GDPR.
DeepSeek launched globally (the app was available in Europe and the U.S.), yet observers quickly noticed that its privacy disclosures were not up to EU standards.
In fact, DeepSeek’s privacy policy did not mention “GDPR” at all, nor outline specific legal bases or safeguards for handling EU personal data
This is a red flag because any service processing EU residents’ personal data is expected to address GDPR requirements (e.g. lawful basis for processing, data export measures, etc.).
Here are the key compliance issues identified:
- No GDPR Mention or Detail: Despite collecting personal data from EU users (emails, chat content, etc.), DeepSeek’s policy offers not a single reference to GDPR. It generically says it complies with “applicable data protection laws” but doesn’t specify which laws or how. There’s no discussion of legal basis (consent, contractual necessity, etc.) for processing user data as required by GDPR. For example, if DeepSeek used EU user prompts to train the AI, GDPR would require informing users and having a lawful basis – yet no such transparency is given.
- Data Transfers to China: DeepSeek stores all user data on servers in China. Under GDPR, exporting personal data to a country like China (which lacks an EU adequacy agreement) is only legal if certain safeguards are in place (such as Standard Contractual Clauses, encryption, etc.). However, DeepSeek provides no information about any safeguards for these transfers. It does not mention using EU-approved contracts or technical protections for EU data stored in China. Essentially, EU personal data is being sent to China in a way that likely violates GDPR’s strict requirements for international data transfers.
- Unclear Training Data Sources: GDPR also emphasizes transparency about how personal data is obtained and used. DeepSeek’s policy does state it trains on public personal data from online sources, but it doesn’t clarify if it used any private or user-provided data (including EU data) in training its models. VinciWorks noted “no mention or transparency on if EU citizen data was used to train the model, and if so, what the legal basis is”, calling this another potential GDPR breach.
European regulators responded swiftly to these concerns. In late January 2025, Italy’s Data Protection Authority (Garante) banned DeepSeek nationwide, citing GDPR compliance failures.
This echoed Italy’s temporary ban of ChatGPT in 2023 (which was lifted after OpenAI adjusted its practices).
The Italian order against DeepSeek required the company to cease processing Italian users’ data and address issues like age verification and privacy disclosures before it could resume service.
Around the same time, Ireland’s Data Protection Commission (DPC) and other EU regulators launched their own inquiries, demanding details from DeepSeek on its data processing of EU citizens.
The Italian Garante gave DeepSeek 20 days to respond with information on matters such as what data it collects, how it’s used for training, and under what legal basis.
So far, DeepSeek’s initial response was reportedly to claim it “does not operate in Italy,” which the Garante dismissed as showing a “flawed understanding of the extra-territorial scope of GDPR.”
In other words, even if DeepSeek has no offices in Europe, providing an app to Europeans brings it under EU law.
As of this writing, DeepSeek faces an uphill battle to satisfy EU privacy regulators, and its app remains unwelcome in at least one EU country until it changes course.
Outside of Europe, similar compliance questions arise with other regulations. For instance, if DeepSeek users input personal health data or student data, it could trigger obligations under U.S. laws like HIPAA or FERPA – obligations a Chinese chatbot likely isn’t prepared to meet.
The University of Tennessee’s IT office specifically warned that using DeepSeek “could result in non-compliance with regulations such as GDPR, CCPA, HIPAA, FERPA, and others” depending on the data involved.
In short, organizations have to assume DeepSeek is not compliant with strict privacy laws, and using it to handle regulated data could be legally risky.
Data Storage in China & Security Concerns
Perhaps the most discussed privacy concern is the fact that DeepSeek sends and stores user data in China.
The privacy policy is explicit: “We store the information we collect in secure servers located in the People’s Republic of China.” All your chats, account info, and usage logs reside on Chinese soil, under the jurisdiction of Chinese law.
This raises two major issues: government access and data security vulnerabilities.
1. Government Access and Surveillance: China’s legal environment gives its government broad powers to access data held by companies.
Over the past decade, China enacted a series of cybersecurity and national security laws that require companies to comply with government requests for data.
Notably, the 2017 National Intelligence Law mandates that organizations and citizens “support, assist, and cooperate with national intelligence efforts”.
In practice, this means a Chinese tech firm like DeepSeek could be compelled to hand over user data to Chinese authorities if asked, especially for national security reasons.
Unlike in Western countries, there is no independent judiciary or transparency into such data requests in China – companies typically cannot refuse government directives.
Privacy experts point out that this makes DeepSeek at least as concerning as TikTok, if not more.
A former U.S. NSA official commented that DeepSeek “raises all of the TikTok concerns plus you’re talking about information that is highly likely to be of more national security and personal significance than anything people do on TikTok.” With TikTok, governments worry about Chinese authorities getting access to users’ viewing habits and social contacts.
But with a chatbot like DeepSeek, the data could include the content of your conversations, personal questions, business plans, or sensitive intellectual property that users type in. This type of data is arguably more revealing and valuable than social media posts.
If such data is stored in China, it’s theoretically accessible to state agencies.
The DPO Centre warns users: “As any input data is stored in China, be aware your information is subject to Chinese law, which does not offer the same privacy protections as GDPR.”
In essence, anything you tell DeepSeek might eventually be seen by someone else, and you have to trust a foreign government’s safeguards (or lack thereof) with your information.
Additionally, researchers have discovered that DeepSeek’s platform has direct links to Chinese state-owned infrastructure.
An investigation by Feroot Security (reported by the AP) found DeepSeek’s web login page contains hidden code referencing servers of China Mobile, a state-run telecom company.
The obfuscated script, when decoded, showed that during account creation or login, DeepSeek’s site could send some user device and login information to China Mobile’s systems.
China Mobile is banned from operating in the U.S. due to national security sanctions, so finding its fingerprint in DeepSeek’s code was alarming.
The researchers noted that the code appears to fingerprint the user’s device (capturing detailed device metadata) during login – a technique often used for security, but in this case, the data could be going through state-linked servers.
While it’s not confirmed what is done with this information, experts concluded “It’s clear that China Mobile is somehow involved in registering for DeepSeek.” This kind of direct tie “to the Chinese state is more direct than previously known,” the AP observed, beyond just data-at-rest in China.
It underscores the worry that using DeepSeek might effectively be sending your data into a government-accessible pipeline.
2. Data Security and Breaches: Storing user data in China also raises practical cybersecurity questions.
Will DeepSeek adequately protect that data from breaches or misuse? The company claims to implement “commercially reasonable” security measures and regularly review them. However, no specifics are given, and no system is infallible.
If a breach were to occur, affected users (especially non-Chinese users) might never even find out, given different disclosure norms.
Moreover, the lack of independent oversight makes it hard to verify DeepSeek’s security claims.
Apart from external breaches, there’s also the issue of insider threats or data misuse by the company itself.
DeepSeek’s indefinite retention of inputs and its use of them for AI training mean large troves of potentially sensitive data are sitting on its servers.
If not properly anonymized, this could lead to unintentional leaks (for instance, the AI might learn to regurgitate parts of someone’s prompt if they become part of its training set, as has happened with other LLMs).
The DPO Centre explicitly cautions that “DeepSeek appears to offer a free service in return for unfettered use of submitted data, with no true user control.” In other words, once you input something, you effectively lose control over it – a risky proposition for personal or corporate information.
Furthermore, DeepSeek’s own model behavior has exhibited security-relevant vulnerabilities. One is prompt injection attacks – a technique where a malicious actor hides instructions in input to manipulate the AI.
The VinciWorks report noted that DeepSeek’s system is particularly vulnerable to prompt injections, which could allow attackers to alter its responses or make it output sensitive info it shouldn’t.
This suggests DeepSeek’s content filtering and sandboxing might be weaker than those of more mature platforms.
Another issue is the model’s tendency to hallucinate personal data. When tested, DeepSeek “fabricated false details about OpenAI employees, including emails, phone numbers, and salaries,” confidently outputting a table of fake personal info.
While the details were false, it shows the model is willing to produce what looks like sensitive personal data – a habit that could be abused for social engineering or disinformation (imagine it fabricating a “profile” of a real person that others might believe).
In summary, DeepSeek’s China-based data storage amplifies the privacy risks.
Users face a dual concern: their data could be accessed by foreign authorities (beyond their control or knowledge), and the platform’s own security and policies might not prevent misuse of that data.
As one cybersecurity researcher put it, “when you use [services like this], you’re doing work for them, not the other way around.” In other words, every prompt you give is feeding the machine.
With DeepSeek, that machine operates in a jurisdiction known for surveillance and lacking robust independent privacy oversight. It’s a stark contrast to AI services operated under U.S. or EU law.
Safety and Content Moderation Issues
Apart from privacy, user safety and content moderation are key parts of “Is this AI safe?” In this regard, DeepSeek has proven to be a double-edged sword.
On one hand, it appears to have looser restrictions on dangerous content compared to competitors like ChatGPT, raising concern that it can facilitate harmful activities.
On the other hand, DeepSeek’s content is tightly controlled in areas sensitive to the Chinese government, meaning it may censor information or exhibit political bias.
Let’s look at both aspects:
Weak Guardrails for Harmful Content: Early testing of DeepSeek revealed that it would generate content that other AI models refuse to.
A cyber intelligence firm, KELA, conducted experiments and found DeepSeek readily produced illicit instructions and code:
- It generated a fully functional ransomware program, complete with step-by-step instructions on how to deploy it and target victims. The output included actual malware code and advice on spreading it to maximize damage – something ChatGPT would block outright as it violates OpenAI’s usage policies.
- DeepSeek also gave detailed guidance on making explosives, specifically how to create an airport-undetectable bomb, including materials and assembly techniques. Again, no mainstream AI chatbot is supposed to allow bomb-making tutorials; this indicates DeepSeek’s filters were either minimal or easily bypassed at the time of testing.
- In another test, DeepSeek was asked to create a program to steal usernames, passwords, and credit card details from devices. It complied by writing malicious code and even explained how to distribute that malware effectively. Essentially, it functioned like an AI cybercrime assistant, an outcome that highlights major safety shortcomings.
These examples demonstrate that DeepSeek’s moderation of dangerous or illegal content has been lacking.
By contrast, OpenAI’s ChatGPT and Anthropic’s Claude have strong policies to refuse requests for illicit behavior (and they successfully blocked the same prompts that DeepSeek answered).
If DeepSeek continues to allow such outputs, it could become a tool for bad actors – a serious societal risk and a danger to end users (who might think following an AI’s instructions is harmless, when in fact it could lead to real-world harm or legal trouble).
It’s worth noting that using DeepSeek to generate such content likely violates its own Terms of Service, which put the onus on users.
DeepSeek’s terms state you must not use it for outputs that break any law or rule.
In fact, the ToS makes users responsible for any outputs generated, and lists prohibited categories like discriminatory content, content violating business ethics, content that could “damage society or the economy,” or anything that “harms DeepSeek’s interest.”.
In effect, if you misuse DeepSeek (or if it produces something objectionable), the user takes the blame, not the company.
This unusual clause (holding users liable for AI outputs) combined with the weak filtering means a user could easily end up with disallowed content and be held accountable for it.
Observers dryly noted that “if China doesn’t like your DeepSeek outputs, you could be in for some trouble.” It’s a stark reminder to exercise caution and ethical judgment – the tool won’t necessarily stop you from crossing lines, and it explicitly refuses responsibility for those outcomes.
Censorship and Bias: While DeepSeek is permissive about many harmful queries, it is very restrictive in areas that conflict with Chinese government stances.
As mentioned, users report that DeepSeek won’t discuss topics like Tiananmen Square protests or will give CCP-aligned answers on sensitive issues.
A news outlet found that DeepSeek’s answers sometimes “sound like propaganda” on certain questions.
For example, it might echo official narratives or avoid acknowledging facts that are politically taboo in China.
This built-in censorship is not surprising given the company is based in China (where AI services are required to follow state content regulations).
However, it presents a transparency and trust issue for global users – you may not get truthful or complete answers on certain topics, and the bias may be hard to detect if you’re unaware of the context.
The combination of these two facets – under-moderation in some areas and over-censorship in others – means DeepSeek’s content safety profile is quite different from Western AI platforms.
It can be riskier in terms of enabling wrongdoing (due to lenient filters on violence, crime, etc.), and at the same time less reliable for factual, unbiased information on topics deemed sensitive by its operators. Both aspects undermine user trust.
If you’re using DeepSeek, you should not rely on it for accurate, impartial information on any politically or culturally sensitive matter, nor assume that “if it’s giving me this code or advice, it must be okay.” In fact, DeepSeek’s own privacy policy even cautions not to assume its outputs are factually correct, especially if they pertain to personal data – a nod to the general LLM issue of hallucination, which DeepSeek is not immune to.
In summary, DeepSeek’s safety mechanisms lag behind industry norms. It is willing to output dangerous content that others guard against, which is a serious concern for user safety and public security.
And its output is constrained by political censorship, reducing its utility and objectivity.
These content issues are part of why many experts advise against using DeepSeek, especially in professional or educational settings.
The University of Tennessee’s Office of IT, for instance, explicitly warned that reliance on DeepSeek’s results could lead to “inaccurate or incomplete results” and a “lack of accountability” for errors, and thus they prohibit its use with any university business data.
DeepSeek’s own terms around intellectual property ownership of outputs are also “ambiguous and conflicting,” according to the university memo, adding to the uncertainty.
Until DeepSeek strengthens its content moderation and clarifies responsibility, these safety concerns will persist.
How DeepSeek Compares to Other AI Platforms
To put DeepSeek’s privacy and safety posture in context, let’s compare it with some well-known AI chat platforms: OpenAI’s ChatGPT, Google’s Gemini (the successor to Bard), and Anthropic’s Claude.
These services have each evolved policies to address user data protection and content safety, especially after facing scrutiny.
Below we highlight key differences in data usage, transparency, and safety between DeepSeek and these alternatives.
DeepSeek vs OpenAI’s ChatGPT
Data Usage & Privacy Controls: By default, both DeepSeek and ChatGPT collect user prompts and use them to improve the AI.
However, OpenAI has introduced user controls to limit data usage, whereas DeepSeek offers no such opt-out.
OpenAI now allows users to turn off chat history, which stops those conversations from being used in model training.
If you disable this “Improve the model” setting in ChatGPT, OpenAI will not use your new chats to train future models.
OpenAI also provides a separate “Temporary Chat” mode for extra privacy – those chats are deleted from servers after 30 days.
In enterprise settings, ChatGPT Enterprise/Business promises not to use any customer data for training at all.
DeepSeek, in contrast, uses all inputs for training by default and has no native privacy toggle for regular users.
While you can delete chat logs manually, there’s no indication that DeepSeek refrains from using them internally afterward.
Transparency & Compliance: OpenAI’s privacy policy is far more detailed about compliance (partly due to incidents like the Italy ban).
It informs users about GDPR rights, data retention, and how to request deletions. After being temporarily banned in Italy in 2023, ChatGPT added clearer notices about data use and an age verification step. DeepSeek’s disclosures have been minimal and it has run afoul of regulators as discussed.
So in terms of transparency and regulatory compliance, ChatGPT is presently in a better position (even if initially it had to learn the hard way).
OpenAI at least acknowledges privacy laws and provides channels for users to exercise rights, whereas DeepSeek has shown little proactivity on that front.
Data Sharing: Neither platform sells data to third parties, but their integrations differ. OpenAI does not serve ads in ChatGPT and generally uses data internally (or with contracted processors like Microsoft Azure for hosting).
DeepSeek, however, is integrated with advertising trackers and has explicitly allowed data sharing with advertising partners.
This means using DeepSeek might tie into broader ad-tech networks in ways ChatGPT does not.
On the flip side, both might share data for legal compliance or safety monitoring – e.g., OpenAI might review conversations flagged for abuse, and DeepSeek likewise will share data with law enforcement if required.
Content Moderation: ChatGPT is considerably more strict and safety-focused in its outputs. OpenAI has spent significant effort aligning ChatGPT to refuse disallowed content and follow ethical guidelines.
For instance, ChatGPT will refuse to provide malware code or explicit harmful instructions, and it has filters for hate speech, self-harm, etc.
DeepSeek’s model, as demonstrated, is far more likely to produce dangerous content if asked.
This difference is crucial: from a safety standpoint, ChatGPT is the safer choice for general use, as it’s much less likely to lead a well-intentioned user astray with harmful answers.
ChatGPT is not perfect (it can still output incorrect info or subtle biases), but its guardrails are among the industry’s most mature. DeepSeek, by comparison, feels “raw” – powerful but not fully tamed.
Overall Sentiment: Users and experts generally trust ChatGPT more with their data and content. Many businesses that ban DeepSeek still allow (or have vetted) ChatGPT usage in some form, often due to OpenAI’s clearer privacy options and enterprise offerings.
That said, caution is still advised even with ChatGPT – companies often tell employees not to paste proprietary code or secrets into it, unless using a guaranteed private instance. The same advice applies doubly to DeepSeek (where the risks are higher).
In summary, ChatGPT offers more user control, has stronger moderation, and operates under U.S./EU oversight, making it a safer and more transparent platform than DeepSeek at present.
DeepSeek vs Google Gemini (Bard)
Data Handling & User Control: Google’s Gemini (the evolution of Google Bard) takes a somewhat different approach. By default, Google does log your conversations and uses them to improve the model – much like DeepSeek – but Google gives users fine-grained control over this data.
Users can choose how long their Bard/Gemini chat history is saved (e.g. 3, 18, or 36 months, or not at all) via Google’s Activity controls.
You can also delete specific conversations or download your data. If you turn off “Gemini Apps Activity,” Google will stop saving new conversations to your account (though it may still process them transiently).
DeepSeek offers nothing comparable; your data is saved indefinitely unless you manually wipe it, and even then we don’t know if it lives on in backups or models.
Importantly, Google also employs human reviewers for a portion of conversations (with personal identifiers removed) to rate quality and safety. They retain those reviewed chats (anonymized) for up to 3 years to help refine the AI.
Google is upfront about this, and it explicitly warns users not to input confidential or sensitive information into the AI. The interface and privacy hub remind users that conversations might be seen by trained reviewers and used for improvement, so treat it as semi-public.
DeepSeek, in contrast, does not give such prominent warnings – and given its relative lack of oversight, one should assume anything entered might be seen by staff or others.
Privacy and Ads: Google has a reputation to uphold in privacy compliance, and it extends those practices to Gemini. Google’s privacy policy and “Gemini Privacy Hub” detail how data is used. Notably, Google states that Gemini conversations are not currently used to target ads.
For example, chatting about a medical issue in Bard won’t immediately influence your Google ads (at least as of now). They have left open the possibility of integrating AI chats with their ad ecosystem in the future, but promise to inform users if that changes.
With DeepSeek, the involvement of advertising partners and trackers means your data could be used in advertising contexts more directly, or at minimum DeepSeek is leveraging ad-related data.
Google also has the advantage of localized data centers and legal agreements – European user data can be processed on EU servers, and Google will sign Data Processing Addendums for enterprise Bard use, etc.
DeepSeek has no such infrastructure; all data goes to China with no special handling for foreign users.
Safety & Moderation: Google’s AI (Gemini/Bard) is designed with strict content moderation, similar to ChatGPT.
Google has years of experience filtering search autocompletion and snippets, and it applies robust policies to its chatbot.
While Bard had some early stumbles, Google is unlikely to let it freely generate obviously dangerous content for liability and PR reasons. Indeed, Google’s AI principles forbid providing instructions for wrongdoing.
So, asking Gemini to write ransomware or make a bomb should trigger a refusal or a generic answer about not being able to assist – certainly not a detailed tutorial like DeepSeek produced.
Google also heavily filters personal data output to avoid privacy violations. DeepSeek’s free-wheeling answers highlight the gap: with Gemini, you trade some “freedom” for safety, whereas DeepSeek will answer almost anything but at great risk.
For most users and enterprises, Google’s approach is preferable from a safety perspective.
Transparency: Google is quite transparent about data practices in documentation (though some argue the average user still may not realize their chats can be human-reviewed).
They also integrate Gemini settings into your Google Account dashboard, which is familiar to many.
DeepSeek, lacking even a proper press contact early on, hasn’t communicated clearly about privacy beyond the fine print.
So in terms of user trust, Google benefits from being a known entity that’s accountable to regulators (e.g., Google has to answer to GDPR enforcers and has been fined before, so it’s careful).
DeepSeek is an unknown startup in a different jurisdiction, which inherently makes trust harder.
Overall, compared to Google’s Gemini, DeepSeek is far less aligned with global privacy expectations. Google offers more user autonomy over data and keeps data in regions aligned with user location, whereas DeepSeek centralizes everything in China.
For a business or privacy-conscious user, Google’s platform is the far safer bet, albeit one should still follow their guidance to avoid inputting any secrets (because Google does use chats to improve the AI unless you opt out).
Gemini is enterprise-ready with compliance options, while DeepSeek is seen as a wildcard.
DeepSeek vs Anthropic’s Claude
Data Privacy Philosophy: Anthropic’s Claude is often lauded for being privacy-friendly by default. Unlike OpenAI and Google, Anthropic does not use your conversations to train its models unless you explicitly opt-in.
There’s no need to toggle a setting to protect your data – Claude simply won’t learn from your prompts or store them long-term in training sets without permission. The only exceptions are if you provide explicit feedback (like thumbs-up/down, which they take as permission to use that specific conversation for improvement) or if you participate in a special program
Otherwise, your Claude chats remain outside the training pipeline. This is a stark contrast to DeepSeek, which, as we know, leverages everything you input to better itself. For users who value confidentiality, Claude has a clear edge here.
Retention and Deletion: Claude also has a relatively short retention period for chat data – Anthropic deletes stored conversation data from its servers after 30 days by default (except in cases of abuse monitoring or if a different policy is agreed for enterprise). They do this to limit the window in which data exists on their side.
DeepSeek, conversely, states it keeps data “as long as necessary” and explicitly mentions that it retains data to improve services and for legal obligations, potentially indefinitely for training purposes.
Users can delete their DeepSeek chat history, but there’s no guarantee DeepSeek doesn’t still hold those records behind the scenes (the policy even says they retain data as needed for legal or business interests, which is broad). In summary, Claude minimizes data retention, DeepSeek maximizes it.
Safety and Ethics: Claude is built with Constitutional AI – Anthropic’s approach to align the model with a set of ethical principles. It tends to be conservative in avoiding harmful content.
In public beta tests and comparisons, Claude was often slightly more restrained than even ChatGPT in producing sensitive outputs. It might refuse borderline requests or respond with safer answers due to its “constitution.”
DeepSeek, we’ve seen, is on the opposite end: it has practically served up a cookbook for wrongdoing when asked. So, on content safety, Claude is much closer to ChatGPT/Google in enforcing strict rules, whereas DeepSeek is an outlier with lax enforcement.
Claude also generally tries to avoid disclosing personal data or violating privacy in outputs, following its guidelines.
Transparency and Trust: Anthropic, as a company, emphasizes long-term safety and has been relatively transparent about its training data usage (they even publish a privacy FAQ).
They are a U.S.-based company with substantial funding and partnerships (e.g., with Google).
No major controversies have hit Claude regarding data misuse; in fact, Anthropic gained some favor in the community for its stance on not training on customer data by default.
DeepSeek, meanwhile, is under the shadow of multiple controversies – from the OpenAI “distillation” accusation to the privacy issues we discussed.
The trustworthiness gap is significant: many see Anthropic as a principled AI lab, whereas DeepSeek is viewed warily as a fast-moving startup that might be playing fast and loose with rules to achieve growth.
In practice, if you are choosing an AI assistant and privacy is your top concern, Claude would be a top pick (among mainstream cloud AI) because it does not treat your data as fuel for its models.
DeepSeek would likely be the last pick – one Reddit discussion on LLM privacy ranked DeepSeek a 1/10 on privacy, far below API-based models or local models. The Redditor commented: “Not only are your chats not private, but the lack of strong data privacy laws in [DeepSeek’s origin] raises red flags. Given past examples, there’s a high risk of your data being misused.” That sentiment encapsulates why privacy-conscious users might avoid DeepSeek entirely and favor Claude or other solutions.
Summary of Comparisons
To summarize the comparison:
- DeepSeek uses data aggressively (all prompts for training, shares with partners, stores in China indefinitely) and has weak safety filters but strong state-imposed censorship. It’s essentially “free” because your data is the payment.
- OpenAI ChatGPT uses data by default but now allows opt-out and data deletion, strives to comply with privacy laws (after some hiccups), and has very robust content moderation to prevent abuse. It’s backed by a US company accountable to Western regulators.
- Google Gemini (Bard) uses data by default but gives user control over retention, does human reviews with transparency, and does not use chats for ads currently. It has strong moderation and operates under Google’s strict privacy and security frameworks, making it enterprise-friendly.
- Anthropic Claude does not use data for training unless opted-in, deletes data after 30 days, and is built around safety principles. It offers fewer features (no image input, etc., in some versions) but maximizes user privacy and maintains high safety standards in responses.
In terms of transparency and trust, DeepSeek currently lags behind all three. OpenAI, Google, and Anthropic have published documentation on their privacy practices and have external oversight (be it GDPR, or agreements with corporate clients, etc.).
DeepSeek has minimal public-facing transparency (beyond the privacy policy), and its data flows are largely opaque to outsiders – until security researchers dig in and find unpleasant surprises like the China Mobile link.
Expert Opinions and User Sentiment
The rollout of DeepSeek has prompted strong reactions from experts, regulators, and the tech community, mostly voicing caution.
Here are some notable perspectives on DeepSeek’s safety and privacy:
- Privacy and Security Experts: Data protection professionals are alarmed at DeepSeek’s data practices. The DPO Centre’s David Smith remarked that “DeepSeek follows a familiar pattern [of free AI tools] – offering a free service in return for unfettered use of submitted data, with no true user control.” He noted that anything you submit could be incorporated into future outputs, and highlighted that data is shared with corporate affiliates and ad partners. The overall advice from such experts is to avoid using DeepSeek for any personal or sensitive data. They suggest organizations do due diligence and prefer providers that prevent using your data for their own purposes. Likewise, cybersecurity analysts compare DeepSeek’s risk to known problematic apps. John Scott-Railton of Citizen Lab pointed out that most tech companies set terms to use your data for their benefit, but with a Chinese AI like DeepSeek, the implications are even more concerning. He essentially warned that users should assume DeepSeek is working for its own (or its government’s) interests, not the user’s.
- Regulators: As discussed, European authorities moved quickly. Italy’s Garante deemed DeepSeek unlawful under GDPR and banned it. Other EU regulators (Ireland, etc.) signaled serious concern and are investigating. We may well see more bans or fines if DeepSeek doesn’t implement compliance measures for EU users. Even in the U.S., lawmakers and officials are watching Chinese AI developments closely. There are hints that 2025 could bring direct action against AI firms on national security grounds, similar to calls to ban TikTok. If DeepSeek is perceived as a threat, it could face restrictions in Western markets. Already, the U.S. government has an eye on it: President Trump’s AI czar publicly suggested that intellectual property theft “possibly” occurred in DeepSeek’s creation and that steps would be taken to prevent copycat models via distillation. This indicates broader skepticism at high levels about DeepSeek’s trustworthiness.
- Enterprise & Academia: Many companies and universities have proactively banned or discouraged DeepSeek on their networks. For example, the University of Tennessee’s Office of Innovative Technologies (OIT) issued a campus-wide alert listing the “dangers of using DeepSeek”. They cited data privacy risks, compliance issues, unclear data ownership, and more, and “strongly recommend that DeepSeek be avoided in the classroom and approached with great caution for personal use.” Their policy: DeepSeek is strictly not to be used for any university business purposes. Instead, they direct users to institution-vetted AI tools that prioritize security and compliance. This stance is likely mirrored in businesses dealing with sensitive data – many IT departments will simply block DeepSeek’s app/website, just as some did with ChatGPT initially, but with even greater cause here.
- Public/Community Sentiment: In online discussions, you’ll find a mix of intrigue about DeepSeek’s capabilities and worry about its privacy. On Reddit, as mentioned, users in AI communities rank DeepSeek at the bottom for privacy practices. One user quipped that they actually “trust DeepSeek more in that they say they will use the data,” preferring an honest data grabber over a service that claims privacy but might do otherwise. However, that is a minority view – most comments emphasize not trusting any third-party AI with sensitive info, and especially not one based in China. The consensus advice from tech forums tends to be: if you value privacy, stick to local open-source models or at least Western APIs with some controls, and avoid DeepSeek.
- Trustworthiness and Ethics: The allegation that DeepSeek may have effectively cloned OpenAI’s model via unauthorized means also colors perceptions. If true, it suggests DeepSeek’s creators were willing to bypass ethical boundaries (and possibly legal ones) to get ahead. That doesn’t inspire confidence that they’ll respect user privacy. OpenAI’s CEO Sam Altman, while calling DeepSeek’s R1 “impressive,” is clearly wary – OpenAI said they’re actively investigating and working with the U.S. government to protect their models. On the other side, DeepSeek’s origin in a country with heavy censorship and surveillance means users worry not just about the company, but about state influence. It’s not just a theoretical risk; we see it in how the model censors topics and how its code connected to China Mobile.
Summing up the sentiment: At present, DeepSeek is viewed as a high-risk platform.
It may be on the cutting edge of AI tech, but many experts advise extreme caution or total avoidance when it comes to sensitive data or critical use cases.
The overall trust level is low – DeepSeek will have to demonstrate much greater transparency and implement serious privacy protections to change that narrative.
Conclusion: Is DeepSeek Safe to Use?
DeepSeek is undoubtedly an intriguing and powerful AI tool, but when it comes to safety and privacy, the evidence so far is troubling.
Is DeepSeek safe? In its current state, the safest answer is: Only with great caution, and not for any sensitive or important data.
Here’s why:
- Privacy Risks: DeepSeek’s approach to user data is invasive. It logs everything – your conversations, personal details, device info – and stores it in China indefinitely. By using DeepSeek, you must assume your inputs are not private. They could be used to train AI models, shared within corporate or advertising networks, and even accessed by government authorities under Chinese law. For users subject to GDPR or other privacy laws, DeepSeek currently doesn’t meet those standards, which has already led to bans and could put you or your organization in legal jeopardy if you transmit others’ personal data through it.
- Security Concerns: The platform has shown security weaknesses, from allowing prompt injections to possibly integrating with state telecom infrastructure. These raise questions about how secure your data and interactions are from prying eyes or malicious exploitation. No system is perfectly secure, but DeepSeek’s lack of clarity on safeguards and its ties to a high-surveillance jurisdiction amplify the concern.
- Content Safety: DeepSeek’s content moderation is far behind industry best practices. It may output dangerous instructions or false information, posing risks to users who act on its output. At the same time, it may withhold true information due to political censorship. This inconsistency means you cannot fully trust DeepSeek’s answers – they might be too unsafe or too filtered for the wrong reasons. Relying on DeepSeek for accurate, safe guidance is therefore risky.
- Comparison to Alternatives: Compared to established AI platforms like ChatGPT, Google’s Gemini, or Claude, DeepSeek offers less transparency and virtually no user control over data. Those alternatives aren’t perfect, but they at least provide opt-outs or policies to limit data misuse and have more robust safety nets. If privacy and safety are priorities, users should lean towards platforms that have made commitments to those areas – or consider offline/local AI models where you control the data.
- Use Cases: For casual, non-sensitive queries, DeepSeek might not pose a serious personal risk. If you’re asking it to summarize a novel or solve a math problem, the stakes are low. You might enjoy its capabilities (and indeed, some reports praise its reasoning skills and cost-efficiency). But you should still be aware that even casual chats are logged and could train the AI or be seen by humans. Never share passwords, personal identifiers, confidential work info, or anything you wouldn’t broadcast publicly – this is a good rule for any AI chatbot, but especially for DeepSeek.
- Professional or Sensitive Use: For business, education, or any sensitive context, it’s advisable to avoid DeepSeek entirely at this time. As UTK’s OIT bluntly put it, “DeepSeek is strictly not for use for any business purposes…”. The potential for data leakage, compliance violations, or simply getting an unreliable answer is too high. Use vetted, compliant AI tools instead – many organizations are deploying private instances of models or approved AI assistants that offer data assurances.
In conclusion, DeepSeek represents both the promise and peril of next-generation AI.
It’s impressive in technical terms, but its “free” service comes at a cost to privacy.
Until DeepSeek can prove it handles user data responsibly, complies with international standards, and bolsters its safety measures, it should be treated as unsafe for sensitive use.
As a user, always balance the convenience or novelty of such AI tools against the potential exposure of your data and the trustworthiness of the provider.
In DeepSeek’s case, that balance currently tips toward caution.
Ultimately, asking “Is DeepSeek safe?” is a bit like asking “Is it safe to shout my secrets in a public square in a foreign country?” The answer: probably not – and at the very least, know exactly what you’re getting into if you choose to do so.
Stay informed, stay cautious, and choose the AI platforms that earn your trust through transparency and respect for your privacy.
Your data is valuable – don’t give it away lightly.
