Tech & AI, Explained

deepseek

9 min read

🇨🇳 China · DeepSeek — ⛔ DO NOT USE

Status: 🟩 COMPLETE 🟦 LIVING Verdict: ⛔ AVOID — use Claude / ChatGPT / Gemini / Cursor instead. Full reasoning below.

â›” This entry exists to explain what DeepSeek is and why the encyclopedia recommends against using it. If you only want the bottom line, skip to vendors-chinese-avoid.md.

Front-matter facts

FieldValue
VendorDeepSeek (Hangzhou, Zhejiang, China) — a subsidiary of High-Flyer, a Chinese hedge fund
Country / origin🇨🇳 China — mainland PRC
Recommended for Australian users?⛔ NO — banned on Australian government devices; not recommended for personal or business use
Why we recommend againstMainland-China-based · Subject to PRC Cybersecurity Law, Data Security Law, National Intelligence Law · Inputs retained and accessible to Chinese authorities · Outputs politically filtered per CAC regulations · No meaningful privacy guarantees vs Western alternatives
Western alternatives by capabilityChat → Claude 🟩 🟦 or ChatGPT 🟩 🟦 or Gemini 🟩 🟦 · Code → Claude Code 🟩 🟦 or Cursor 🟥 or GitHub Copilot 🟥 · Reasoning → Claude Opus or GPT-5 Pro or Gemini 3 Deep Think
First releasedNovember 2023 (DeepSeek Coder); v3 January 2025 made global headlines
Last reviewed2026-06-25
Official site(not linking — pick a Western alternative)

What it is (factually)

DeepSeek is a Chinese AI research lab and consumer product. It’s a subsidiary of High-Flyer, a quantitative hedge fund founded in 2015 by Liang Wenfeng. The lab released two model families that gained Western attention:

  • DeepSeek-V3 (December 2024) — a general-purpose chat / reasoning model claimed to be trained on a smaller hardware budget than US competitors (the “trained for US$5.6M” claim that drove the January 2025 market-stir)
  • DeepSeek-R1 (January 2025) — a reasoning model competitive with OpenAI’s o1 on some benchmarks

Both are released as open-weight models — anyone can download the model files and run them on their own hardware or on a Western cloud — alongside a consumer chat at chat.deepseek.com and a paid API.

The consumer apps and direct API route your data to servers in mainland China. The open weights, when run on Western infrastructure (Together AI, Fireworks AI, AWS Bedrock, Hugging Face Inference, locally via Ollama), do not — but the political-filtering baked into the weights remains.

Why we recommend against it

The full reasoning is in vendors-chinese-avoid.md — DeepSeek is the canonical example of why that guidance exists. The short form:

1. Your data goes to China

If you use chat.deepseek.com, the mobile app, or the DeepSeek API directly, your inputs are transmitted to and stored on servers in mainland China operated by a Chinese company. This is true even with an English UI; UI language is unrelated to server location.

2. Chinese law gives the government access

Under the PRC Cybersecurity Law (2017), Data Security Law (2021), and National Intelligence Law (2017), DeepSeek can be compelled — secretly — to hand any data to Chinese state agencies. It cannot refuse and cannot tell you it received the request. There is no Western-style court-supervised process; there is no transparency report.

3. Outputs are politically filtered

The Cyberspace Administration of China requires Chinese generative AI services to align outputs with “socialist core values.” Independent testing has confirmed DeepSeek refuses or evades on:

  • Tiananmen Square 1989
  • Taiwan’s status / Taiwan as a country
  • Tibet, Xinjiang, the Uyghur situation
  • Hong Kong’s National Security Law and protests
  • Falun Gong
  • Xi Jinping’s personal life or any criticism
  • The South China Sea
  • COVID origins (when framed unfavourably to the PRC)
  • Russia-Ukraine framing
  • Various other politically sensitive Western topics

Even if you don’t ask about politics, this training affects how the model handles adjacent technical and historical questions.

4. Privacy guarantees are weaker

DeepSeek’s privacy policy (in Chinese, with selective English translation) is meaningfully weaker than the major Western providers’:

  • Limited data-deletion rights (you can request deletion; verification depends on company policy)
  • No SOC 2, ISO 27001, HIPAA, or GDPR audit certifications meaningful in the Western sense
  • Mobile apps request more device permissions than Western equivalents (contacts, device fingerprinting)
  • The DeepSeek mobile app was found in February 2025 by Wiz Research to have an exposed ClickHouse database leaking chat history, API keys and back-end logs — a real-world security failure of the kind Western frontier-lab products have not had at this scale

5. You don’t gain anything you can’t get elsewhere

DeepSeek-V3 and R1 are genuinely strong models on some benchmarks. So are Claude Sonnet 4.6, GPT-5, Gemini 3, Mistral, Llama, Qwen-hosted-by-Western-cloud. There is no DeepSeek-specific capability you can’t get from a Western provider without the privacy and political-filtering downsides.

6. Real-world bans tell the story

  • Australia banned DeepSeek from all federal government systems on 4 February 2025 (Department of Home Affairs Protective Security Policy Framework Direction 001-2025).
  • The US Department of Defense banned it from internal use.
  • Italy’s data-protection authority ordered DeepSeek to suspend processing of Italian user data in late January 2025.
  • South Korea removed it from app stores.
  • Taiwan, India, Belgium, Canada (federal), the Netherlands, Germany, and others issued formal advisories against government / sensitive use.
  • Many ASX-listed companies and AUS universities followed with their own bans.

When government CIOs and data-protection regulators across the Western world simultaneously ban a tool, the prudent personal and business decision is to follow their lead.

What to use instead

If you want to use DeepSeek for…Use this Western alternative
General chat / writingClaude.ai 🟩 🟦 · ChatGPT 🟩 🟦 · Gemini 🟩 🟦 — all free tiers excellent
Cheap powerful reasoningOpenAI GPT-5 mini or GPT-5 nano · Google Gemini 3 Flash · Anthropic Haiku 4.5 — all cost less than DeepSeek V3 at comparable quality
Open-weight reasoning modelMeta Llama 4 / 5 · OpenAI gpt-oss (open weights released 2025) · Google Gemma 3 · Mistral Mistral Large — all open-weight, all Western, no political-filter training
CodingClaude Code 🟩 🟦 · Cursor 🟥 · GitHub Copilot 🟥 · Codestral 🟥 (Mistral)
“I just want it cheap”OpenRouter (some Western models free), Groq (fast cheap inference of Llama models), Together AI (cheap open-weight inference) — all Western infrastructure
”I want to play with DeepSeek’s open weights for research”Use the open weights via Together AI, Fireworks AI, Hugging Face Inference, or AWS Bedrock instead of DeepSeek’s own API. The political-filtering training is still baked into the weights, but at least your data doesn’t go to China.

”But the open weights are okay, right?”

This is the most nuanced part of the DeepSeek story. Open weights means anyone can download the model files (the trained neural-network parameters) and run them on their own hardware.

If you download DeepSeek-V3 weights and run them locally with Ollama on your laptop, OR host them via a Western cloud (Together AI, Fireworks AI, Hugging Face, AWS Bedrock, Groq), then:

  • âś… Your data does NOT go to China
  • âś… You’re not subject to PRC data-handover law via the inference
  • ❌ The political-filtering training is still baked into the weights themselves — running locally doesn’t remove it
  • ❌ You cannot inspect the 14+ trillion tokens of training data DeepSeek used
  • ❌ Subtle research findings have shown open-weight Chinese models sometimes produce slightly different code on sensitive topics (insecure cryptography defaults in niche cases) — this is an active research area, not a confirmed exploit, but worth knowing

The encyclopedia’s recommendation: even for open-weight use, prefer Western open-weight alternatives. Meta’s Llama 4 / 5, Mistral’s models, Google’s Gemma 3, and OpenAI’s gpt-oss releases cover the same use cases without the training-trust questions.

See open-weights-vs-closed.md 🟥 for the longer treatment.

”I’m a researcher and need to study it — how to do it safely”

If you genuinely need to evaluate DeepSeek (writing a comparison article, doing AI safety research, evaluating competitive products) and not just casually using it:

  1. Never sign up with your real email. Use a disposable address (SimpleLogin, Fastmail aliases, addy.io).
  2. Never paste real work, code, customer data, or anything sensitive. Synthetic test prompts only.
  3. Don’t install the mobile app on your personal phone. Use a separate device or an Android emulator on a Mac/PC.
  4. Don’t pay with your real card. Use a single-use virtual card (Wise, Revolut, ANZ Plus, UP Bank virtual cards).
  5. Treat every session as 100% public to Chinese state agencies. That is the realistic threat model.
  6. Don’t grant browser permissions (camera, microphone, location, clipboard).
  7. If studying the open weights, use a Western inference provider (Together AI, Fireworks AI) — never deepseek.com.

This is standard security-research hygiene for any potentially-hostile software.

Gotchas

  • The English UI fools people. chat.deepseek.com is polished and English; that has nothing to do with where the servers are or who operates them.
  • App Store availability isn’t a safety vetting. Apple and Google check for malware and TOS compliance, not for data-residency-to-China concerns. DeepSeek’s apps have been periodically removed from various app stores under government order, but availability changes don’t reflect underlying safety.
  • “Hosted on AWS Tokyo” doesn’t help. If DeepSeek (the company) hosts an API endpoint on AWS, Chinese law still applies to DeepSeek the company — they can still be compelled to hand data over.
  • DeepSeek the company is not Open-source in any meaningful Western sense. The weights are open; the training code, training data, and corporate practices are not.
  • “DeepSeek is funded by a quant fund, not the Chinese state.” True; doesn’t matter — Chinese law applies regardless of ownership.
  • Subtle benchmark inflation. Several independent reproductions of DeepSeek-V3’s headline coding-benchmark scores have found smaller gaps to Western models than the original claims. Treat benchmark hype with the same scepticism you’d apply to any vendor.
  • VPNs do not change the legal situation. A VPN hides your traffic from your ISP; it does not change what the receiving company is legally required to do with your data.

See also

Sources

Graph View

Backlinks