Written by Ayush Yajaman, January 30th, 2025

The AI world has been set ablaze and how! Not even in the faintest of his imaginations would one have thought that the AI juggernauts, the likes of Nvidia and Open AI, would be rattled by a small Chinese company, wiping out a whopping $1 trillion from the US Stock market. From the past few days Deepseek-r1 has been all the buzz. Some are calling it a game-changer — an open-source disruptor challenging OpenAI’s monopoly on large language models (LLMs). Others are skeptical — how did they train a model at just $6 million when OpenAI spends ten times that? But one concern stands out above all: privacy. And that’s exactly what we’re diving into today — how Deepseek’s privacy policies stack up against ChatGPT’s.

Reading through both companies privacy policies, I realised that not much separates the two. Both of them collect your personal data, your inputs, and device information. But Deepseek’s privacy policy does have some unsettling points that stood out:

Keystroke tracking: They collect “keystroke patterns or rhythms” i.e you typing.

Getting data about you from third parties: They take information from third parties about actions you have taken outside their service.

No explicit “No Sharing” statement: Unlike OpenAI, Deepseek’s policy doesn’t clearly state that they don’t sell or share personal data.

Data storage: All your data is stored in servers within the “People’s Republic of China”.

Now, let’s be honest — none of this necessarily means your data is compromised, but it’s enough to make anyone pause. I certainly did.

The Open-Source Paradox: Deepseek’s Biggest Strength

Ironically, Deepseek is fully open-source, unlike OpenAI, which has strayed far from its “open” roots. This means there’s a way to sidestep these privacy concerns entirely — run it locally.

Using Ollama and an AI-chat framework like Chatbox, you can deploy Deepseek-r1 directly on your system. No internet connection, no data leaks, no hidden risks. That’s something you simply can’t do with OpenAI’s models, which require API keys and an active online connection.

Running Deepseek Locally: The Hardware Challenge

Ollama provides a range of model sizes, from a 1.5B parameter model (tiny) to a monstrous 671B parameter model. But as you’d expect, bigger models = bigger requirements. Here’s what I found:

• 1.5B model: Only 1.1 GB of storage, but it was hallucinating more than a sleep-deprived student before finals.

• 671B model: Performs amazingly well, but requires a jaw-dropping 404 GB of storage — not exactly laptop-friendly.

So unless you have some serious hardware, you’re probably going to depend on online access — and that brings us back to privacy concerns.

The Real Takeaway

So, what’s the best move?

• If privacy is your top concern, invest in hardware and run Deepseek r1 locally.
• If you want to be economical, OpenAI’s GPT-4o is still a solid choice.
• If you would rather bear ongoing costs instead of one-time hardware investment, OpenAI’s o1 model is your guy.

Deepseek brings a lot to the table, but it’s really important to know what you’re getting yourself into before using it. So, take your time to research and make the choice that’s best for you.