Manufactured outrage over LLMs’ bias is getting old

You wouldn’t ask a rabbi about the New Testament or an imam about the Torah and expect unbiased responses. So why ask an LLM influenced by the CCP about things you already know you won’t get an unbiased answer to?

I’ve seen a lot of hand-wringing lately on LinkedIn about how supposedly biased or CCP-influenced large language models (LLMs) like DeepSeek-R1 provide unsatisfactory responses to inquiries about politically and geopolitically touchy (for China) topics. The app is obviously censored and–I can tell you–the self-host-able DeepSeek-R1-distilled Qwen2.5 models are censored too (more on that below).

Are you really surprised? Could you not see the bias coming from miles away? Really?

What did you really expect?

Asking an LLM out of China about politically charged issues is just… pointless. What are you proving, exactly? And even if it weren’t censored or biased, should you really rely on an LLM for insights into complex historical and geopolitical matters? These are topics shaped by decades or centuries of context, national identity, and cultural perspectives–expecting an LLM to provide a universally agreeable answer is naive at best.

Asking DeepSeek-R1 about Taiwan is as silly as asking Meta’s LLaMA models about DEI. You’ll get responses shaped by the biases of the people who oversaw their training according to what is euphemistically called “AI safety” (i.e.: censorship). Maybe the next LLaMA model will be more “Trumpist”–so what? Does it matter whether the bias comes from a tech giant, a government, or a particular administration’s policies?

Complex topics with tons of subjectivity and nationally and culturally charged interpretations and decades to centuries of historical context, answered by an LLM in a way that pleases not only either side of the fence, but also those with knowledge on the matter?

Surely, you jest!

Information is “colored” by its source and its moment in history

All information (both stored and consumed) is shaped by its source and moment in history–and those aspects correlate. If you know the source, you can infer the bias and judge the information accordingly. Often this means being ready to take it with a grain or boulder of salt, without making a surprised Pikachu face. That’s part of becoming competent not only in prompting an LLM, but also in the subject matter you are prompting it about, so that you can qualify the output. Only someone in the crap or danger zones would be foolish enough to trust the output of any LLM, especially when it pertains to such tricky topics.

If you’re upset that a Chinese LLM provides China-friendly answers, ask yourself: would you expect a Western LLM to provide neutral answers about Western policies? Of course not. Every model reflects the biases and priorities of its creators.

Manufactured outrage–all for the “engagement”

But here’s the thing: I get it. it. Controversy generates engagement. People love a good drama, China has been in the West’s mind’s eye for years, and anything that’s supposedly controversial (is it, really, though?) makes for great clickbait. So when a Chinese model gives out “controversial” answers (i.e. what you would expect it to, if you know the model’s provenance), it feeds into that endless cycle of reposts of reposts and comments that are so breathless and contain so little that’s unexpected, that I suspect it’s all purely about performative histrionics. Regardless of whether you side with China or not on any of those topics, did you really expect a different outcome from querying DeepSeek-R1 about these things?

Wow, truly, what a big surprise! What a revelation! Next you’ll make a viral post that “exposes” ChatGPT’s refusal to provide you with instructions on making a Molotov cocktail. So insightful; so _groundbreaking! And so vintage 2023!

The actual, practical use of a self-hosted DeepSeek-R1 (actually, Qwen2.5)

I self-host the 14B variant of DeepSeek-R1 with 4-bit quantization on an NVIDIA RTX 4070 Super. It fits within the meager 12GB of VRAM and runs at a more-than-acceptable inference speed. Thankfully, I’m neither in the business of political or geopolitical analysis, nor in the business of “personal branding” through me-too posts on social media about trivialities, in order to build a following of chickens clucking in the LinkedIn coop.

In my case, I primarily use (various) LLMs for code suggestions with the VS Code Privy extension, for summarization, for thinking through ideas about software architecture, and during debugging and refactoring. I also use it with my own Elixir library in an app I’ve been building for fun and personal use, to “interrogate” and tag Linux kernel changelogs . For all these applications, any of the available models on ollama.com/models works to degrees of quality ranging from “barely sufficient, but acceptable” to “excellent”.

DeepSeek-R1:14b, in particular (actually, Qwen2.5 distilled by the full-blown, actual DeepSeek-R1), is excellent. The information provided in the <think>...</think> part of the response can be useful. Certainly, it’s preferable to NOT having such information. The ability to self-host it, even though limited by the amount of VRAM (or RAM, for much slower CPU-based inference) means that I don’t need to worry about API keys, surprise bills, or have to deal with ChatGPT’s lacking UI, which is far behind that of Open WebUI.

LLMs Are tools–stop treating a screwdriver like a political oracle

Instead of using these tools for their actual purpose—like coding, brainstorming, or even just casual conversation, some are turning them into generators of content for weird online drama about nonsense.

Put simply: you wouldn’t use a screwdriver as a hammer, so why waste time complaining about DeepSeek-R1’s obvious bias? It refuses to answer politically charged questions–so what? That’s irrelevant to 99.999% of what you can actually do with it.

The real conversation should be about how LLMs enhance productivity, not about performative outrage over predictable (and thus utterly boring) biases. If you ignore the practical applications of LLMs and focus on political theatrics instead, you’re missing the point of LLMs entirely.

Besides, even without deliberate bias in training, do you truly believe any dataset represents an objective “ground truth” rather than a historically contingent collection of texts with its own inherent biases?

Think critically, use LLMs for what they’re designed for, and stop acting shocked when they reflect the views of their creators and training overseers.

Ignore the practical uses of any LLM and focus on performative histrionics at your own peril.

Besides: even without deliberate training bias, do you seriously believe that any training corpus represents actual “ground truth” and not merely a sample of historically path-dependent documents with its own inherent distribution of biases?