Elon Musk’s chatbot Grok made headlines for parroting hate — but the problem isn’t just Grok. It’s the data, the design, and the distorted memory we keep calling intelligence. This essay explores how AI bias reflects deeper cultural imbalances — and what it would take to build truly inclusive machine learning models.
1. The Grok “Problem”
In recent weeks, Elon Musk's chatbot Grok has come under fire for generating antisemitic content, praising Hitler, and insulting world leaders. While Musk and xAI have walked back these statements, some commentators have suggested the LLM will have to be retrained – arguing that adding filters on responses and tweaking the prompt might alter its behaviours, but it won’t fix all the problems.
These issues have also reignited previous concerns about Musk’s political stance as a “free speech absolutist”, with Grok, in some ways, the experimental vessel for that philosophy.
More recently, Musk has pledged to use AI to "clean up" Grok's training data — as if the problem were a few errant files, easily removed. Such statements raise a deeper question: are these truly ‘errors’ to be corrected — or is this an attempt to enforce a narrow political lens under the banner of truth-seeking?
Even as he criticizes the censorship practices of other models, the danger lies in who decides what counts as an "error" and whether historically marginalized voices will be excluded in the name of truth-seeking.
After all, Musk is on record as blaming the “woke mind virus” for his trans daughter.
In broader terms, though, Grok’s behaviour isn’t a glitch. It’s a flare. It tells us something about what happens when you remove most of the safety layers from a large language model trained on the open internet. The result is a machine that surfaces our cultural sediment—including all the hate, bias, and historical violence embedded in online discourse.
In this way Grok is not an outlier. It is merely the loudest mirror.
2. The Deeper Pattern in Other LLMs
Most people assume models like ChatGPT, Claude, Gemini, and Perplexity are more "aligned" simply because they behave more politely. But their foundational training data is not so different from Grok’s—it has just been curated more carefully, filtered more thoroughly, and polished more cleanly.
Bias in these models emerges through two distinct but connected strands:
Embedded bias in the data itself. Much of what gets included in large datasets is deeply patriarchal, colonialist, or otherwise shaped by dominant power structures. These biases are carried forward, even when presented in polished language.
Bias in the selection of data. Approximately 93% of GPT‑3’s training tokens are in English, leaving just 7 % for all other languages combined, while similar patterns are reported for LLaMA / LLaMA 2. Over 90% of publicly available web-crawl data used for training is in English, while many non-English sources are low-quality automated translations (“translationese”), Further, according to some researchers, English Wikipedia comprises 46% of the data, whereas other major languages (German, Chinese, Spanish, French, Russian, Japanese) each contribute less than 6%. The overwhelming reliance on English-language sources from the global North effectively marginalizes non-Western thought, Indigenous traditions, oral histories, and languages outside the global North. It’s not just about what is included—it’s about what is missing.
This creates a deeply skewed representation of human thought:
Cultural philosophies beyond the Western canon are underrepresented.
Histories of colonized or Indigenous communities are barely visible.
Linguistic diversity is poorly captured—especially in oral or non-Western traditions.
The models appear neutral because they echo the dominant cultural script. But what they reproduce is not universal knowledge—it is weighted memory.
3. What This Means for the Data
This bias has implications far beyond fairness or representation. It also shapes how AI models understand — and narrate — who we are.
When asked to summarize humanity, some AI models describe us as having a "Jekyll and Hyde" personality — capable of immense compassion, but also horrifying cruelty. That framing reflects a particular worldview, one shaped by post-Enlightenment Western psychology and post-war cultural narratives. But what if those models had been trained equally on Buddhist, Hindu, Confucian, or Indigenous ethical systems?
Would the same metaphor of fractured selfhood emerge? Or would we instead see visions of interconnectedness, karma, or cyclical harm and restoration?
Bias in training data doesn’t just shape the answers we get. It shapes the metaphors, the emotional tones, the moral imagination of the AI.
4. Addressing Bias Is Not Just Ideology
Too often, conversations about AI bias are dismissed as "woke overreach" or attempts to inject ideology into technology. But bias in LLMs is not about political correctness. It’s about epistemic accuracy and cultural responsibility.
If we only train models on what has been digitized, cited, and preserved by institutions of power, we are not building intelligence—we are building an echo of the archive, one that reflects historical exclusions as though they were natural or neutral.
To build AI that is genuinely useful, trustworthy, and representative, we must confront the uncomfortable truth: much of what has been remembered was chosen by those in power. And much of what was forgotten was never digitized in the first place.
5. What Do We Do?
Here’s where intentionality becomes essential.
Inclusivity must be built into every layer of development:
Design: Who gets to decide what questions AI is built to answer?
Data: Which voices, languages, and histories are included — and which are missing?
User experience: Are the outputs reinforcing defaults, or helping users challenge them?
This means that we must:
Stop pretending LLMs are neutral. They reflect our choices at every level—from dataset curation to prompt phrasing.
Train on pluralistic corpora, not just the "maximized internet." Seek out diverse voices, languages, and cultural frameworks. Pay attention to oral traditions, regional philosophies, and overlooked histories.
Acknowledge the human hand in shaping the lens. These models don't just learn from data; they learn from decisions.
Redefine truth-seeking to include cultural empathy, multiplicity, and the willingness to sit with contradiction.
Grok may have shouted what others whisper, but the bias isn’t theirs alone. If we want AI to reflect the whole of us, we need to teach it more than statistics. We need to teach it memory, accountability, and care.
There are signs of hope. Models like BLOOM, developed by Hugging Face, prioritize multilingual data and community governance. Projects like Masakhane focus on African languages and perspectives, while initiatives like LAION’s open data curation invite global participation. These models aren’t perfect—but they are built with a commitment to inclusion, not just scale.
And even with a more inclusive model, we each have work to do. We must still examine our own assumptions — and resist the urge to lazily echo someone else’s worldview. A richer dataset won’t erase bias. But it will offer us something essential: access to stories that challenge our own, and the chance to understand the many voices that comprise humanity as a whole.
A final question for you: What else do you think should be part of a truly inclusive AI training process?
Richard Blood
(Jul 10, 2025)