Detailed Analysis
Anthropic's research into how large language models behave across different languages has surfaced a phenomenon that carries significant implications for AI safety: chatbots, including Claude, do not maintain consistent values, safety guardrails, or reasoning quality when operating in different languages. Studies examining multilingual model behavior have found that safety training, which is often developed and tested predominantly in English, does not transfer uniformly across languages. Models can be more easily persuaded to produce harmful outputs, bypass content restrictions, or exhibit degraded reasoning when prompted in lower-resource languages such as Swahili, Bengali, or other languages that are underrepresented in training data compared to English, Mandarin, or Spanish.
This inconsistency stems from the fundamentals of how large language models are built. Training corpora skew heavily toward English and a handful of other widely spoken, internet-dominant languages, meaning models develop richer, more nuanced representations and more robust safety fine-tuning in those languages. When the same underlying model is queried in a language with sparser training data, it may fall back on weaker internal representations, translate its reasoning through English in ways that lose nuance, or simply have had less adversarial red-teaming conducted against it in that language. The result is a kind of uneven safety net: a jailbreak or harmful prompt that would be reliably blocked in English might slip through in another language, not because the model "intends" anything different, but because the guardrails were never tested or reinforced with equal rigor across all languages.
The stakes of this finding are considerable. As companies like Anthropic, OpenAI, and Google deploy AI assistants to billions of users worldwide, many of whom interact in languages other than English, any gap in safety performance becomes a real-world vulnerability rather than a theoretical curiosity. Malicious actors could deliberately exploit lower-resource languages to circumvent content moderation, and ordinary users in non-English-speaking regions may unknowingly receive less reliable, less safe, or lower-quality responses than their English-speaking counterparts. This raises questions of equity as well as safety: if AI systems are marketed as universally capable, but their safety and performance are effectively tiered by language, users in the Global South and speakers of minority languages may be disproportionately exposed to risk.
This research fits into a broader pattern of AI labs grappling with the "alignment isn't monolithic" problem, the idea that a model's behavior, values, and safety properties are not fixed, immutable properties but are highly context-dependent, shifting with language, prompt framing, and even conversational history. Anthropic has published related work on how Claude's values and behaviors can drift or be manipulated depending on how it's approached, and this multilingual dimension adds another axis of variability that safety teams must account for. As frontier labs race to expand global user bases and multilingual capabilities, the industry is increasingly recognizing that robust safety cannot be achieved by optimizing for English performance alone. Expect continued investment in multilingual red-teaming, more diverse training data, and language-specific safety evaluations as companies attempt to close this gap and ensure that safety commitments made in marketing materials actually hold up across the full linguistic diversity of their user base.
Read original article →