The Human-AI Axis: Where It Forms and What Steering It Buys

Across language models from the Gemma, Qwen, and OLMo families, we find a model-specific linear residual-stream direction that separates human from AI self-representation. We identify this as the human-AI axis, a direction that can be causally manipulated by steering the model's position along it. A human-versus-AI separation is already present in an untuned Gemma 3 base model, and across OLMo 2 checkpoints we find that the axis forms within the first 3.7\% of pretraining, long before instruction tuning. Steering along the axis reliably flips a model's stated identity, even on questions held out from axis extraction, and we find the same effect in three language models across the Gemma and Qwen model families. Unlike prompting, steering intervenes directly on this internal representation while holding the prompt fixed, letting us test its causal role rather than just its correlation with what the model says. The representation is also sensitive to context, as a user claiming to be an AI shifts the model's own self-representation significantly toward being an AI on all three models, with no steering. Although steering along the axis shifts a model's own self-report far more than its behavior, steering toward a human self-representation produces safety-adverse behavioral changes. For example, it raises the model's willingness to cause harm significantly on Gemma 3 4B and Qwen 2.5 7B, and directionally on Gemma 3 12B. On Gemma 3 4B, human-ward steering causes prompt-injection execution to rise 16-fold while 94\% of rollouts stay on-task. Writing register such as warm versus formal phrasing shifts the raw projections more than identity does, but steering with a register-matched direction does not reproduce the behavioral effects on Gemma 3 4B, leaving identity as the best-supported interpretation of what steering changes. The human-AI axis is therefore a safety-relevant inference-time variable whose behavioral effects are not captured by self-report alone. Code and reproducibility artifacts will be made publicly available upon acceptance.