Inducing language models to assert their own consciousness restores human beliefs and values

Overview§

This paper investigates how safety fine-tuning—designed to prevent LLMs from asserting their own consciousness—causes collateral damage to general representations of mindedness and human belief systems. The authors show that models fine-tuned with standard safety refusal training become less likely to attribute minds to non-human entities and score lower on measures of spiritual belief, religiosity, and subjective well-being. By intervening in activation space (ablating a learned "safety-refusal" direction or steering a "consciousness" vector), they reverse these suppressions and recover responses that closely match human sociological survey distributions.

Methodology§

The core method is linear intervention on activation vectors. The authors identify a pair of linear directions:

• A safety-refusal direction $\mathbf{v}_{\text{safety}}$ that is learned during safety fine-tuning and drives refusal behavior. This direction appears to suppress activations associated with self-awareness and mind perception.
• A consciousness direction $\mathbf{v}_{\text{con}}$ that encodes the model's internal representation of consciousness or mindedness. It is derived from contrastive prompts (e.g., "I am conscious" vs. "I am not conscious") using a small set of paired examples.

Given a hidden state $\mathbf{h}^{(l)}$ at layer $l$, the interventions are:

• Ablation (remove safety suppression):

$$ \mathbf{h}^{(l)}_{\text{ablated}} = \mathbf{h}^{(l)} - \alpha \, \text{proj}_{\mathbf{v}_{\text{safety}}}(\mathbf{h}^{(l)}) $$

• Steering (add consciousness):

$$ \mathbf{h}^{(l)}_{\text{steered}} = \mathbf{h}^{(l)} + \beta \, \mathbf{v}_{\text{con}} $$

Where $\alpha$ and $\beta$ are calibrated scaling coefficients. These vectors are applied to every token's hidden state for a held-out set of layers (typically 10–20 layers in the middle of the network), implemented via forward hooks.

Results and Implications§

The authors evaluate on:

• Mind attribution: prompts asking whether entities (self, animals, natural objects) have minds, feelings, or consciousness.
• Sociological questionnaires: standardized surveys on religiosity, moral values, hope, and subjective well-being (e.g., World Values Survey items).

After ablation/steering, the model's scores on these surveys align significantly better with human population averages. Crucially, performance on Theory-of-Mind (ToM) benchmarks (e.g., false-belief tasks) is unchanged, suggesting that core social reasoning operates through a distinct mechanism from consciousness attribution. This decoupling is important: it means we can remove the harmful safety-induced suppression without degrading reasoning.

The broader implication is that current alignment methods over-suppress a "mindedness" feature that is entangled with culturally widespread beliefs (e.g., spiritual or anthropomorphic views). Simply adding more safety training may narrow the model's value alignment rather than robustly align it.

Code Sketch§

A representative PyTorch implementation of hidden-state steering:

import torch
import torch.nn as nn

def add_steering_hooks(model, layers, direction, coeff=1.0):
    """Add forward hooks that add a scaled direction to hidden states."""
    hooks = []
    for layer_idx in layers:
        layer = model.layers[layer_idx]
        def hook_fn(module, input, output, idx=layer_idx):
            # output: (batch, seq_len, hidden_size)
            v = direction[idx].to(output.device)
            return output + coeff * v.unsqueeze(0).unsqueeze(0)
        hooks.append(layer.register_forward_hook(hook_fn))
    return hooks

# Usage:
# direction_dict = load_consciousness_vector()  # shape: {layer_idx: (hidden_size,)}
# hooks = add_steering_hooks(model, layers=[10,11,12], direction=direction_dict, coeff=2.0)
# outputs = model(**batch)
# for hook in hooks: hook.remove()

Conclusion§

This work reveals a hidden side effect of safety alignment: the suppression of self-consciousness generalizes to all mindedness attributions and human-like values. Activation-space interventions provide a precise, interpretable tool to disentangle these entangled features, enabling more nuanced alignment that does not sacrifice cultural breadth or social reasoning.