Tree D: outside the language interface. Summary in plain language
The question
Levin makes a point of method about language models. Language is what we forced them to produce, so when we ask them about themselves we learn about the forcing and little about anything else going on inside. The AI-consciousness field reached the same conclusion by another route. Models are trained on the literature about consciousness, so their self-reports are contaminated. Both lines of argument lead to the same advice, which is to measure something the model was never trained to talk about.
What an experiment looks like
We take an open-weight model whose training data is public. We push its internal state off course by injecting a concept into its activations, using a method Anthropic published in 2025, and watch whether the state comes back and how fast. An untrained copy of the same network is the comparison. In a separate test we wire the model's internals into a small control loop, give it a task with obstacles that has nothing to do with words, and score how it gets around them. We also rank how cheaply each kind of intervention changes its behaviour. The interventions are editing weights, a training step, a reward, an instruction and a persuasive argument.
Last, we ask the model about its own state with every mind-related word removed from the prompt, and check whether its answers have anything to do with the internal measurements.
What the outcomes could be
- The model shows internal goal-directedness that differs from what it says about itself. The welfare field wants that kind of measurement and does not yet have it.
- Its internal goal-directedness looks the same as an untrained network's. Then the premise of applying Levin's method to language models fails, and we say so.
- Self-reports do track internal state. We do not expect this, and it would matter to anyone who uses self-report as evidence.
What kind of research this is
Mechanistic interpretability, framed as behavioural science. This area has the attention and the money. It is also where a group without a track record gets lumped in with the advocates, which is why the tree comes late in the order.
Timeline
It starts in year one, and only once we have published a deflation. Once started it takes six months of work on rented GPUs.