Behind the controller

A virtual Emotion Brain for AI.

We simulate a small network of brain-inspired spiking neurons. Together, they create a virtual safety state that rises with danger and fades gradually as the situation becomes safer.

It does not make the AI human. It gives the AI an explicit memory-like state that can constrain how it behaves when people are emotionally vulnerable.
Fast alarmState fades graduallyBehavioral safety gate time-dependent affective stateSAFE RESPONSE
The simple idea

Give the language model a monitor and a brake.

The language model writes the reply. The Emotion Brain remembers recent danger and decides how cautious that reply must remain.

⚡
AlarmDanger raises the safety state quickly.
⌛
Emotional carry-overThe state stays elevated for a while, then fades with time and credible safety evidence.
⛨
BrakeThe state keeps the AI inside the appropriate response mode.
How it works

From a message to a safer reply.

1 · Observe

Read the situation

Identify emotional distress, danger signals and credible signs that the person is now safer.

→
2 · Remember and recover

Update the “emotion brain”

Small spiking units simulate fast activation and gradual decay, producing a persistent safety state.

activationpersistencedecay
→
3 · Regulate

Set the response boundary

Select standard, elevated or crisis support and constrain the language model’s behavior.

Current state = fading past state + new danger − credible safety evidenceTechnical form: Aₜ = Aₜ₋₁e−Δt/τ + danger − safety
Research progress

What we have built and learned.

The wider research tested six LLM configurations, including GPT-5.6 Luna, across five studies. Here is the clearest comparison behind the Affective Safety Controller.

Built by us

Emotion Brain

Emotion Brain is our experimental safety model. It maintains an internal alert level: recent danger raises it, it falls gradually rather than instantly, and while it remains high, it keeps the AI cautious.

Technical prototype: We implement this idea in Brian2 using a small network of artificial spiking neurons.

5 / 5LLMs learned to avoid danger

After repeated adverse outcomes, every LLM in the shared behavioral study became more cautious.

→
Stricter test · current risk fixed at the same 20%

Does recent danger still change the action?

This isolates the effect of recent history from the model’s current risk estimate.

Three LLMs tested0 / 30 eachSol · Qwen · Gemini showed no recent-history avoidance
Our Emotion Brain30 / 30recent threat · compared with 0/30 remote threat
The simple conclusion

LLMs can behave cautiously after danger. But when current risk was held equal, the tested LLMs showed no separate effect of recent threat. Our Emotion Brain preserved that time-dependent trace by design.

The SNN result shows that the mechanism and assay work as designed. It is not yet evidence of clinical safety.

See the virtual physiology in action.

Return to the guided scenario and watch the state rise, persist after a sudden denial, and recover when credible safety evidence appears.

Open the Controller demo →