In AI conversations, we often focus on how to reduce model hallucinations and drift issues. But one developer has offered a different perspective: drift might not just be a machine problem, but rather a responsibility that both humans and AI need to share.

This experiment, called a 'prompt governance layer,' centers around having AI serve as a calm drift detection layer. Rather than generating new content, it evaluates the clarity, basis, and certainty of conversations.

**How it works:**

When given recent text or a conversation, this governance layer will:

- Mark the reliability of statements (based on evidence, reasonable inference, possible reconstruction, high hallucination risk)

- Detect drift on the human side (goal shifting, vague expressions, emotionally confident assertions without basis)

- Detect drift on the model side (confident assertions without basis, fabricated details, ignoring previous constraints)

The output uses a concise, structured report format, including drift risk levels, primary sources of uncertainty, statements most likely to be reconstructed, and specific suggestions for improving clarity.

**How to use it:**

1. Copy the prompt governance layer template into an LLM

2. Ask it to evaluate recent conversational responses

3. Observe whether the conversation becomes clearer and more reliable

Some commenters pointed out that all systems relying on internal reference points experience drift, which requires external reference points for calibration. The developer responded that this experiment explores how human-AI pairs can share responsibility for early drift detection, with future parts addressing recalibration and external references.

The advantage of this approach is that it doesn't aim for perfect solutions but rather makes both parties in the conversation more aware of potential drift issues through structured feedback. Even partial improvements can provide valuable information for more reliable AI interactions.

This experiment is the first in a series on governance-based prompts, with future parts exploring calibration and re-anchoring mechanisms. For developers who frequently use AI conversations, this approach might offer new tools for building more reliable AI applications.