Qualcomm Patents an AI System That Shows Its Work

By Patentlyze Team · Updated Jun 12, 2026

Most AI tools give you an answer without explaining their thinking. Qualcomm is patenting a system where the AI essentially narrates its own reasoning process as it goes — not just the final result, but the intermediate steps it took to get there.

What Qualcomm's self-explaining AI actually does

Imagine asking a contractor to renovate your kitchen. They hand you a finished invoice, but you have no idea how they broke the job into tasks — plumbing, electrical, cabinets — or why. Most AI assistants work the same way: you ask a question, you get an answer, the internal logic stays hidden.

Qualcomm's patent describes a system where a second, smaller AI watches the main AI's internal activity while it works, then translates that internal activity into plain instructions explaining what sub-task the AI was handling at that moment. So when the main model is mid-process — say, figuring out a scheduling step inside a larger planning request — the auxiliary system surfaces what's happening and why.

The result is an AI that can produce two outputs at once: the final answer to your request, and a running explanation of the steps it took. That's genuinely useful for anyone who needs to trust or audit what an AI is doing — think enterprise tools, on-device assistants, or safety-sensitive applications.

How the auxiliary network reads the AI's hidden state

The patent describes a two-model architecture. The first model — the main language model — takes in a user's query as a set of input tokens (chunks of text the AI processes) and does two things simultaneously: it builds up an internal representation called a hidden state, and it generates output tokens that form the final answer.

The hidden state is the interesting part. It's the model's internal working memory at a given point in processing — not visible to the user, but rich with information about what the model is focused on. Qualcomm's system adds an auxiliary network (a secondary, lighter-weight AI) that reads this hidden state and translates it into a human-readable set of instructions describing what sub-task the main model was solving at that moment.

So the pipeline looks like this:

• User sends a query (e.g., "Plan a product launch")
• Main model processes it, building a hidden state tied to a specific sub-task (e.g., "Draft a timeline")
• Auxiliary network reads that hidden state and outputs instructions for that sub-task
• Main model's output tokens are also processed into the full task answer

The key claim is that these two outputs — the sub-task explanation and the final answer — are generated from the same inference pass, meaning no extra round-trip or separate call to the model is needed for the explanation.

What this means for AI transparency on-device

AI explainability is one of the hardest open problems in the industry right now. Regulators in the EU and elsewhere are pushing for "right to explanation" requirements on automated decisions. Most current approaches to explainability are bolted on after the fact — they approximate or reconstruct the model's reasoning rather than capturing it directly. Qualcomm's approach, if it works as described, would capture the reasoning from inside the model as it runs.

For Qualcomm specifically, this aligns with its push to run AI models on-device in phones and chips rather than in the cloud. An on-device AI that can explain its own steps becomes far more trustworthy for enterprise, medical, or personal-assistant use cases where users need to verify what the AI is actually doing — not just accept its output on faith.

Editorial commentary on a publicly published patent application. Not legal advice.