IBM · Filed Feb 6, 2025 · Published Aug 6, 2026 · verified — real USPTO data

IBM Patent Proposes Rival AI Panels to Evaluate Each Other for Ethics

IBM is patenting a system where an AI's output doesn't just get checked by one safety filter, but by a whole panel of different AI models, each trained differently, voting on whether the response passes an ethics test.

IBM Patent: AI Outputs Judged by a Panel of Other AIs — figure from US 2026/0228214 A1
Figure from the official USPTO publication.
See all 5 drawings from this filing ↓
Publication number US 2026/0228214 A1
Applicant International Business Machines Corporation
Filing date Feb 6, 2025
Publication date Aug 6, 2026
Inventors Nixon CHEAZ, Christopher Thomas LEONARD
CPC classification 707/692
Grant likelihood Medium
Examiner PENG, HUAWEN A (Art Unit 2169)
Status Final Rejection Mailed (Jun 11, 2026)
Document 20 claims

How IBM's AI-judging-AI ethics check works

Imagine asking an AI a question and, before the answer reaches you, it gets sent to a committee of other AIs for review. Each committee member has its own background and training, so they bring different perspectives. If enough of them flag a problem, the original question gets rewritten and sent back to the AI for a better answer.

That's essentially what IBM is describing here. Instead of relying on a single built-in safety filter, this system uses a diverse panel of large language models (the same kind of AI behind chatbots) to score the output for ethical issues. Each model rates how appropriate the response is.

If the scores suggest something went wrong, the system rewrites your original prompt to steer the AI toward a safer answer, then tries again. The goal is to catch problems that any single filter might miss, by getting multiple independent opinions before the response ever reaches you.

Inside IBM's multi-model scoring and prompt rewrite loop

The system sits between a user and a generative AI model, acting as an automated quality-control layer. Here's the sequence:

  • A user submits a prompt and the AI generates an initial response.
  • That response is forwarded to a panel of diverse LLMs (large language models), chosen specifically because they have different architectures and were trained on different datasets. Think of it like convening judges from different legal traditions rather than cloning one judge.
  • Each panel member returns a score reflecting how well (or poorly) the output handles an ethical dimension, such as bias, harm, or factual integrity.
  • Based on those scores and the original output, the system rewrites the user's prompt and resubmits it to the generative AI, aiming to coax a better response.

The key design choice is diversity: using models with different training backgrounds means a blind spot in one model is more likely to be caught by another. A single safety classifier can be fooled or share biases with the model it's checking; a panel with genuinely different foundations is harder to fool uniformly.

The claim covers the full loop: receive, evaluate, score, rewrite, resubmit.

We find one patent like this every day. Get the best of each week in your inbox, free →

What this means for AI safety in enterprise software

Enterprise customers buying AI tools from IBM (or anyone else) need assurance that outputs won't embarrass them legally or publicly. A single safety layer is a single point of failure. A panel-based approach raises the bar considerably, and the automatic prompt rewrite means you never have to see the bad answer or manually rephrase your question.

For IBM specifically, this fits squarely into its watsonx AI governance pitch: the idea that business AI needs auditable, trustworthy guardrails, not just speed. If this system ships in a product, it could become a selling point for regulated industries like banking, healthcare, and government contracting, where a single inappropriate AI output can carry real consequences.

Editorial take

This is a genuinely interesting architectural idea: using model diversity as a defense against shared blind spots is a smarter approach than stacking more layers of the same thing. The automatic prompt rewrite is the part worth watching, because it closes the loop without requiring the user to do anything. Whether IBM can make the panel fast enough to be practical in real-time applications is the open question.

There are more where this came from

We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.

The drawings

5 drawing sheets from US 2026/0228214 A1 · click any drawing to enlarge

Patent filing page

Source. Full patent text and figures from the official USPTO publication PDF.

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