Get the newsletter
IBM · AI / ML · Filed Dec 13, 2024 · USPTO published Jun 18, 2026 · Jun 18 release

Red Hat Patents a System for Making AI Models Forget Specific Data

When an AI learns something it shouldn't have — say, your personal data — deleting that data from a database doesn't erase it from the model's brain. Red Hat is patenting a way to fix that.

By · Updated · 3 min read

Red Hat Patent: AI Model Unlearning Explained, figure from US 2026/0170341 A1
Figure from the official USPTO publication.
Publication number
US 2026/0170341 A1(US20260170341A1)
Applicant
Red Hat, Inc.
Filed · Published
Dec 13, 2024 · Jun 18, 2026
Inventors
Leigh Griffin, Dimitri Saridakis
Status when we published
Waiting for an examiner
Document
20 claims
US classification
706/12

What Red Hat's AI 'unlearning' system actually does

Imagine a doctor trained on thousands of patient records. Even if you delete one patient's file from the hospital's system, the doctor still remembers what they learned from it. AI models have the same problem — once they've trained on a piece of data, that data's influence is baked into the model, even after you delete the original.

Red Hat's patent describes a method for making an AI model actively forget a specific piece of training data, not just remove it from a database. The system adjusts the model's internal settings to reduce the influence of that particular example, then tests whether the forgetting actually worked by feeding the model a question based on that same data.

If the model still seems to "remember" something it shouldn't, the system can keep adjusting. Think of it like a memory-erasure process with a built-in quiz at the end to make sure it took.

How the model tests its own forgetting before finishing

The patent describes a multi-step machine unlearning process — a field of AI research focused on selectively removing the influence of specific training examples from an already-trained model.

Here's how it works:

  • Parameter adjustment: The system modifies the model's internal weights (the numerical values that encode what the model has learned) to reduce how much a specific training example affects the model's outputs.
  • Verification test: After the adjustment, the system feeds the model a test input derived from the same training example that was supposed to be forgotten — essentially asking the model about what it was told to forget.
  • Feedback loop: The model's response to that test is then used to guide further adjustments. If the model still shows signs of remembering, another round of updates follows.

This loop-based approach distinguishes it from simpler unlearning techniques that only perform a single-pass adjustment with no verification step. The goal is to confirm erasure, not just attempt it.

What this means for AI privacy and data-removal rights

Privacy laws like GDPR in Europe and CCPA in California give people the right to request that companies delete their data. But for AI systems, "deleting" a record from a database doesn't actually erase what the model learned from it. That gap between legal compliance and technical reality is a serious problem for any company deploying AI trained on user data.

A working unlearning system would let organizations actually honor those deletion requests in a meaningful way — without retraining an entire model from scratch, which is enormously expensive. For enterprise software, where Red Hat plays, this is a real operational headache. If this approach works reliably at scale, it could become a standard tool in how companies manage AI compliance.

Source: US 2026/0170341 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time IBM patent news breaks.

Add Patentlyze on Google One click.
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.

More from IBM

All 470 IBM patents →

More in AI / ML

All 2,637 AI / ML patents →