Get the newsletter
IBM · AI / ML · Filed Nov 14, 2024 · USPTO published May 14, 2026 · May 14 release

Red Hat Patents a Self-Updating Background Dataset for AI Explainability

Getting an AI model to explain its own decisions is hard enough — but doing it accurately when your data is constantly changing is even harder. Red Hat's new patent tackles exactly that problem with a system that rebuilds its explanation context fresh with every single prediction.

Red Hat Patent: Dynamic Background Datasets for AI Explainability, figure from US 2026/0134341 A1
Figure from the official USPTO publication.
Publication number
US 2026/0134341 A1(US20260134341A1)
Applicant
Red Hat, Inc.
Filed · Published
Nov 14, 2024 · May 14, 2026
Inventors
Rui Miguel Cardoso De Freitas Machado Vieira, Robert Geada
Status when we published
Waiting for an examiner
Document
20 claims
US classification
706/12

What Red Hat's rolling AI explanation system actually does

Imagine you ask an AI system to approve or flag a job application, and it does — but you want to know why. Most explainability tools (like SHAP or LIME) need a reference set of "normal" examples to compare against, kind of like needing to know what a typical application looks like before you can say why this one is unusual. The catch: that reference set is usually built once and then frozen, which means it goes stale as the world changes.

Red Hat's patent describes a system that never freezes that reference set. Every time the AI makes a prediction, the system generates fresh comparison examples on the fly, runs them through the same model, and uses the results to update a living dataset. That updated dataset feeds straight into an explanation engine.

The result: you always get an explanation grounded in the AI's current behavior, not its behavior from six months ago when the reference data was last compiled. For teams running AI in production — where data drifts and models get retrained — that's a meaningful improvement over the status quo.

How the distribution sampling loop builds context per inference

The system operates as a continuous loop, one iteration per incoming prediction. Here's the sequence:

  • Receive a target input — the actual data point the model needs to evaluate (a loan application, an image, a log entry).
  • Update a distribution summary statistic — the system maintains a statistical model of all inputs seen so far (think: a running mean and variance). That model gets nudged with each new input.
  • Sample the distribution — the system draws a batch of synthetic "what-if" inputs from that distribution. These are plausible neighbors of the real input, not random noise.
  • Run everything through the target model — both the real input and the synthetic samples get scored. The model returns a real output and a set of counterfactual outputs.
  • Update the background dataset — the real input-output pair and all the synthetic pairs get logged. This is the reference context that explanation tools need.
  • Query the explanatory model — the fresh background dataset, the real input, and the real output are bundled into a request to an explanation layer (likely a SHAP-style or LIME-style model), which explains why the target model produced that output.

The key innovation is that the background dataset is never static. It grows and shifts with every inference, staying aligned with the model's actual behavior over time. The distribution used for sampling also updates continuously, so the synthetic examples remain statistically representative — not artifacts of an old data snapshot.

What this means for real-time explainable AI in production

For most production AI systems, explainability is a post-hoc add-on that relies on a fixed reference dataset baked in at training time. When the model's input distribution shifts — seasonal trends, changing user behavior, a model retrain — the explanations quietly become less accurate, and nobody notices. Red Hat's approach makes the explanation infrastructure adapt at the same cadence as the model itself.

This is especially relevant for regulated industries like finance, healthcare, and HR, where you may be legally required to explain individual AI decisions to affected users. A system that always explains based on current context is a much more defensible audit trail than one frozen to a snapshot from last quarter. For Red Hat's OpenShift AI and Trusty AI customers, this fits neatly into the explainability tooling they already ship.

Source: US 2026/0134341 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.

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 in AI training & infrastructure

All 316 AI training & infrastructure patents →

Watchlists this filing belongs to

All watchlists →