Get the newsletter
IBM · AI / ML · Filed Mar 31, 2025 · USPTO published Oct 1, 2026 · Oct 1 release

IBM Patents a System That Adjusts AI Prompts to Match How Different People Talk

Different people talk in different ways, and AI that ignores that tends to produce text that feels off. IBM has filed a patent for a system that measures how "surprising" AI-generated prompts are to a reader, then dials that up or down to match a specific audience's communication style.

A system for generating and refining AI prompts, showing how human input and various filters adjust prompt perplexity and diversity. Drawing from patent filing US 2026/0300630 A1.
A system for generating and refining AI prompts, showing how human input and various filters adjust prompt perplexity and diversity.All 6 drawings ↓
Publication number
US 2026/0300630 A1(US20260300630A1)
Applicant
International Business Machines Corporation
Filed · Published
Mar 31, 2025 · Oct 1, 2026
Inventors
Bo WEN
Status when we published
Examiner raised objections; the company can respond
Document
20 claims
US classification
704/9
Examiner
WEAVER, ADAM MICHAEL (Art Unit 2658)

What IBM's communication-style AI prompt tuning actually does

Have you ever read something that felt like it was written for the wrong person? A legal notice explained like a children's book, or a casual explainer dressed up in stiff corporate language? That mismatch is a real problem when AI systems generate text meant for specific groups of people.

IBM's patent describes a way to fix that automatically. The system uses multiple AI language models to produce a variety of draft prompts, then measures a quality called "perplexity" for each one. Perplexity is essentially a score for how unexpected or confusing a piece of text would feel to a given reader. A high score means the words are surprising or hard to follow; a low score means they're predictable and easy.

By tuning that score up or down, the system tries to match the writing style of a target population, whether that's teenagers, medical professionals, or people who communicate in very informal ways. The goal is AI-generated content that actually fits how its intended audience speaks and reads.

From the filing · CLAIM 1
generating, based on different large-language models (LLMs), prompts that are diverse; and adjusting a level of perplexity in the generated prompts to correspond to one or more communication styles of a target population.

Translation: The system creates varied AI prompts and tweaks how unpredictable the wording is to fit how specific groups of people actually talk.

How perplexity scores shape the generated prompts

The patent describes a two-step process. First, the system pulls from multiple large language models (LLMs), the kind of AI behind tools like ChatGPT, to generate a diverse pool of candidate prompts. Using more than one model helps ensure variety rather than a batch of text that all sounds identical.

Second, the system measures and adjusts the perplexity of those prompts. In AI and linguistics, perplexity is a statistical measure of how well a language model predicts a sequence of words. In plain terms: low perplexity means the text is smooth and predictable; high perplexity means the word choices are unexpected or complex. Think of it as a readability score with a more precise mathematical foundation.

The system then shifts that perplexity level to match the communication style of a target population. If the target audience uses simple, direct language, prompts get tuned toward lower perplexity. If the audience is technical or uses specialized vocabulary, prompts can be kept at a higher perplexity level.

  • Multiple LLMs generate a diverse starting set of prompts
  • Each prompt is scored for perplexity (how surprising its word choices are)
  • The score is adjusted to align with the target audience's communication norms

What this means for AI tools aimed at specific audiences

If this approach works at scale, it could make AI-generated content meaningfully more useful in contexts where audience mismatch causes real harm. Think of public health communications that need to reach low-literacy communities, or customer service tools deployed across different regions with very different speech patterns. Right now, most AI text generation treats the audience as a constant. A system that actively accounts for who is reading would be a genuine shift in how these tools get deployed.

For everyday users, the practical question is whether the output actually becomes more natural. IBM keeps filing on language model customization and enterprise AI Perplexity is a well-understood metric in research, but turning it into a reliable real-world tuning dial for broad audience groups is an unsolved challenge, and the patent does not spell out how that gap gets closed.

IBM's 37th filing in our Language AI coverage since May adds to a run that includes one on drug molecule structure and one that rewrites prompts.

Source: US 2026/0300630 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.

The drawings · 6 sheets

6 drawing sheets from US 2026/0300630 A1 · click any drawing to enlarge

Computer system components and network., drawing sheet 1 from US 2026/0300630 A1
Sheet 1 Computer system components and network.
AI prompt generation and refinement., drawing sheet 2 from US 2026/0300630 A1
Sheet 2 AI prompt generation and refinement.
Validating synthetic data for AI., drawing sheet 3 from US 2026/0300630 A1
Sheet 3 Validating synthetic data for AI.
Generating and adjusting AI prompts., drawing sheet 4 from US 2026/0300630 A1
Sheet 4 Generating and adjusting AI prompts.
Filtering diverse AI prompts., drawing sheet 5 from US 2026/0300630 A1
Sheet 5 Filtering diverse AI prompts.
Analyzing AI prompts for patients., drawing sheet 6 from US 2026/0300630 A1
Sheet 6 Analyzing AI prompts for patients.
Patent filing page

Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.