IBM · Filed Jan 3, 2025 · Published Jul 9, 2026 · verified — real USPTO data

IBM Patents a System That Makes AI Chatbots Check Photos Before Answering

Instead of just answering your question, IBM's proposed AI system would first search for images and videos to make sure the answer holds up. If the visual evidence doesn't match, the chatbot stays quiet.

IBM Patent: Using Images and Video to Fact-Check AI Chatbots — figure from US 2026/0195306 A1
Figure from the official USPTO publication.
See all 5 drawings from this filing ↓
Publication number US 2026/0195306 A1
Applicant International Business Machines Corporation
Filing date Jan 3, 2025
Publication date Jul 9, 2026
Inventors Uri Kartoun, Sophie Batchelder, Krishnan Sugavanam, Mauro Martino
CPC classification 707/690
Grant likelihood Medium
Examiner NGUYEN, CAM LINH T (Art Unit 2161)
Status Final Rejection Mailed (Jun 29, 2026)
Document 20 claims

How IBM's chatbot fact-checker uses images and video

Imagine you ask an AI assistant whether a famous athlete competed in a specific event, and the chatbot confidently says yes, but gets the year, the person, or the event completely wrong. This is called hallucination, and it happens more than anyone would like.

IBM's newly filed patent describes a system designed to catch those mistakes before they reach you. After a chatbot comes up with an answer, the system runs a separate search for images and videos related to your question. A machine learning model then checks whether that visual evidence actually backs up what the chatbot is about to say.

If the pictures and videos don't line up with the answer, the system treats that as a red flag and withholds the response rather than sending you something wrong. The goal is to make AI assistants more trustworthy, especially in business settings where a bad answer can have real consequences.

How the visual verification pipeline filters bad answers

The patent describes a multi-step pipeline built on top of a standard chatbot architecture. When a user submits a query, the system first analyzes the meaning of the question (what IBM calls semantic analysis, meaning it tries to understand the intent, not just match keywords).

It then searches a knowledge base for an answer. That part is fairly conventional. The novel piece comes next: the system runs a separate search across a predefined list of image and video sources specifically related to the activity or person mentioned in the query.

  • A machine learning model is trained to judge how well the visual evidence aligns with what the chatbot found.
  • The image and video search results are fed into that model, which outputs a confidence score indicating how accurate the answer is likely to be.
  • If the score falls below a set threshold, the answer is excluded entirely from the reply.

In plain terms, the system treats photos and videos as a secondary witness. If the visual record contradicts or fails to support the text answer, the chatbot holds back rather than guessing.

What this means for AI hallucination in enterprise tools

AI hallucination is one of the biggest practical barriers to deploying chatbots in high-stakes enterprise environments, think healthcare queries, legal research, or customer support where a wrong answer isn't just annoying, it can cause real harm. IBM's approach is notable because it uses an entirely different modality (visual search) as a cross-check, rather than just running the text answer through another language model.

For everyday users, the most visible effect would be a chatbot that occasionally says 'I don't have a confident answer for that' rather than inventing one. That's a behavior change many people would welcome. Whether the image-search approach is reliable enough at scale is an open question, but the core instinct of building in a refusal mechanism is sound.

Editorial take

This is a real and important problem IBM is trying to solve, and the idea of using visual evidence as a sanity check on text answers is genuinely interesting. That said, the approach has an obvious weakness: image and video searches can be wrong, missing, or misleading too, so you're relying on one imperfect source to validate another. IBM would need to show strong results on real-world queries before this counts as a solved problem.

The drawings

5 drawing sheets from US 2026/0195306 A1 · click any drawing to enlarge

Patent filing page

Which company should we read for you?

We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.

Get one Big Tech patent every Sunday

Plain English, intelligent commentary, no hype. Free.

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

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