Adobe's New Patent Covers a System That Fact-Checks AI-Generated Image Captions
AI systems that describe images often make things up, confidently calling a red apple green or adding people who aren't there. Adobe has filed a patent for a method that catches those errors by breaking captions apart and checking each piece individually.
How Adobe's caption cleanup actually works
An AI writes a caption for a photo: "A woman in a blue dress stands next to a white car on a sunny day." The trouble is, the car is actually red, and the AI made that up. That kind of confident, plausible-sounding error is called a hallucination, and it happens constantly with AI image tools.
Adobe's patented approach attacks this by splitting the caption into smaller, testable pieces. Instead of checking the whole sentence at once, the system checks things like "blue dress," "white car," and "sunny day" as separate claims, then compares each one against the actual image. Any piece the image doesn't support gets dropped from the final caption.
The result is a description that only includes what can be visually confirmed. You get a shorter, more careful caption instead of a fluent-sounding one full of invented details.
… decomposing, using a language generation model, the input caption into a plurality of partial captions, wherein each of the plurality of partial captions includes a portion of the input caption; …
Translation: The system breaks a full caption down into smaller, bite-sized pieces for analysis.
Inside the decompose-verify-rebuild pipeline
The system uses two types of AI working together: a standard large language model (an LLM, the kind that generates text) and a multimodal large language model (an MLLM, which can look at both images and text at the same time).
Here is the sequence:
- Step 1, Decompose: The LLM takes the original caption and breaks it into partial captions, short fragments that each make a single verifiable claim about the image.
- Step 2, Verify: The MLLM looks at each partial caption alongside the actual image and judges whether that claim is supported or not.
- Step 3, Rebuild: Only the partial captions that pass verification get assembled into the final output caption. Any fragment flagged as a hallucination is excluded.
The key design insight is that long, flowing sentences are hard to check all at once. Shorter, isolated claims are much easier to verify against visual evidence. The patent describes this decomposition step as central: without it, the verification step would struggle to pinpoint which part of a sentence is wrong.
… where each partial caption can be independently verified to detect hallucinations.
Translation: Each smaller piece is checked separately to see if the AI made anything up.
What this means for AI tools that describe images
AI image captioning is already built into tools used for accessibility (screen readers for people with visual impairments), digital asset management, and automated content tagging. A hallucinated caption in any of those contexts is not just an embarrassment, it can actively mislead. Adobe's steady investment in AI-generated content tools suggests this fits a broader effort to make AI output reliable enough for professional workflows.
For everyday users, the practical payoff is straightforward: if you use an AI tool to describe your photos or auto-tag a library of images, you would get descriptions you can actually trust rather than ones you have to fact-check yourself.
That makes this Adobe's seventh filing we've tracked since July in our AI safety guardrails watchlist, after one on teaching AI to admit uncertainty and one on flagging risky data for human review.
The entire approach runs on software, using two AI models that already exist as widely available tools. No new hardware, no specialized equipment, nothing that has to be invented before this can ship.
The main friction is time. Checking a caption by breaking it into pieces and verifying each one against the image takes more steps than generating a single caption the old way. That overhead matters less in a behind-the-scenes tagging job and more in a live tool where someone is waiting for a result.
The problem being solved, AI-generated descriptions that confidently state things the image does not show, is well-documented and already embarrassing enough that companies have reason to fix it. The architecture described here looks like a straightforward path to doing that.
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
11 drawing sheets from US 2026/0268064 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →