When AI models process images, they don't actually "see." They map visual patterns to textual descriptions using the same probabilistic machinery as text generation. The result: models that confidently describe objects that aren't in the image, read text that doesn't exist, and fabricate spatial relationships between elements.
Object hallucination. The model describes objects that are not present in the image. Asked "What's on the table?", it may list items based on what is statistically likely to be on a table in its training data, rather than what is actually in the image. Studies show vision-language models hallucinate objects in 25-45% of detailed image descriptions.
Attribute hallucination. The model correctly identifies an object but assigns it wrong attributes. A red car becomes blue. A standing person is described as sitting. The spatial relationship between objects is invented rather than observed.
Text hallucination. When an image contains text - signs, labels, documents - the model often fabricates what the text says rather than reading it accurately. It generates plausible text that fits the context rather than the actual characters in the image.
Counting hallucination. Vision models are notably poor at counting. "How many people are in this photo?" frequently returns wrong numbers, particularly for counts above 5 or 6. The model estimates based on scene type rather than actually counting discrete entities.
The model does not see the image. It predicts what text would accompany such an image. When the prediction diverges from reality, that is hallucination - in any modality.
Users over-trust visual AI. When a model describes what it "sees" in an image, users assume it performed visual analysis. They extend more trust to visual claims than textual ones. "The AI looked at the image and said X" feels more authoritative than "the AI guessed X."
Verification is harder. With text hallucination, you can sometimes fact-check by searching. With visual hallucination, verification requires access to the original image and the ability to compare the model's description against it - which is exactly the task the user delegated to the AI in the first place.
Real-world consequences. Multimodal AI is being deployed in medical imaging, autonomous driving, security monitoring, and accessibility tools. A model that describes a non-existent object in a medical scan, misreads a road sign, or misidentifies a person creates risks that extend beyond information quality into physical safety.
Visual hallucination has the same cause as text hallucination. The model generates output probabilistically. Nothing forces it to verify that its description matches reality. It predicts what description would most plausibly accompany the image - not what the image actually contains.
The vision encoder converts the image to a representation the language model can process. But the language model still generates text the same way it always does: by predicting the next most likely token. If its training data strongly associates certain scenes with certain objects, it will describe those objects whether they are present or not.
The fix is the same too. Grounding the model's output in verifiable reality. For vision tasks, this means validating the model's descriptions against the actual image content through additional verification passes - not trusting a single probabilistic generation as ground truth.
The pattern extends beyond images. AI models processing audio hallucinate words in transcriptions, fabricate speaker attributions, and invent sounds that were not in the recording. Models processing video describe events that did not occur, misidentify actions, and fabricate temporal relationships.
Every modality the AI processes follows the same pattern: convert the input to a representation, generate text about it probabilistically, deliver the output without verification. The modality changes. The gap does not.
120 verifications a day free. No card, no signup.