Why AI Image Models Still Make Things Up, and What Apple's Researchers Are Doing About It
A new study from Apple ML Research digs into why multimodal AI models hallucinate, meaning they describe images with confident-sounding details that simply are not there, and how a training technique called preference alignment could fix it.
Key points
- Multimodal large language models, AI systems that process both images and text, frequently "hallucinate" by generating descriptions that contradict what is actually in a picture.
- Preference alignment, a training method that teaches a model to prefer accurate answers over plausible-sounding ones, is well-established for text-only AI but far less studied for image-understanding models.
- Apple ML Research published a study examining how alignment techniques affect image-language models, covering what works and what does not.
- These models power features like image captioning and accessibility tools that millions of people use every day.
Tell an AI to describe a photo and it'll usually get most things right. But sometimes it confidently adds objects or text that simply aren't in the picture. Researchers call this hallucination, and it's a known, stubborn problem with a class of AI systems called multimodal large language models, or MLLMs, which are AI models trained to understand both images and written language at the same time.
For a text-only chatbot, hallucination means stating a wrong fact. Image models can do that too, but they can also contradict the visual evidence in front of them: describing a red car as blue, inventing a sign on a shop wall, or claiming a person is smiling when they aren't. The model isn't lying. It just has no reliable anchor to the image.
Why does this keep happening?
Text-only AI models have benefited enormously from a training step called preference alignment. After the initial training, researchers show the model pairs of answers and teach it to prefer the more accurate, helpful one. Over thousands of examples, the model learns to self-correct.
That correction step has been studied far less for models that handle images. A model trained mostly on text develops strong instincts about what sentences should sound like. When you add images, those instincts don't automatically transfer to checking visual facts.
What did the Apple study find?
Apple ML Research set out to map the problem systematically. The team tested how different preference alignment strategies change the behaviour of image-language models, identifying which approaches reduce hallucination and which fall short.
The sharpest finding is that alignment for MLLMs must handle a second dimension of accuracy: not just "is this sentence factually correct?" but "does this sentence match what is in the image?" Those are different questions, and training methods built for text alone don't automatically handle the visual check. Our earlier report on Apple's work helping AI recover mid-task from 24 July 2026 showed a similar pattern: fixes designed for one modality rarely transfer cleanly to another.
The study is positioned as a resource for the wider research community, mapping open problems and pointing toward the most promising directions.
What does this mean for ordinary users?
If you use an app that reads out image descriptions or answers questions about pictures you upload, you're relying on this technology. Hallucinations in those contexts are more than a curiosity. A wrong description in an accessibility tool can genuinely mislead a visually impaired user.
Better alignment training is one of the clearest paths toward models that stay honest about what they actually see. Progress happens inside the training process, long before a product reaches your phone. Users don't need to do anything differently.
Common questions
Is hallucination the same problem in image models as in text chatbots?
Not quite. Text chatbots hallucinate by stating wrong facts. Image models can do that too, but they can also contradict the visual evidence right in front of them, describing things that aren't there or missing things that clearly are.
Does this study change any products I can use right now?
No. This is foundational research, not a product announcement. Its value is in giving other researchers a clearer map of the problem so that future models, across many companies, can be built with stronger alignment training from the start.
