Grounding and hallucination across modalities
Grounding means making sure a claim the model makes about an image is actually verifiably tied to what's in the image, rather than a plausible-sounding guess drawn from general world knowledge. A model that has seen thousands of kitchen photos during training might describe a kitchen in an image with plausible-sounding details, a toaster on the counter, say, even if no toaster is actually present, because kitchens in its training data commonly have toasters. That's a grounding failure, the claim is disconnected from the actual visual evidence in front of it.
This shows up as hallucination across modalities, where a model confidently describes something in an image that isn't actually there. It's a distinct failure mode from the kind of hallucination you saw in earlier text-only courses, because here the model has real visual input it could check against, and the error is specifically a failure to stay tied to that input rather than falling back on statistical priors from training.