Intelai
Multimodal Models Do Not Weigh Evidence
A multimodal system can read the scan, the note, and the lab value. Ask what it does when they disagree, and the demo ends.
The demo is a scan, a clinical note, and a lab value, and a paragraph that summarizes all three. It is impressive, and it is the wrong test.
The test that matters is what happens when the three disagree. The note says one thing, the lab value says another, and the image is ambiguous. A clinician resolves that with training, context, and a phone call. A model resolves it with whatever the training data taught it about which input to trust, and that lesson is usually invisible.
Fusion is not adjudication
Multimodal systems learn a shared representation across inputs. Image and text get projected into a space where they can be compared, usually through contrastive training on paired data. That produces a model that is good at associating a picture with a description.
Association is not adjudication. Nothing in the training objective teaches the model that a lab value measured this morning outranks a note written last week by a resident who had not seen the result.
So the system produces a fluent summary that averages the inputs, and the average is the one thing a decision maker must not accept. When two sources conflict, the correct output is usually a question, not a paragraph.
A model that averages conflicting evidence produces confidence where a human would produce a phone call.
Where the disagreement hides
The conflicts are rarely dramatic. They show up as small inconsistencies that a domain expert catches and a model smooths over.
A form where the OCR read a 3 as an 8 and the total does not reconcile. A scanned table whose column headers were split across a page break. A photo of equipment with a handwritten annotation that contradicts the printed label. A video where the timestamp on the overlay does not match the sequence of events.
Each of those is a case where one modality is more reliable than another, and the reliability depends on the specific document, not on the modality in general. A model trained on clean pairs learns that text and image agree. Real documents are where they do not.
The failure is quiet because the output is well written. A wrong answer that reads badly gets caught. A wrong answer that reads like a competent summary gets forwarded.
The evaluation you actually need
Build a contradiction set. Take twenty cases from your own work where two sources disagreed and a human resolved it. Record which source won and why. That set is worth more than any benchmark, because it tests the exact behavior your deployment depends on.
Then score three things.
Does the system surface the conflict. Not resolve it. Surface it, in the output, in a form a human can see.
Does it attribute each claim. A summary where every sentence can be traced to a specific input is reviewable. A summary with no provenance is a paragraph you have to re-derive by hand.
Does it abstain when the inputs are insufficient. An abstention is a feature. A system that always answers is a system that will answer when it should not.
Run the same case three times. If the answer changes between runs, the system is guessing, and the guess is not stable enough to build a process on.
What this means for design
Keep the modalities separable until a human decision point. The combined view is useful for review. The separate views are what let a reviewer check the model's work.
Ask for citations at the sentence level, and reject outputs that cannot produce them. This is a product requirement, not a research problem. If the vendor cannot show which input produced which sentence, they have built a summarizer, not a decision aid.
Log the inputs with the output. The most common failure in production is not a bad model. It is a good model given a bad input, six weeks ago, with no record of what it saw.
Put a threshold on disagreement. If the sources conflict beyond a defined tolerance, route to a person regardless of the model's confidence. That rule is cheap to implement and it catches the class of error that costs the most.
Where it works today
The strongest near-term case is document-heavy back office work, not clinical diagnosis.
A loan file, an insurance claim, a contract package, an inspection report. These are bundles of text, tables, images, and signatures where the task is extraction and reconciliation, and where a human reviewer is already in the loop. A multimodal system that reads the bundle, flags inconsistencies, and cites its sources saves real time, and it fails in a way the reviewer can see.
That last property is the whole argument. The value is not that the model is right more often. The value is that when it is wrong, the error is visible at the point where a person is already looking.
Show me a case where two inputs disagreed, and show me what your system did.
If the answer is a paragraph that resolves the conflict without mentioning it, the system is not ready for a decision that matters. If the answer is a flagged conflict with both sources cited and a recommendation to check, you are looking at something that can be deployed.
The difference is not model size. It is whether anyone thought about disagreement when they designed the output.
Reading three inputs is easy. Saying which one you believe, and why, is the product.