The report
An assistant can compare dozens of satellite images yet still treat a corrupted measurement as the deciding fact. EarthRecord tested multimodal agents on questions involving Landsat, MODIS, VIIRS and Sentinel-2 records. After evaluating normal performance, the researchers altered data from one instrument while leaving other evidence sources available. Of 200 formerly correct answers selected for the intervention, 107 became incorrect, according to the study.
What a conflicting image did to the answer
The 53.5% deterioration rate applied to the previously correct subset, not all 238 questions in the larger study. Most changed answers followed the direction of the manipulated sensor rather than displaying an arbitrary decline. The researchers also ran a placebo comparison that changed substantially fewer outcomes. Yet the agents rarely identified the disagreement as a possible instrument defect: only a minority mentioned inconsistent evidence, and very few explicitly attributed failure to the suspect instrument.
The experiment examined ten task families across different geographies and several decades of observation. These are valuable controls for an evidence-processing failure, but the data manipulation was part of a deliberately constructed evaluation. No criminal actor tampered with a public satellite archive, and no real land-use decision was shown to fail. The EarthRecord Case keeps that distinction explicit.
A multi-source answer needs independent provenance
Having three sensor names in a summary is not the same as demonstrating independent agreement. Measurements can disagree because of calibration, collection time, weather or tampering. A defensible analytical pipeline should retain instrument identity and processing lineage, flag conflicting measurements and refuse to elevate one corrupted series without a stated reason. The next test is a blinded, realistic sensor-drift scenario with traceable source selection and independently measured decision errors.