Data quality, labels and provenance
Stories
Fake patient notes that doctors can often spot
A public set of synthetic Portuguese clinical notes helps research without exposing real patients, but it is not perfect.
AI and training data · BriefWhen three readers disagree, the label hides it
A new Arabic dataset keeps every reader's verdict, and shows what a single majority label throws away.
Policy and lawSame farms, five data sets, five different answers
Which data set a regulator picks can change which crops are told to cut their fertiliser.
Life sciencesAuthors pull their own mouth-cancer data paper
A 37-author team withdrew a dataset paper after finding bugs in its script, saying the reported figures may mislead.
AI and training dataEven coral experts disagree on about one label in twelve
A reef dataset built from 76 scattered sources shows what it takes to make labelled images fit for training.
AI and training dataFourteen experts label bias in 448 Hindi news articles
A small open data set shows how a language left out of the tools gets its first measuring stick.
AI and training dataThe most accurate post moderator is often not the fairest
A new test set shows that judging posts well and judging them fairly are different skills.
TechnologyTeaching an AI model to say it does not know
A new method nudges a few internal switches so a model declines to guess.
AI and training dataMachine-written tests catch more AI failures than human ones
A method that grounds test questions in real-world evidence makes chatbots fail more often on medical advice and self-harm.
Data valueNot every stale record is worth the cost of re-checking
A short paper shows that choosing which data to refresh by the task at hand beats refreshing blindly.
AI and training dataReweighting training data can shift harm to another group
A preprint finds a common fairness fix lowers missed cases for one group by raising them for another.
Data valueImage checks cannot prove where a medical scan came from
A test of synthetic chest X-rays at dataset intake says provenance has to be recorded, not guessed from pixels.
Life sciencesA curated set and shared test make protein-model building comparable
CryoAtlas filters public structure records into one training set with a fixed benchmark.
AI and training dataA text-reuse test set built only from open sources
Researchers release a bilingual set for testing plagiarism detectors, with every source passage openly licensed or in the public domain.
Life sciencesA tool that cleans health records for AI is put to the test
A new benchmark checks a chat agent that turns raw patient records into datasets ready for AI, and reports where it falls short.
Life sciencesA public trial registry trains a simulator that ignores the drug
A new test shows models built on registry data can score well without using the one field that matters.
TechnologyOne aviation safety dataset built from 128 sources
A new paper merges national accident records into one deduplicated, traceable corpus.
Physics and frontier scienceA detector's full record can tell more than the single number it is usually boiled down to
A new preprint finds that using the whole simulated record of a particle shower sharpens the energy reading by a factor of 1.7 to 1.9.