Spot the problem

“Because a transformer model trained on a curated medical-text dataset reached 92% accuracy on a multiple-choice licensing benchmark, it follows that the model truly understands clinical reasoning and can be safely deployed as an autonomous diagnostic agent in any hospital setting.”

“Across the ImageNet, COCO, and Open Images datasets, a recent audit found that images of people from several regions are heavily under-represented compared with North American and Western European images. It also found that object categories common in lower-income households are mislabelled by widely used vision classifiers far more often than similar objects from higher-income households.”

“Because one state-of-the-art language model needed hundreds of MWh of electricity to train in a single reported setup, it follows that every deep-learning project — regardless of model size, task, or hardware — is environmentally indefensible and should be abandoned immediately.”

“Since one commercial face-analysis system was shown to misclassify the gender of darker-skinned women far more often than lighter-skinned men on a single benchmark, it follows that machine learning as a field is fundamentally unable to produce fair systems, and that no algorithmic intervention could ever meaningfully reduce demographic disparities.”

“The claim that current code-generation models systematically reproduce insecure patterns from their training data — including hard-coded credentials and SQL-injection-prone string concatenation — is supported by controlled studies of model completions on security-sensitive prompts. It is further supported by national survey data showing that the average job tenure of professional software engineers has declined over the past decade.”

“Although a large body of work has shown that tiny, almost invisible changes to inputs can reliably cause state-of-the-art image classifiers to misclassify, and that current certified-defence methods only offer narrow robustness guarantees, practitioners who deploy such classifiers in safety-critical perception pipelines should keep treating their outputs as fully reliable and need not add any input validation or monitoring.”