A remote-work AI scored nearly 0.99—yet its authors warn employers not to use it
TabNet predicted a performance label across 100,000 public workforce records, but the dataset's provenance and representativeness remain too weak for real employment decisions.
The short version
- What happened
- The peer-reviewed study used 100,000 public workforce records and reported R² 0.988 on its main split; five-fold validation averaged 0.9723.
- Who it affects
- A workforce score could influence monitoring, promotion or dismissal. Without trustworthy labels and workplace-specific validation, a precise-looking output may give weak evidence unjustified authority.
- What remains unproven
- Dataset provenance and representativeness are insufficient for real-organisation claims
Keep in mind
Dataset provenance and representativeness are insufficient for real-organisation claims
Fairness checks cover only the recorded groups and errors
Prediction does not establish valid performance measurement or better employment decisions
No independent operational deployment was evaluated
Source analysis · 1 linked source. Primary analysed source date: 25 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
A high score survived several stress tests—but weakened as the test became more realistic
Researchers built an explainable TabNet model using 100,000 workforce records assembled from eight public data providers. The model predicted an employee-performance score from multidimensional workplace variables. On the main test split it reported R² 0.9880, mean absolute error 0.1223 and root-mean-square error 0.1549; five-fold validation averaged R² 0.9723 with a 0.0201 standard deviation.
The team also tested repeated random, group-aware, chronological and out-of-distribution splits. R² declined from 0.9618 in repeated random splits to 0.9483, 0.9326 and 0.8874 respectively. An ordinal version achieved 94.18% accuracy and macro-F1 0.9257. Subgroup error differences were generally small for the demographic groups assessed.[1]
The data gap matters more than the decimal places
For workers, the important question is not whether a model can reproduce labels in a compiled dataset. It is whether those labels reflect valid performance assessment, whether relevant groups are represented and whether a prediction changes pay, promotion or monitoring fairly. The paper cannot answer those questions.
The authors explicitly say it is far too early to connect the result to real organisational environments. A deployment case would need documented label creation, consent and lawful purpose, independent validation inside each workplace, threshold-specific errors, appeals and evidence that the system improves decisions without intensifying surveillance or discrimination.[1]
The distinction that matters
Why a near-perfect score is not deployment evidence
Each test asks a harder question than ordinary random splitting.
Random split
Predict records similar to training data
R² 0.9880
Chronological split
Predict later records
R² 0.9326
Out of distribution
Predict less familiar data
R² 0.8874