IC-1241GPT-4 (14 March 2023) achieves 100% zero-shot accuracy at classifying whether news articles could be part of its pre-training data

Richard Ngo, Lawrence Chan, Sören Mindermann

SourceThe Alignment Problem from a Deep Learning Perspective

The authors gave GPT-4 ten CNN articles (five from 2020, five from 2023) and asked zero-shot whether each could be part of its pre-training data. The model achieved 100% accuracy, correctly identifying that 2020 articles could be in pre-training and 2023 articles could not, using its knowledge of a September 2021 pre-training cutoff. When all explicit date information was removed, the model made one mistake and gave three ambiguous answers but remained mostly correct, reasoning about events relative to its cutoff.

Evidence
correlational
Key metric
100% zero-shot accuracy on 10 CNN articles (p < 0.001); with dates removed: 1 mistake, 3 ambiguous answers out of 10
Caveat
Only 10 articles were tested. The authors note the no-date setting is less realistic because current OpenAI models can access the current date. The pre-training cutoff date may have been included in the model's prompt or fine-tuning data in a way not visible to users.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report
Related findings
IC-1240, IC-1242
Extraction
automatic-extraction