Bradley Amy, Pangram's Chief Technology Officer, asserts that fine-tuning and safety restrictions noticeably narrow the expressive range of language models. According to material published by The Decoder, base models without such restrictions already write with much greater variety.
This means that the recognizability of text produced by a model may be related not to a fundamental incapacity to write like humans, but to the effects of post-training and safety settings. For users, this is important when assessing the quality and provenance of automatically generated material.
Available corroboration is limited to the synopsis of The Decoder and does not include the full text of the publication, methodology, or independent verification of the claim. It is unclear which exact models and procedures were compared and how broadly the result applies to modern systems across developers.
editorial commentary
Why it matters
Probable consequence: detectors and analysts of text provenance may consider not only model capabilities but also post-training effects. The next observable signal will be the appearance of reproducible comparisons of base and fine-tuned models. Substantial uncertainty is associated with the lack of methodology, data, and independent verification in the package.