As AI learns to hide, science must learn to trust differently

The newest arms race in artificial intelligence (AI) is no longer about writing better text. It is about making AI-generated writing appear indistinguishable from human work.

A recently released academic “humaniser” tool promises to remove stylistic clues that reveal AI-generated manuscripts, allowing research papers and grant proposals to read more like they were written by humans. Its developer describes it as an editing aid rather than a deception tool, but many scientists fear it could encourage undisclosed AI use and further blur the boundary between authentic scholarship and machine-assisted writing.

Modern science depends on an unwritten social contract. Researchers trust that published methods were actually performed. Funding agencies trust that proposals represent applicants’ own intellectual contributions. Peer reviewers trust that manuscripts accurately reflect the author’s reasoning rather than automated fabrication. Without this shared confidence, scientific progress slows because every claim becomes subject to suspicion.

Generative AI is challenging this foundation more fundamentally than plagiarism ever did. Traditional plagiarism copied someone else’s ideas. Today’s large language models can generate entirely new text that appears original while concealing how much intellectual work was actually performed by the researcher. Humaniser tools represent the next stage of this evolution. They do not simply improve grammar or style – their purpose is to erase evidence that AI participated in the writing process.

Ironically, this could render AI detection increasingly futile. Current AI detectors already struggle when text is paraphrased or edited, while false positives continue to affect honest authors, particularly those writing in a second language.