Automated models often outperform humans at detecting deception but remain vulnerable to adversarial attacks—subtle alterations of statements (i.e., words or phrases) that preserve meaning but change the classification of a model. After training a DistilBERT classifier on 80 percent of the statements from a dataset of autobiographical truths and lies (Hippocorpus), humans and GPT-4o each rewrote 153 test statements (of the remaining 20%) up to 10 times, attempting to flip the model’s prediction. This can be understood as a paraphrasing attack, the statement is rewritten in a way that it means the same, but in an attempt to fool the classifier. Nearly 70% of paraphrased statements succeeded in changing the model’s prediction (i.e., lie to truth or truth to lie). While humans and LLM were similarly effective and efficient overall, humans induced a greater change in model confidence and in fewer iterations for truthful statements. This highlights a key vulnerability: Models can be tricked through benign changes without changing the underlying content.