Brain-computer interfaces (BCIs) hold significant promise for enabling communication in individuals with severe disabilities. EEG-to-text models, which convert brain signals into written language, are at the forefront of this technology. Recent advancements in machine learning have enhanced the accuracy and speed of these models, yet their true effectiveness is still uncertain due to shortcomings in evaluation methods. A new study evaluates the performance of EEG-to-text models, emphasizing their learning capabilities from EEG signals versus merely memorizing patterns. The researchers introduced a method to compare model performance on actual EEG data and noise inputs. Results indicate that many models perform similarly or better on noise, suggesting potential memorization rather than genuine learning from brain signals. This finding underscores the necessity for improved benchmarking and evaluation practices in EEG-to-text translation. By addressing these methodological limitations, researchers can create more reliable systems that better utilize the potential of BCIs for effective communication.