In August, I was facing an important deadline and felt reassured knowing that Codex could work alongside me through the night on a research related task. Unfortunately, that confidence did not last.
Without warning, Codex stopped and displayed:
This content cannot be displayed.
We take special care with requests involving biological research and applications that could pose security risks. Eligible researchers can apply for Trusted Access.
I then asked it to simply show whatever portion of the response was allowed. I received exactly the same warning.
What makes this particularly frustrating is that I reported essentially the same problem about a year ago, when legitimate scientific protocols were being blocked because they contained supposedly “actionable steps”:
One year later, I am encountering the same issue.
I have been an OpenAI subscriber for roughly three years and have also spent considerable time helping other users and researchers in this community. I use these tools for bioinformatics, genomics, sequencing workflows, teaching, scientific writing, and laboratory planning. These are ordinary parts of academic biomedical research.
I fully understand the need for safeguards around genuinely dangerous biological applications. What I do not understand is a system that appears unable to reliably distinguish legitimate academic research from requests that represent an actual biosafety or biosecurity risk.
The suggestion that researchers should seek “Trusted Access” also raises another concern. Access to useful scientific capabilities should not effectively become restricted to researchers affiliated with selected institutions or organizations while ordinary academics, particularly researchers working in universities and laboratories outside major research centers, receive blanket refusals.
This becomes particularly problematic for scientists in Brazil and other countries where access to computational infrastructure, specialized software, laboratory reagents, and institutional resources is already unequal.
Tools such as Codex and Rosalind could substantially reduce those barriers. Instead, excessively broad biological restrictions risk creating another one.
After encountering the same problem repeatedly over more than a year, I am seriously considering cancelling my subscription; because a research assistant that unpredictably refuses routine biology tasks becomes difficult to rely on when deadlines, experiments, teaching, and scientific projects depend on it.
I hope OpenAI revisits how these biological safeguards are implemented, particularly the distinction between legitimate scientific research and genuinely hazardous requests.