Bioinformatician Highlights Frustrations with Overly Restrictive AI Safety Filters in Research
A bioinformatics researcher shared real-world examples of proprietary LLMs like Claude falsely blocking routine scientific tasks, including file transfers for viral genomic analysis and basic questions on phage-host interaction prediction. In contrast, open and unaligned models like DeepSeek executed the requests without unnecessary refusals. This highlights a growing conflict between commercial AI safety alignment and legitimate scientific inquiry, where false positives stall academic research in biology and computer science. It underscores the critical necessity of open-weights and unaligned local models for specialized research communities. The user reported that commercial guardrails flagged benign terms like 'exfil' in script output logs and blocked queries regarding viral host infections as bio-hazards. Open models provided straightforward technical assistance, such as recommending Graph Convolutional Networks (GCNs) over CNNs for predicting phage-host interactions.
## BACKGROUND
Safety alignment in commercial LLMs aims to prevent malicious misuse, such as pathogen synthesis or cyberattacks, but often relies on overly broad classifiers that produce high false-positive rates for legitimate scientific terms. Open-source tools like detectEVE are standard bioinformatic pipelines used to identify endogenous viral elements in genomic assemblies. Additionally, machine learning approaches such as Graph Convolutional Networks (GCNs) are widely researched for modeling biological interactions between bacteriophages and host bacteria.