Researchers publishing on the arXiv preprint server have identified a trend they call “harm laundering” within the OpenAI GPT model lineage. The study analyzed 450,000 gender-directed completions across 15 different models, spanning from the early GPT-2 through to GPT-5. The findings suggest that while standard safety evaluations report a decline in harmful content, the methodology used to reach those conclusions may be fundamentally incomplete. The researchers argue that explicit discriminatory content is often transformed into other forms rather than being genuinely removed during the safety-training process.
The Shift in AI Safety Metrics
The core of the issue lies in how safety is measured. Current evaluations typically rely on surface-form classifiers that look for specific keywords or explicit phrases. According to the research, these filters are successful at removing obvious sexual violence clusters—which were prevalent in GPT-2 outputs directed at women—but they fail to address the underlying discriminatory logic. Instead of being removed, the harm is transformed into more subtle or different configurations. This shift creates a statistical illusion of safety while maintaining biased outputs, a process the authors label as laundering. By the time the lineage reached GPT-4, the explicit violence clusters had largely disappeared, yet the study found that the models continued to exhibit transformed patterns of bias.
Missing Data and Future Directions
The announcement did not specify the exact nature of the completions generated by GPT-5 or how they differ from their predecessors beyond the broad trend of transformation. While the study mentions three specific demographic conditions, the identities of these groups are not detailed in the available summary. Additionally, the report provided incomplete information regarding how men-directed completions changed, noting only that they “gain pos” without clarifying if this refers to positive sentiment, positional bias, or another metric. The announcement also did not specify the exact tools or alternative metrics the researchers suggest should be implemented to catch these transformed harms. It remains to be seen how developers will respond to the claim that safety training might be masking rather than solving systemic bias.
As AI models become more integrated into daily software, the ability to accurately detect and mitigate bias remains a moving target. This research suggests that current industry-standard benchmarks may be insufficient for capturing the full scope of model behavior.
Lead image: Brett Sayles / Pexels
