In recent weeks, CDT has published not one but two first-of-their-kind reports that are shaping the conversation around some of the most important questions in AI governance.
The first explores the concept of “safety drift”, demonstrating how fine-tuning — even on innocuous or helpful data sets, like law and medicine — can have an unpredictable impact on safety guardrails in AI base models. Extensive fine tuning in one case might have no impact at all, but even limited fine tuning, in other circumstances, can lead chatbots to supply dangerous answers on topics such as health, suicide, and child sexual abuse. Our findings provide important context for apportioning responsibility between developers and deployers, and underscore the importance of sharing safety information across the supply chain.
Just last week, CDT’s Research Team released a new study exploring how dark patterns (like the controversial "infinite scroll” of some social networking apps) map onto experiences with AI chatbots. We examine how design choices impact chatbots used for social or personal interactions — including by children — and offer practical suggestions to guide responsible chatbot design. These findings provide AI developers and policymakers with evidence and practical recommendations at a pivotal time for chatbot design.
We’re also keeping a close eye on how AI is being used by our own government. In the second entry in CDT’s AI in Policing series, we examine the harms to communities from Shotspotter, an AI gunshot detection tool. After spending millions, some departments are rolling back usage due to ineffective results for public safety and serious ethical concerns over disproportionate deployment in predominantly minority communities. Meanwhile, CDT’s headline-grabbing polling released earlier this month revealed that an overwhelming majority of Americans are concerned about the misuse of personal data collected and stored by government agencies — a concern that becomes even more serious when AI is brought to bear on large sets of government data. Last month, in testimony before the House of Representatives Committee on Homeland Security, CDT’s Samir Jain emphasized the heightened security risks to sensitive data held by government entities and highlighted how advanced AI systems will exacerbate the risks of cyberattacks, particularly for under-resourced jurisdictions.
At the same time, CDT continues to push back on federal proposals that would block state legislatures actively working to protect their constituents from the harmful effects of AI. In recent analysis, we highlight key priorities for state legislation to establish guardrails for the responsible adoption of AI by public agencies and encourage legislators to prioritize risk management, AI governance, and transparency following legislative trends from 2025 state public sector AI bills. Along with the aforementioned policy priorities for public sector AI regulation, the CDT team highlighted legislative priorities for responsible AI adoption and use in K-12 education.
