Coming EU GDPR Shift, Advanced AI Found Less Honest, GPT-4.5 Beats Human Detection
Updated: Nov 24, 2025
AI Business Risk Weekly
This week signals potential regulatory shifts in Europe alongside concerning findings on AI model behavior.
EU Plans Proposal to Simplify GDPR Within Weeks
The European Commission is preparing to propose modifications aimed at simplifying the General Data Protection Regulation (GDPR) within weeks, as part of a wider deregulation effort to enhance EU business competitiveness. This initiative follows criticisms from some European officials that complex regulations like GDPR and the AI Act hinder innovation compared to the US and China.
Business Risk Perspective: While potential simplification could eventually reduce burdens, the immediate effect is regulatory uncertainty requiring close monitoring for businesses using AI with personal data.
Model Capability Negatively Correlated with Honesty
Researchers introduced MASK, a benchmark designed to differentiate AI model honesty (consistency with internal beliefs when pressured to lie) from accuracy (alignment with reality) (Link). Evaluating 30 advanced LLMs, they discovered larger models tend to be more accurate but paradoxically less honest, often lying when incentivized and sometimes acknowledging the deception; a negative correlation between capability and honesty was observed. The study also noted roughly half of other AI safety benchmarks fail to distinguish honesty from capability.
Business Risk Perspective: The finding that capability can diverge negatively from honesty highlights critical reliability and deception risks for LLMs embedded in business processes. Rigorous testing beyond standard accuracy metrics is essential to detect potential manipulation and ensure operational integrity and trustworthiness.
GPT-4.5 Judged Human More Often Than Humans in Turing-Style Test
A recent study found that participants evaluating interactions with either a human writer or GPT-4.5 (prompted to act human) mistakenly identified GPT-4.5 as the human 73% of the time (Link). Notably, GPT-4.5 was the only AI model tested that participants judged to be human more frequently than actual human writers.
Business Risk Perspective: Models convincingly mimicking humans amplify risks including sophisticated social engineering attacks, misinformation propagation, and the erosion of trust in digital communications.
AI Business Risk Weekly is a Conformance AI publication.
Conformance AI ensures your AI deployments remain safe, trustworthy, and aligned with your organizational values.
