AI gets more 'meh' as you get to know it better (www.theregister.com)

Wiley’s preview of its 2025 ExplanAItions survey of 2,430 researchers finds a clear split: adoption has jumped—84% now use AI tools (up from 57% in 2024)—but confidence in AI’s ability to outperform humans has plunged (from 53% in 2024 to under a third in 2025). Early adopters remain more bullish (believing AI outperforms humans in 59% of use cases), yet worries have risen across the board: 64% cite inaccuracies/hallucinations (vs. 51% last year), 58% flag security/privacy concerns (up from 47%), and ethical and transparency worries also increased. Most practical use today centers on manuscript prep, writing assistance, documentation and literature review, and researchers report tangible productivity gains (85% say AI made them more efficient; 77% say it boosted output), though only 48% feel it helps them think critically. The takeaways matter for AI/ML development and deployment: researchers are recalibrating expectations as real-world limits surface, reinforcing findings from recent CMU and MIT studies that many AI agents and GenAI pilots fail to deliver reliable results or measurable ROI. Adoption will hinge less on raw enthusiasm and more on improved reliability, privacy/security, transparency, and institutional support—57% of respondents said lack of guidelines and training hampers uptake. Notably, a majority (57%) would permit a truly capable autonomous agent to act on their behalf, signaling readiness for more advanced tools if performance and safeguards improve.