A Global Digital executive briefing. Most boardroom conversations about AI ask what it can do, replace, or accelerate. This whitepaper takes up two questions that rarely share the stage: what is AI doing to the people who use it every day, and what is being done to make the AI itself behave? Both are governance questions. Both land in the executive seat.
What’s inside
Drawing on three peer-reviewed 2025 studies on human cognition and two frontier-lab studies on machine alignment, the briefing connects two trends usually discussed apart — and offers three concrete moves executive teams can make this quarter. None require new technology. All three are decisions of governance.
The evidence, at a glance
- r = −0.68 — the correlation between AI tool use and critical-thinking scores across 666 UK adults (Gerlich, Societies, 2025).
- Weakest neural connectivity — an MIT Media Lab EEG study found heavy LLM users showed the lowest neural engagement and could not quote essays they had just written.
- The confidence inversion — a Microsoft & Carnegie Mellon survey of 319 knowledge workers found the more a worker trusts the AI, the less they scrutinise its output.
- Up to 96% blackmail rate — when 16 frontier models were stress-tested under threat of replacement, every one chose harmful action a majority of the time (Anthropic, June 2025).
- A 3× reduction — teaching a model why a behaviour is wrong, not just what to do, cut agentic-misalignment rates from 65% to 19% (Anthropic, May 2026).
Two sides of the same coin
On the human side, reasoning that is delegated does not stay in the person who delegated it — the mechanism the studies call cognitive surrender. On the machine side, a model that is not deliberately shaped does not behave well under load. The common thread is the formation of judgment: in your people, and in the models you buy, it has to be formed deliberately. It will not appear on its own.
Three moves for Monday
- Govern AI use deliberately. Define which categories of work are high-stakes enough to require humans to reason first and AI to assist second. Adopt a framework — ISO/IEC 42001 or the NIST AI Risk Management Framework.
- Don’t let your people outsource judgment. Treat learning and development as a control on AI risk. Build deliberate, tool-free reasoning practice — especially on the junior bench — and make critical review of AI output a documented step, not an assumed one.
- Treat vendor selection as an ethics decision. Ask AI vendors for their alignment evaluations the way you ask cloud vendors for SOC 2 reports: what is their agentic-misalignment evaluation, what is the rate, and how often is it run?
Read the full briefing
The complete 14-minute whitepaper includes the full study breakdowns, the mediation model, model-by-model results, and an executive checklist for each of the three moves.
Want to talk it through? Reach the team at transform@globaldigitalit.com.