Google’s AI co-scientist uses agents to generate, critique, rank and refine research hypotheses, with scientists supplying goals, ideas and feedback. The approach brings AI into research design and makes the distinction between a promising proposal and an established finding especially important.
Experiments involving 1,401 participants show that human–AI feedback loops can amplify biases in perception, emotion and social judgement. Participants often underestimate the influence. The work extends attention from individual mistakes to how repeated interaction may reshape the people using a system.
Across more than 100 experiments, human–AI teams did not, on average, beat the better standalone performer. Content creation showed more promising gains than decision tasks. The findings invite attention to how work is divided and coordinated, rather than assuming that adding AI automatically improves a team.
METR reports that developers’ reluctance to work without AI and selective task submission weakened its follow-up productivity study. Concurrent agent work also complicated time measurement. As AI changes which tasks people attempt and how they work, established experimental designs may no longer capture the benefits reliably.
The authors propose five principles covering research goals, procedures, knowledge sharing and public communication under uncertainty about AI consciousness. Their proposal separates convincing appearances from scientific assessment and asks what commitments research organisations should make before the question is settled.
With client consent, Morgan Stanley’s Debrief produces meeting notes, action items and draft follow-up emails, saving notes to Salesforce. Advisors edit and decide what to send, placing AI inside an existing service workflow.
Define success as a business change, make ownership explicit, then decide which technology the work needs.
ERPBench brings pricing, production, procurement and finance into a shared simulation to examine how market conditions change agent performance.
Supplementary study · 11
Do AI gains become lasting organizational capability? This study links recent productivity gains to durable, firm-specific knowledge built through practice.
Core reading · 09
Valuations can influence investment and capital formation, rather than merely reflect productive capacity. Can temporary optimism leave lasting productive assets?