0727

AI Insights & Research

Working paper | Surveys and administrative dataNBER·

Work can change before earnings do

Danish surveys linked to administrative records show task reorganization around content generation, AI oversight and integration. Difference-in-differences estimates find no average earnings or recorded-hours effects larger than 2% within two years of ChatGPT’s launch. These early, context-specific results do not imply that future labor-market impacts will remain small.

Source
Research report | Platform observation and measurement frameworkAnthropic Research·

Measuring AI’s economic impact at the task level

The report describes AI use through task complexity, skills, purpose, autonomy and success, incorporating success rates into efficiency estimates. Tasks within one occupation can differ substantially, so usage counts alone cannot measure automation. Many metrics rely on model assessment of platform conversations rather than causal measurement of economy-wide productivity.

Source
Technical announcement | Robotics system demonstrationsGoogle DeepMind·

From instructions to coordinated physical action

DeepMind presents Gemini Robotics 2, combining visual-language understanding with whole-body control, dexterity and coordination. Demonstrations extend AI assistance into physical environments. Evidence mainly comes from developer demonstrations and evaluations; it does not establish universal reliability, safety or commercial returns in uncontrolled workplaces.

Source
Journal paper | Retrospective probabilistic forecast evaluationNature·

AI weather forecasting needs probabilities, not just one answer

GenCast produces ensembles of possible weather trajectories up to 15 days ahead. Retrospective evaluations showed improvements over the comparator on many weather, cyclone-track and wind-power targets. Probabilistic information can support risk decisions, but retrospective skill does not guarantee accurate prediction of every extreme event or operational outcome.

Source
Journal paper | Empirical workplace rolloutThe Quarterly Journal of Economics·

AI in customer support: who benefits from shared expertise?

A staggered rollout covering 5,172 support agents found about 15% more issues resolved per hour with AI assistance. Less experienced and lower-skilled workers benefited most; experienced staff saw smaller speed gains and some quality declines. Evidence from one firm does not establish economy-wide employment or wage effects.

Source
Journal paper | Preregistered controlled behavioral experimentNature Human Behaviour·

AI persuasion and personal data: what the corrected study shows

In controlled debates involving 900 participants, personalized GPT-4 was more persuasive than the human baseline. A 2026 correction clarifies that its direct advantage over non-personalized GPT-4 was not statistically significant. The study therefore cannot isolate a personalization benefit, nor establish long-term persuasion effects on real platforms.

Source
Working paper | Global occupational-exposure measurementInternational Labour Organization·

Occupational AI exposure is not a job-loss forecast

The ILO combines task data, worker assessments, expert input and model scoring to refine occupational exposure estimates. Around one in four workers holds a job with some generative-AI exposure, with clerical work highly exposed. The index maps potentially affected tasks rather than forecasting layoffs; reorganization and full replacement are different outcomes.

Source
Annual report | Interdisciplinary evidence synthesisInternational AI Safety Report·

Governing advancing AI under incomplete evidence

This synthesis examines general-purpose AI capabilities, risks and mitigation, separating documented harms from uncertain future risks. It highlights an evidence dilemma: premature interventions may fail, while waiting for certainty can leave serious risks unmanaged. A research synthesis does not make every discussed scenario an inevitable outcome.

Source
Journal paper | Randomized screening non-inferiority trialThe Lancet·

AI-assisted mammography: evaluating follow-up outcomes

In a Swedish randomized trial involving roughly 106,000 women, AI-supported screening improved sensitivity, maintained similar specificity and reduced reading workload. Follow-up interval-cancer rates met non-inferiority criteria, but the between-group reduction was not statistically significant. This evaluates a specific workflow, not mortality reduction or fully autonomous screening.

Source
Journal paper | Educational field experimentPNAS·

Better assisted performance does not always mean better learning

In a high-school mathematics experiment, both a standard GPT interface and a tutoring variant improved assisted practice. After tool removal, the standard-interface group performed worse independently, while tutoring safeguards largely mitigated this loss. Design affected short-term learning in this setting, not necessarily long-term outcomes across all subjects.

Source
Stay curious. See you in the next read.

Join the conversation after approval.

Comments, replies and messages are reserved for approved, signed-in collaborators.

  1. Submit your collaborator profile
  2. Wait for profile review
  3. Sign in after approval to participate
Already a member? Sign inBecome a collaboratorThis is a flow preview. Account sign-in and approval checks are not connected yet.