Original analysis. AI reliability. 18 min read.
Does AI Know When It Is Wrong?
What confidence, calibration, semantic entropy, and human judgment reveal about whether a language model can recognize its own mistakes.
Read the full articleIndependent AI research, made useful
We read the research, separate evidence from ambition, and show where human judgment still matters.
Our premise
AI is most useful when it expands human judgment, agency, and possibility.
Latest articles
Clear explanations of what the research supports, where it stops, and what to do with it.
View all articlesOriginal analysis. AI reliability. 18 min read.
What confidence, calibration, semantic entropy, and human judgment reveal about whether a language model can recognize its own mistakes.
Read the full articleArticle 02. Research review. 5 min read.
Can agents improve their own prompts, tools, orchestration—and eventually learn how to improve that process itself? A critical review of an ambitious framework that has not yet been empirically tested.
Read the reviewCurated research library
Fifteen independent briefings. Each shows what the source answers, what remains open, and where the research points next.
01
What AI changes about jobs, skills, productivity, and who gets to do valuable work.
Work & economy. Source: One Useful Thing, 2026.
A practical argument for AI agents that increase human agency instead of quietly taking it away.
Read our briefingWork & economy. Source: Anthropic, 2026.
A careful attempt to measure actual AI use at work rather than treating every technically possible use as adoption.
Read our briefingWork & economy. Source: OpenAI, 2026.
Usage evidence suggests AI is moving tasks across occupational boundaries, not merely automating a fixed job description.
Read our briefing02
When AI strengthens learning—and when a convenient answer quietly removes the learning.
Learning & education. Source: One Useful Thing, 2025.
Why friction, effort, and productive struggle still matter when a model can produce the answer instantly.
Read our briefingLearning & education. Source: OpenAI, 2026.
A proposal to measure whether AI produces lasting knowledge, not just a good interaction or completed assignment.
Read our briefingLearning & education. Source: Google DeepMind: The Podcast, 2025.
A conversation about personalized tutors, teacher judgment, and the choices that determine whether classroom AI helps.
Read our briefing03
How models fail, how we notice, and why confident language is not evidence.
Reliability & safety. Source: OpenAI, 2025.
A training-and-evaluation explanation for why models often guess instead of admitting uncertainty.
Read our briefingReliability & safety. Source: Anthropic, 2025.
Why a visible chain of thought should not automatically be treated as a faithful audit trail.
Read our briefingReliability & safety. Source: Google DeepMind, 2025.
Human strategies can reveal failures that automated safety tests miss, especially in large generative systems.
Read our briefing04
Where AI is becoming a serious research instrument, from proteins to proofs.
Science & discovery. Source: Google Research, 2025.
A multi-agent research system that generates, critiques, ranks, and revises hypotheses with scientists in the loop.
Read our briefingScience & discovery. Source: Google DeepMind, 2025.
How accurate protein-structure predictions changed research workflows—and where experiments still matter.
Read our briefingScience & discovery. Source: Google DeepMind, 2026.
A reasoning system positioned as an expert collaborator for difficult proofs, conjectures, and scientific problems.
Read our briefing05
The human choices hiding inside supposedly automatic systems.
Society & human agency. Source: Anthropic, 2025.
A large qualitative study of people’s hopes and fears, from professional excellence to time freedom and loss of control.
Read our briefingSociety & human agency. Source: The Gradient, 2025.
A case for evaluating AI futures by human and social flourishing, not capability or economic growth alone.
Read our briefingSociety & human agency. Source: IEEE Spectrum, 2025.
Why deliberate pauses, approvals, and escalation points can improve autonomous systems rather than merely slow them down.
Read our briefingHow we work
Our briefings are independent summaries, not substitutes for the originals. We state the questions a piece answers, identify what it leaves open, and connect those gaps to primary research.
About our approachThe useful newsletter
Get concise, evidence-aware notes for people who want to use AI thoughtfully—without letting it run the show.
Subscribe free