Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

consequence-aware evaluation

Imagine testing a self-driving car only in empty parking lots; it might pass every test there, but that doesn't tell you if it's safe on a real highway. Consequence-aware evaluation means judging an AI system not just on whether it gets the right answer or executes a task correctly, but on what happens next, the real-world cost of its mistakes, whether that's a medication error, a robot dropping a heavy object, or a misleading financial report.

Across the papers, a clear pattern emerges: standard evaluation metrics, like accuracy or isolated safety scores, often miss serious failures. Adding user context, operational consequences, or future physical interactions reveals that systems thought to be safe or effective can in fact cause harm. Whether in health conversations, robot control, or multi-agent finance, the gap between local success and downstream consequence is large and systematic, and closing it requires explicitly measuring what happens after the AI's output.

Papers
33
Released code
1
First seen
Aug 2026
Latest
Nov 2026

30 papers in the last two months, against 3 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1ChatGPT's health answers are high-risk for one in five real users

    Real ChatGPT health conversations in four Global South countries received high-risk responses about one-fifth of the time, with user context exposing many additional harms.

    Oct 2026
  2. 2AI agents’ clicks marked wrong can work a third of the time

    Many clicks that GUI-grounding benchmarks reject are actually correct because they produce the same action as the annotated box.

    Oct 2026
  3. 3Coding agents fail most long chains of dependent tasks

    Coding agents solve at most 18.5% of long-horizon multi-step software engineering tasks.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

agentic evaluationMeasuring AI agents' performance in complex tasks, highlighting failures in planning, monitoring, and real13 togetheragent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.9 togetherbenchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.9 together

Who is working on it

Tencent 3Carnegie Mellon University 2

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.