MSD InternationalMSDINTERNATIONAL← Home

MSD / Methodik / RESEARCH EDITION

Neue Werkzeuge, neue Auswahlverzerrung

METR describes a follow-up experiment in February 2026. Evidence of faster work is difficult to interpret due to selection of participants and tasks as well as problems with time measurement.

Jetzt eintauchen ↓RECHERCHESTAND 12.11.2026 · FAKTEN + EINORDNUNG + SZENARIEN
RESEARCH NOTE / 2026Methodik

KI & Agenten

External research · MSD classification
Keine Behauptung eines eigenen Versuchs.

01 / EINORDNUNG

What this finding means.

Grenzen des Nachweises

The authors consider a stronger acceleration to be plausible, but describe the data on the size of the effect as weak evidence.

Our derivation for a pilot operation

Renew experimental design after technical leaps. Anyone who rejects certain tasks without AI changes the sample examined.

This derivation is an editorial recommendation from MSD. It is not a further result of the study cited.

Source and version

METR · Methodenupdate 24.02.2026 · Publikations-/Bezugsjahr 2026. Recherchiert am 12.11.2026.

Read primary source ↗

External research, not an MSD customer reference. We will only publish our own experiments with their setup, date, measurements and documented failed attempts.

02 / LINKS

Think ahead. Keep checking.

AI & Agents / Contrary finding / 2025

When AI makes work take longer: METR's contrary finding ↗

In a randomised trial with 16 experienced developers and 246 tasks, early-2025 AI tools increased completion time by 19%. Participants perceived their benefit much more favourably.

METR · Randomised trial
KI & Agenten / Grundlage / 2017

Attention: Language can be processed in parallel ↗

The Transformer paper presents an attention-based architecture for sequence processing. It is a methodological building block of later language and multimodal models.

Vaswani et al. · Attention Is All You Need