# mollick-real-ai-agents-work-oneusefulthing-2025-09-29

## Veille

Ethan Mollick - AI Agents and Real Work - Economic Impact - OpenAI Study - Research Replication - Future of Work - One Useful Thing

## Titre Article

Real AI Agents and Real Work: The race between human-centered work and infinite PowerPoints

## Date

2025-09-29

## URL

https://www.oneusefulthing.org/p/real-ai-agents-and-real-work?utm_campaign=post&utm_medium=web

## Keywords

AI Agents, real work, economic impact, OpenAI, AI capabilities, transformation of work, research replication, agentic AI, future of work, AI adoption, task automation, AI productivity, scientific research, human-AI collaboration

## Authors

Ethan Mollick, Professeur à la Wharton School, University of Pennsylvania

## Ton

**Profile:** Academic-accessible | First-person testimonial | Analytical-forward-looking | Intermediate-expert

Mollick adopts his characteristic professor-communicator style: he grounds a forward-looking thesis (AI agents crossing the threshold of economically relevant work) in concrete empirical evidence (OpenAI study, METR measurements) and in his own experimentation (replicating an economics paper with Claude Sonnet 4.5). The title, opposing "human-centered work" and "infinite PowerPoints," crystallizes the central tension in a memorable phrase. The tone is measured: neither catastrophism nor naive enthusiasm, with sustained attention to limitations ("jagged" capabilities, absence of real agency, risk of mediocre content overproduction). Target audience: executives, researchers, and professionals seeking to understand the real impact of AI agents on work.

## Pense-betes

- OpenAI study: experts (14 years of average experience) vs. AI on realistic 4-to-7-hour tasks, blind evaluation
- Humans still win, but narrowly; recent models are progressing fast
- AI's main weakness: output formatting and instruction-following (improving rapidly)
- Prediction: the next generation of models should surpass human experts on average
- AI replaces tasks, not entire jobs; "jagged" capabilities (excellent here, deficient there)
- Mollick's experiment: Claude Sonnet 4.5 autonomously replicates a sophisticated economics paper (reading, file sorting, STATA → Python conversion)
- Cross-validation: spot-check by Mollick and re-replication by GPT-5 Pro
- What takes humans hours is accomplished in minutes → potential lever against the scientific replication crisis
- METR study: the length of tasks AI can accomplish autonomously grows exponentially (from GPT-3 to GPT-5)
- Agents lack agency in the human sense: humans define objectives and boundaries
- Three productive uses: automating routine tasks, augmenting human capabilities, creating new opportunities
- Central risk: overproduction of low-value content ("infinite PowerPoints") if organizations lack imagination

## RésuméDe400mots

Ethan Mollick argues that AI agents have crossed a critical threshold: they are now capable of performing economically relevant work. He first draws on a recent OpenAI study comparing AI models to human experts (14 years of average experience in finance, law, retail) on realistic tasks designed to take 4 to 7 hours, evaluated blind by a third group of experts. Result: humans still win, but narrowly, with margins that vary by sector. The most recent models are progressing rapidly, and AI's main weakness — output formatting and instruction-following — is improving fast. Mollick predicts that the next generation of models will surpass human experts on average. He qualifies this, however: AI replaces tasks, not entire jobs, and its capabilities remain "jagged," excellent on some tasks and deficient on others.

To illustrate the value of this real work, Mollick recounts his own experience: he asked Claude Sonnet 4.5 to replicate a sophisticated economics paper involving multiple experiments, providing the full text and the replication data archive. Autonomously, the model read the paper, sorted the files, converted the STATA code to Python, and methodically verified all the findings, including complex interactions. Mollick spot-checked the results and had the replication re-replicated by GPT-5 Pro. What would take experienced researchers hours is accomplished in minutes — a major avenue for addressing the scientific replication crisis, by enabling large-scale verification that was previously impossible.

This leap in capability stems from improved model accuracy: even small gains substantially reduce failures across long chains of tasks, and recent "thinking" models incorporate self-correction. The METR study shows that the length of tasks AI can accomplish autonomously has grown exponentially since GPT-3.

Mollick nonetheless notes that agents lack agency in the human sense: they do not decide their own objectives, and humans must define goals and boundaries. He identifies three productive uses: automating routine tasks (reports, presentations), augmenting human capabilities (research replication), and creating new opportunities. The symmetrical risk is the "infinite overproduction of PowerPoints": unimaginative organizations using agents to generate ever more low-value content. The future of work will depend on our ability to conceive of uses that complement human capabilities rather than mechanically replacing them.

## GrapheDeConnaissance

- Ethan Mollick —publie→ Real AI Agents and Real Work (DOCUMENT, 0.98)
- Ethan Mollick —travaille_chez→ Wharton School (ORGANISATION, 0.98)
- OpenAI —publie→ étude capacités agents IA (DOCUMENT, 0.95)
- agents IA —permet→ travail économiquement pertinent (CONCEPT, 0.93)
- Ethan Mollick —utilise→ Claude Sonnet 4.5 (TECHNOLOGIE, 0.95)
- Claude Sonnet 4.5 —permet→ réplication d'articles de recherche scientifique (CONCEPT, 0.92)
- GPT-5 Pro —soutient→ réplication réalisée par Claude (CONCEPT, 0.88)
- METR —mesure→ longueur des tâches accomplies de façon autonome par l'IA (CONCEPT, 0.9)
- agents IA —améliore→ longueur des tâches autonomes (croissance exponentielle) (CONCEPT, 0.9)
- agents IA —remplace→ tâches (pas emplois entiers) (CONCEPT, 0.88)
- réplication automatique —résout→ crise de réplication scientifique (CONCEPT, 0.85)
- Ethan Mollick —recommande→ augmentation des capacités humaines par l'IA (CONCEPT, 0.87)
- Ethan Mollick —prédit→ un risque de surproduction de contenu à faible valeur par les agents IA (AFFIRMATION, 0.85)
- Ethan Mollick —prédit→ la prochaine génération de modèles dépassera en moyenne les experts humains (AFFIRMATION, 0.88)

---
Canonical: https://www.thekb.eu/en/fiches/mollick-real-ai-agents-work-oneusefulthing-2025-09-29/