---
id: "lib-2026-03-00-ai-part-1-lecture-slides"
type: "library"
title: "AI Part 1 lecture slides"
lang: "EN"
org: "MIT Economics"
source_type: "Report"
date: "2026-03-01"
week: "2026-W09"
url: "https://nosov.ai/library/lib-2026-03-00-ai-part-1-lecture-slides/"
original_url: "https://economics.mit.edu/sites/default/files/inline-files/14.662-Spring-2026-Lectures-Slides-10-AI-Part-1-Canvas_0.pdf"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "David Autor's MIT graduate labour economics slides curate the main experiments and models on AI and work with effect sizes and his own critical notes; expert teaching material that summarises others' research rather than new evidence."
summary_basis: "full-text"
original_minutes: "60"
tags: ["ai economics", "labor markets"]
---

# AI Part 1 lecture slides

Tier B (Useful, limited): David Autor's MIT graduate labour economics slides curate the main experiments and models on AI and work with effect sizes and his own critical notes; expert teaching material that summarises others' research rather than new evidence.

Summary based on the full text.

## In short

Lecture slides 10 for MIT course 14.662 (Spring 2026), "Some (Labor) Economics of Artificial Intelligence", by David Autor (MIT and NBER), dated March 11, 2026. The deck asks where AI substitutes for or complements labour, whether task models are enough, how organisations change, whether AI builds or erodes expertise, and what the evidence on job displacement shows. It combines formal task and assignment models with recent experiments. A recurring thread is the difference between automation, which removes expertise, and collaboration, which removes supporting tasks while expertise remains.

## Key numbers

- 0.548 SD to 0.139 SD: education-based productivity gap without vs with AI access (Cruces et al. 2026, N=1,174)
- 28.4% to 22.4%: adenoma detection rate of non-AI colonoscopies before vs after AI introduction (Budzyn et al. 2025)
- 11 of 15: endoscopists whose unaided detection rate fell after AI exposure
- d = 0.65: learning gain from contrastive AI explanations vs no AI (Bucinca et al., N=628)

## What it reports

- Computing cost fell about seven trillion-fold between 1940 and 2006 (Nordhaus); AI differs from traditional computing by handling tacit knowledge, in Polanyi's sense of "we know more than we can tell".
- Noy and Zhang (Science 2023): 453 college-educated professionals with randomised ChatGPT access; lower initial performers gained most.
- Cruces et al. (NBER WP 34851, 2026), N=1,174 adults in Argentina: the high-vs-low education gap falls from 0.548 SD to 0.139 SD with AI; treatment effect +1.24 SD for low-education vs +0.83 SD for high-education. Low-education users fully copy-pasted AI output more often (70% vs 60%).
- Ide and Talamas (2025, JPE): autonomous AI raises total output but benefits the most knowledgeable; copilot-only AI lowers output but helps the least knowledgeable. Autor's note: "formally elegant, profoundly reductive".
- Budzyn et al. (2025, Lancet), 4 centres in Poland, 1,443 non-AI procedures, 19 endoscopists: unaided detection fell from 28.4% to 22.4% (-6 pp, p=0.009) after AI tools arrived; 11 of 15 endoscopists declined.
- Bucinca et al. (N=628): contrastive AI explanations improved learning vs no AI (d=0.65) and vs one-sided explanations (d=0.35) without losing accuracy; one-sided explanations did not improve learning.
- Employment evidence: Brynjolfsson, Chandar and Chen (2025) on early-career headcount in exposed jobs, and Iscenko and Millet (2026) on job postings by AI exposure; Autor calls current conclusions "controversial".

## How they got it

A teaching synthesis. Autor presents task-model and assignment-model theory, then reproduces figures and headline statistics from randomised experiments, observational medical studies, payroll data and job-postings data, adding his own assessment of each.

## Limits

- Slides without lecture audio; some arguments are only outlined.
- Several cited studies are working papers; the endoscopy result is observational and has no external ground truth.
- Autor states that research on organisational design for AI is almost absent: "we know almost nothing about this so far".

## Why it matters for companies

The evidence that AI narrows performance gaps suggests the largest gains may come from less experienced staff, if the tools are designed as copilots. The endoscopy result is a warning that expert skills can decay when AI does the routine part, so firms need deliberate practice and backup capability. How explanations are designed affects whether employees learn or only depend on the tool. The "experience Catch-22" also raises a question for workforce planning: how juniors build judgement when AI does the grunt work.

Original: https://economics.mit.edu/sites/default/files/inline-files/14.662-Spring-2026-Lectures-Slides-10-AI-Part-1-Canvas_0.pdf