🤖 Lecture 9

Generative AI, Agents, and the Job Market

Byeong-Hak Choe

bchoe@geneseo.edu

SUNY Geneseo

November 9, 2026

🧠 From predictive AI to deep learning

📈 What machine learning did well

🏨 Hotels

Forecast demand from weather, events, and rival prices

📦 Amazon

Forecast orders, arrange shelves, run warehouse robots

📺 Your feed

Guess what to show you next

• Its limit: narrow tasks. It couldn’t handle language well.

0:00–2:00. The 33-minute plan includes brief section transitions. For a 30-minute class, define tokens on the prompt diagram and skip the separate token slide and the agent data-task example. In our course, forecasting tomorrow’s fine-particle concentration is another example of a narrow prediction task.

Mollick’s 2010s story: big firms with big data. Weak on situations outside its training data and on producing coherent text. That gap sets up Part 2.

Source: Mollick (2024), Chapter 1. Slides selected from DANL 101 · AI and Data Analytics. DANL 101 companion reading.

🕸️ What is a neural network?

• Input layer takes in the data.
• Hidden layers transform it.
• Output layer gives the answer.

2:00–4:00. Deep learning uses networks with many layers. Connect the cat example to classifying a satellite image.

Inspired by the brain, but only loosely. Nobody should leave thinking it works the way neurons do.

✨ Inside Generative AI

🛠️ Generative AI in an analyst’s work

Prepare

Draft a data dictionary. Organize text.

Analyze

Suggest R code. Explain an error.

Communicate

Propose chart titles. Critique a claim.

• The bottleneck moves from doing to directing and checking.

4:00–6:00. Generative AI produces text, images, and code. Ask where one of the three uses might fit in an environmental project.

Ask for a task students have already done in this course and what success would mean. AI can speed this up, but that doesn’t mean the code runs, the facts are correct, or the conclusion fits the data. Follow the instructions for the assignment or activity.

Source: Mollick (2024), Introduction.

🔮 How an LLM writes

🧱 What is a token?

cat playing extraordinary !

• The smallest unit of text an LLM reads and writes.
• A word, part of a word, or a symbol.
• Exact splits vary by model.

6:00–7:00. Optional on the 30-minute route.

🔁 From prompt to next token

7:00–9:00. If the token slide was skipped, define a token as a word, part of a word, or symbol.

Words become lists of numbers (embeddings), so words used in similar contexts get similar numbers. Position information is added so that “the dog chased the cat” differs from “the cat chased the dog.”

Change from last year’s slides: the 2017 Transformer had an encoder and a decoder because it was built for translation. Chat LLMs generate with a decoder-style stack, so this slide drops the two-box picture.

🏋️ Training, context, and tools

🏋️ Two phases of training

• Pretraining reads odd things: the web, books, even Enron’s emails.
• Models learn biases and errors along with facts.

9:00–11:00. The figure shows one way to train a chat model. Use the mention of books to introduce Project Panama.

Pretraining is unsupervised in Mollick’s telling: no labels, just text. It is also why training costs so much in chips and energy. Copyright and bias come back in Part 3.

Source: Mollick (2024), Chapter 1.

📖 Anthropic’s Project Panama

Washington Post reporting · Fair-use order in Bartz v. Anthropic

11:00–13:00. Anthropic acquired print books, cut off their spines, and scanned the pages for Claude’s training data. Distinguish how the books were obtained, how they were digitized, and how the text was used. In its June 2025 summary judgment, the court found that the training uses at issue and digitizing purchased print copies for a library were fair uses. The court denied fair use for building a library of pirated copies. Claims about those downloads were subsequently settled.

Ask who supplied the data and what rights came with it. For an environmental dataset, document its source, permitted uses, and transformations.

Image from Exhibit 43, p. 5, in Bartz v. Anthropic, as published in the Washington Post’s Project Panama reporting. See the June 2025 fair-use order and the Authors Guild’s summary of the final settlement.

🛡️ Working with AI, Responsibly

🎯 Alignment at a smaller scale

A narrow target

“Make the average recorded wait as low as possible.”

What if difficult orders are turned away?

The real goal

Serve customers well with reasonable waits.

Track service quality, and who is left out.

• Alignment: directing AI toward human goals and limits.

16:00–18:00. Ask what might go wrong if the system optimizes only for recorded wait times.

A small-scale analogy to the paperclip story: a measurable target can leave out what people value. Hypothetical scenario, not a claim about a deployed system. Adapted from the November 7 draft.

Source: Mollick (2024), Chapter 2.

📜 Working with AI: four rules

📜 Four rules for co-intelligence

1. Always invite AI to the table

2. Be the human in the loop

3. Treat AI like a person (but tell it what kind of person it is)

4. Assume this is the worst AI you will ever use

18:00–20:00. Apply the first rule within assignment permissions. Ask students to give the model a role, task, audience, and limits for reviewing an environmental chart.

Written in late 2023 and still a useful frame. Correction from last year’s slide: Rule 3’s full wording is “but tell it what kind of person it is.”

Source: Mollick (2024), Chapter 3.

🧑‍⚖️ Responsible use of AI

🧑‍⚖️ Responsible use: four questions

20:00–22:00. For this course, follow the assignment’s AI and disclosure rules. Ask what the class would need to check before sharing a partner’s dataset.

This is the frame for the next several slides. It is my synthesis of Mollick’s Chapters 2 and 3, the syllabus, and the evidence below, so treat it as a teaching device. Tell students this is what they should remember from today.

🔄 AI agents

🔄 What is an AI agent?

22:00–24:00. An agent can run an R script, read the error, and choose its next step.

Mollick predicted in late 2023 that the next phase would be semi-autonomous AIs given a goal. The early agent he tried was BabyAGI. Definition to remember: an agent has a goal and tools, and it decides its next step from what it sees.

Source: Mollick (2024), Chapters 3 and 6.

🧰 An agent on a data task

The agent…	You check…
reads sales.csv	Right file? All the rows?
writes and runs dplyr code	Does it answer the question?
fixes its own errors	What did it drop or change?
makes a chart	Are the axes and labels honest?
writes a summary	Can you trace every number?

24:00–26:00. Optional on the 30-minute route.

Connect to their work this semester. Ask which checks students already use in their environmental data projects. Illustrative workflow, not one specific product. Ask for the script and chart alongside any report, so there is something to check.

🔍 Andrej Karpathy’s AI gap

The funnel uses Karpathy’s back-of-the-envelope estimates.

26:00–28:00. Karpathy’s observation is about people’s understanding of what current systems can do. People see different capabilities depending on the tools they can access and the tasks they try. In code and data work, runnable checks make some results easier to verify. The funnel’s widths and counts are illustrative. Ask what an agent with access to files and tests can show that a chatbot demonstration might miss.

Image based on Karpathy’s October 11, 2026 post. Readable copy of the post.

💼 AI and jobs

📦 A job is a bundle of tasks

• AI takes tasks. Whether it takes the job is a separate question.

28:00–29:00. Ask which tasks a wildfire analyst could delegate and which require their judgment.

Spreadsheets changed accountants’ tasks without ending accounting. Jobs also sit inside systems such as liability rules, customers, and managers, which change slowly. Telephone operators are the counterexample: when direct dialing arrived, operator jobs fell 50 to 80 percent, and the most experienced took the largest long-term earnings hit. So it does sometimes happen. The example sort is mine.

Source: Mollick (2024), Chapter 6.

📊 What does each study measure?

Measure	What the evidence shows
Exposure	The ILO estimates that 1 in 4 workers globally are in occupations with some tasks potentially affected by generative AI.
Productivity	At one firm, customer-support workers resolved 15% more issues per hour on average after receiving AI assistance.
Employment	The Yale Budget Lab’s October 2026 analysis found no clear AI-related shift in the overall U.S. labor market using its measures.

ILO exposure index · Productivity study · Yale tracker

29:00–31:00. Ask for the unit in each study. The ILO scores tasks and maps exposure to occupations, which estimates potential exposure. The QJE study uses a staggered rollout among 5,172 customer-support workers at one firm and measures issues resolved per hour. The gains were larger for less experienced and lower-skilled workers. Yale’s October 14 update uses national labor-force survey data through September 2026. Its overall estimates can coexist with differences for a particular group.

ILO’s occupational exposure study, Brynjolfsson, Li, and Raymond’s productivity study, and The Budget Lab at Yale’s labor-market tracker.

💼 But watch young workers

• From November 2022 to June 2026, employment among workers ages 22–25 fell 11% in the most AI-exposed occupations and rose 10% in less exposed occupations, a gap of 21 percentage points.

ADP payroll study

31:00–33:00. The comparison is descriptive. Discuss how employment can fall for a subgroup while the overall labor market shows little movement.

The chart connects two endpoints. The highest two exposure quintiles fell by about 11% and the lower three grew by about 10%, a difference of 21 percentage points in growth. Check the relative comparison with students. Dividing 89 by 110 gives about 0.81. The study finds that slower hiring accounted for most of the gap. Experienced workers show no comparable gap. Same paper’s Fact 1: no widespread, economy-wide displacement. A monthly dashboard is at digitaleconomy.stanford.edu/project/indicators/canaries-dashboard/.

Source: Brynjolfsson, Chandar, and Chen, Canaries in the Coal Mine? (rev. Aug 12, 2026).

Then transition to Classwork 7.