AI Integration Begins (2020)

Two shocks arrived together. In 2020 the pandemic moved a huge share of knowledge work into the home and proved, to employers who had insisted otherwise, that the office was a choice. In the same season, large language models stopped being a research curiosity. OpenAI's GPT-3 was shown in 2020. Image and text generators reached the public in 2022. By 2023, drafting, summarizing, coding, and drawing had a competent assistant that did not need to be hired.

The pandemic reorganized where work happens. The models began reorganizing what the work is. Neither change has finished.

The home as a workplace

Remote and hybrid arrangements emptied downtowns on some days and loaded home internet connections with the working day. A set of jobs appeared just to make that tolerable: collaboration-tool administrators, workplace-experience managers trying to invent a reason to come in, and a boom in the people who outfit, clean, and deliver to houses that were now offices. Middle managers had to supervise people they could not see, and a visible minority of them responded with surveillance software. The monitoring tools are a job category of their own, sold as productivity and experienced as distrust.

The split was brutal and geographic. People who could work through a screen went home. People who moved goods, cared for bodies, or stocked shelves became essential, a word that did not reliably come with hazard pay. The decade made the class line of contemporary work obvious: some jobs are a laptop, and some jobs are a place you have to be.

What the models touched first

Generative systems are best at the producible middle of knowledge work. First drafts, routine code, standard illustrations, meeting summaries, customer-service replies, and the sort of analysis that reassembles existing text. Occupations heavy with that middle, junior copywriting, routine graphic production, entry-level programming tasks, basic customer support, felt the pressure earliest. The senior work of deciding, checking, and taking responsibility did not automate neatly, which is why the first new roles were editors of the machine's output.

Prompting, evaluation, and workflow design became skills faster than they became stable job titles. Companies hired AI leads, forward-deployed engineers, and trainers to wire models into existing processes. They also hired, often through contractors, the annotators who label data and rate outputs, a global piecework force that looks a great deal like the moderation workforce of the previous chapter. The glamour of the model and the labor under it should be described together.

A beginning, not a verdict

It is too early to list this chapter's extinctions with the confidence we can list the typing pool. Adoption is uneven, quality is uneven, and many announced replacements have turned out to be rearrangements. What is already true is that a large class of tasks no longer requires a beginner to spend a day producing the first version. The beginner role was where people learned. If the first version is free, the ladder's bottom rung is the thing at risk.

The museum's bet, visible in the future exhibits, is that the loss of the rung produces new work rather than only less work. The next entries on this timeline take that bet seriously, and they label it as a projection. 2020 is the last chapter we can write as history.

AI Integration Begins

The home as a workplace

What the models touched first

A beginning, not a verdict

Work that grew

Work that lost its place

Companies of this chapter

Largest

New at the time