The claim: AI's real value is in making people better at their work, not in replacing them — and choosing between those two is the whole ballgame. It's the middle position between automation displacement and full-blown optimism, and it's the one with the most hard evidence behind it.
The core of it
- The "Turing Trap." Erik Brynjolfsson's argument that building AI to imitate humans aims at the wrong target. Imitation competes with people; complement makes them worth more. The trap is that imitation is the more obvious thing to build.
- It lifts the floor. The consistent finding is that AI assistance helps the least experienced worker most, and the expert barely at all. That's the opposite of how earlier computer technology worked, which mostly rewarded people who already had skills.
- Gains arrive late. The Productivity J-curve: output often dips before it rises, because the organization has to rebuild how work flows before the technology pays off.
- The technology is the cheap part. Process redesign, training and management changes are what actually capture the value, and they're slow.
- Codifying expertise. David Autor's more recent argument: if AI packages expert judgment so a mid-skill worker can use it, it could compress wage gaps instead of widening them.
Who argues it
- Erik Brynjolfsson — lead theorist. The Turing Trap, the Productivity J-curve, and the workplace studies that supply most of the real evidence.
- David Autor — the task framework, plus the strongest historical counterpoint to doom: roughly 60% of US employment is in job titles that didn't exist in 1940. New work appears; it just doesn't appear on schedule.
- Andrew McAfee — Brynjolfsson's longtime co-author on the machine-and-human-together argument.
- Nikhyl Singhal — the view from inside a big tech company. Senior product managers get amplified because taste, empathy and problem definition haven't been automated. "Craft still wins" — but note that this is the same evidence the displacement camp cites, because the junior rungs are what disappear.
- Peter Diamandis — the abundance version: every knowledge worker becomes many times more productive within a few years.
- Demis Hassabis — the most convincing case, and it's in science. AlphaFold didn't compete with biologists for the same task, it did something no human could. That's augmentation at its least ambiguous.
Who pushes back
- Daron Acemoglu — argues augmentation is mostly rhetoric covering a replacement reality, and that only a small share of tasks will be profitably automated in the next decade anyway, so the productivity story is oversold in both directions.
- Carl Benedikt Frey — augmentation is an outcome you have to fight for, not a default. Historically displacement comes first, and generative AI extends that pattern into white-collar work for the first time.
- David Shapiro — augmentation is the middle phase, not a resting place. If AI eventually competes on all four things humans sell, there's no stable position to augment into.
- Marc Andreessen — doesn't dispute augmentation so much as find it unambitious. Markets invent whole new categories of work, not just better versions of existing jobs.
Adjacent
- Tyler Cowen — augmentation is real but unevenly claimed. The people already good at directing their own work benefit most; the median worker sees a modest bump.
- Pope Leo XIV — asks about the quality of augmentation. Does the deployment leave the worker's judgment intact, or reduce them to approving what the machine produced?
The evidence
This is the thesis's strong suit. It has actual field studies, not just arguments.
- Brynjolfsson, Li and Raymond (2023). Customer-support agents with an AI assistant resolved about 14% more issues per hour. Novices improved most, roughly 35%. The best agents gained close to nothing.
- Comparable 10–40% gains measured in coding, writing and support work.
- The counter-evidence, which matters. Several of those productivity gains showed up as reduced headcount rather than higher pay. Augmentation for the people who stayed; displacement for the rest.
- Autor's "new work" data: about 60% of 2018 employment in post-1940 job titles.
- Agent tools are now blurring the line the thesis depends on. A copilot that finishes the task by itself isn't a copilot.
Open questions
- Does augmentation survive more capable agents, or turn into replacement as autonomy grows?
- Who captures the gains? The call-center study showed the firm capturing them, not the workers.
- Is the thesis stable, or is Shapiro right that it's a waypoint?
- How long is the J-curve lag this time, and is it shorter than in past technology waves?
- Does codifying expertise genuinely help mid-skill workers, or does it just make them easier to swap out?
All theories · clickbridge.com · Rich Price · last reviewed September 2026 · People linked here have their own pages at people.clickbridge.com