Centaurs and Cyborgs on the Jagged Frontier
AI's ability is a wall with an invisible outline: 758 consultants did far better on the tasks inside it and worse than no AI at all on the one task outside, because they could not see which side they were on.
Read the original ↗The most quoted piece of evidence that AI helps knowledge workers, and the part nobody quotes is the part worth reading. Eighteen tasks chosen to sit inside the model's ability, where everyone with AI did better. One task built to sit outside it, where everyone with AI did worse and none of them noticed. The line between those two sets of tasks is invisible, and that is the finding.
Included with a plan
This discussion is included with a plan. The first three lessons of every track are free, with their boards and videos. A plan unlocks every lesson, board, video, lab and paper for one price.
Related on chalk
The same idea as a board, lab, video or lesson.
- BoardHow Do You Know It's Goodan eval is a way of finding the wall before your users do
- BoardThe Harnessthe code around the model that decides which side of the wall a task is on - a centaur, written down
- BoardEvals for Appsthe task the model fails confidently is the one no eval was written for
- BoardRouting and Cascadesrouting by task difficulty is the division of labour the essay calls a centaur, done by a program
- BoardCognitive Biasesautomation bias - trusting the confident answer - is the mechanism behind the 19-point drop
- BoardThe Edge of Abilitywhere your own frontier is, and why working just past it is how the skill leveller works
- BoardLearning and Expertisewhat falling asleep at the wheel costs the person, not the task
- LabThe 99 percent testa tool that is right most of the time and the one case where trusting it is exactly wrong
- VideoEvals: how you know an agent is goodhow you know an agent is good, which is the same question as where the wall is
- VideoThe confidence curve: the chart everyone quotes was never measuredconfidence and competence coming apart, on the human side of the desk
- VideoAgents from first principles: build the loop yourselfthe loop a cyborg runs by hand and an agent runs on its own
- LessonEvaluating Models (how do you *know* it's good?)Uses the board How Do You Know It's Good
- LessonPrecision and Recall (where to set the spam filter's bar)Uses the board How Do You Know It's Good
- LessonThe Harness (everything around the loop that the model never touches)Uses the board The Harness
- LessonThe Returns Desk, For Real (a real model, a real harness, and a dress rehearsal that fails)Uses the board The Harness
- LessonEvals for Apps (the test suite for a program that is a string)Uses the board Evals for Apps
- LessonRouting and Cascades (not every request needs the big model)Uses the board Routing and Cascades
- LessonThe Sale Made at the Bottom - the behaviour gapUses the board Cognitive Biases
- LessonCognitive BiasesUses the board Cognitive Biases
- LessonThe Room That Agreed With the Wrong LineUses the board Cognitive Biases
- LessonWhen the Mind HurtsUses the board Cognitive Biases
- LessonDeliberate Practice (how expertise is actually built)Uses the board The Edge of Ability
- LessonThe Alarm That Won't Switch Off (stress after failure)Uses the board The Edge of Ability
- LessonLearning & ExpertiseUses the board Learning and Expertise
- LessonThe Mind Being BuiltUses the board Learning and Expertise
- LessonMeasuring a PersonUses the board Learning and Expertise
Further reading
Read the original for the full argument and its evidence.
- Centaurs and Cyborgs on the Jagged Frontier ↗Mollick's own write-up, 16 September 2023, twenty minutes. Read the two sections named Inside and Outside the Jagged Frontier together, because reading the first without the second is how this study got misquoted for three years.
- Navigating the Jagged Technological Frontier ↗The study itself, by Dell'Acqua, McFowland III, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon and Lakhani, published in Organization Science in 2025 after circulating as Harvard Business School Working Paper 24-013. The three conditions and the design of the nineteenth task are the parts the essay compresses hardest.
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity ↗METR, July 2025. Sixteen experienced developers on their own repositories were 19% slower with AI and believed they had been 20% faster. The strongest single piece of evidence that the BCG headline does not travel to work people already know how to do.
- The Cybernetic Teammate ↗The 2025 follow-up by overlapping authors, 776 professionals at Procter and Gamble. One person with AI matched a two-person team without it, which is where the research programme actually went after 2023.
- The Impact of Generative AI on Critical Thinking ↗Microsoft Research and Carnegie Mellon at CHI 2025, surveying 319 knowledge workers. The mechanism behind falling asleep at the wheel, described by the people it happened to.