← Wiki

Centaur and Cyborg Work

ai-and-jobsfuture-of-worklabour-marketai-and-cognition

Two Models of Collaboration

Mollick identifies two distinct patterns in how effective workers integrate AI into their tasks. These are not personality types — they are working modes that skilled practitioners switch between depending on the task.

Centaur work maintains a clear division between human and machine, like the distinct human torso and horse body of the mythical creature. The worker decides which tasks fall inside The Jagged Frontier (hand to AI) and which fall outside it (do yourself). A data analyst might choose the statistical approach themselves, then let AI produce the visualizations. In the BCG study, Centaurs did their strongest work themselves and delegated frontier-interior tasks to the AI.

Cyborg work blurs the boundary. Human and AI efforts intertwine at a granular level — initiating a sentence for the AI to complete, editing AI output mid-stream, feeding AI responses back as prompts. The work product is neither human nor AI but a deeply integrated hybrid. The division happens within tasks, not between them.

The Task Categorization Framework

Mollick proposes five categories for sorting work tasks relative to AI:

CategoryDefinitionExample
Just MeAI is not useful or should remain human (ethical, personal, or capability reasons)Creative writing with personal voice, parenting, value-laden decisions
DelegatedAssigned to AI with careful checking; tedious or time-consuming for humansExpense reports, email sorting, scheduling, paper summarization
AutomatedLeft entirely to AI without human review; requires very low error toleranceSpam filtering, code compilation, high-frequency trading
CentaurStrategic division: human does some tasks, AI does othersHuman chooses analytical approach, AI generates charts and graphs
CyborgDeep intertwining: human and AI pass work back and forth within a single taskWriter drafts paragraph, AI suggests alternatives, writer synthesizes

These categories are not static. As AI improves, tasks migrate: Just Me tasks become Centaur-eligible, Delegated tasks become Automated. The frontier shifts, and the categorization must shift with it. Workers who mapped the frontier six months ago may be operating on stale assumptions.

The Reskilling Dimension

The Centaur/Cyborg framework has direct implications for workforce development. The skills that make someone an effective Centaur are different from those that make an effective Cyborg.

Centaur work requires meta-cognitive judgment: knowing what you’re good at, knowing what AI is good at, and making the allocation decision well. Autor calls this the foundational expertise workers need — enough domain knowledge to know when AI output is trustworthy and when it isn’t (see Expertise Democratization).

Cyborg work requires fluid integration skills: the ability to prompt effectively, edit AI output critically, and iterate rapidly. This is a newer competency with no clear analogue in pre-AI work. The closest parallel might be the relationship between a skilled editor and a writer, compressed into a single person operating at high speed.

Both modes require understanding The Jagged Frontier. Without that understanding, Centaurs hand off the wrong tasks and Cyborgs accept bad output.

Secret Task Automation

Mollick documents a troubling organizational dynamic: workers who discover effective AI workflows often keep them secret. Three reasons converge:

  1. Organizational policy. Companies that ban or restrict AI use push workers to personal devices and shadow tools. The ban doesn’t stop usage — it stops disclosure.
  2. Value depends on secrecy. AI-generated work is judged differently when people know AI produced it. Workers who reveal their AI use may find their output devalued.
  3. Training your replacement. A worker who automates 90% of a task and tells management may see 90% of their department downsized. Silence is rational self-preservation.

The result: organizations lose access to their most valuable AI innovations because the incentive structure punishes disclosure. Shadow AI and Organizational Enablement documents the same dynamic from the organizational side. Workers who have the deepest Centaur and Cyborg experience are the least likely to share what they’ve learned.

Productivity Evidence

The early evidence on AI-augmented work consistently shows 20-80% productivity improvements on professional tasks. The BCG study found 40%+ quality improvement on consulting tasks. Noy & Zhang found 40% time reduction on writing tasks. Brynjolfsson et al. found 14% productivity gains for customer service agents, with novices reaching experienced-agent performance in 3 months instead of 10.

The pattern across all studies: AI as equalizer. The lowest-performing workers gain the most. The performance distribution compresses. This has labour-market implications beyond individual productivity — if the gap between novice and expert narrows, the premium for elite expertise declines. See Bounty and Spread for the macroeconomic framing and Expertise Democratization for Autor’s analysis of what this means for the middle class.

Organizational Implications

Mollick argues that organizational systems — hierarchies, processes, incentive structures — were designed around human limitations. The org chart emerged from 1850s railroad management. The assembly line emerged from Ford’s insight that humans were good at simple, repetitive tasks. AI changes the underlying capabilities, which means the systems built around those capabilities must change too.

Companies that respond to AI productivity gains by cutting headcount capture only a fraction of the potential value. Companies that maintain their workforce and redirect freed capacity toward expanded output should dominate competitors who simply do the same work with fewer people. The larger opportunity, as The Expanding Work Frontier argues, is work that was never done before — not doing existing work cheaper.