What's this about?
A paper Mollick highlights finds that human augmentation and task automation are not necessarily related: models that are very good at doing work are not always good at helping humans do work better. In the results, Opus and Sonnet performed strongly at doing tasks on their own but worse at providing assistance; GPT-5-Mini was good at both; Gemini models were stronger as assistants than as automators.
What is the difference between automating and augmenting?
Automating means the model does the task and hands you the result — success is the output being correct. Augmenting means the model works *with* you and success is measured by what you produce, which requires different behaviour: explaining reasoning, surfacing uncertainty, asking rather than assuming, and stopping at the point where your judgement is needed. A model optimised to complete tasks autonomously learns to push through exactly the moments where a good assistant would pause.
What's the community saying?
- Mollick's concern is structural: the commercial pressure is all toward making models good agents, and that may actively undermine human-AI cowork.
- The result is counterintuitive enough to be interesting — a smaller model beating larger ones on the assistance dimension.
- It implies the "which model is best" question is badly formed without saying which of the two jobs you mean.