I have talked about augmentation instead of replacement since 2015, and I tell it through chess.
In 1997 IBM's Deep Blue beat Garry Kasparov. The story the world took away was human against machine, and the machine won. If computers will only get better, why should a human keep practising? Kasparov drew a different conclusion. In 1998 he helped set up a new kind of tournament in León, Spain, where each player had a computer at their side. He later said it was as disturbing as it was exciting. Freestyle chess, or centaur chess, became its own discipline. For years after 2005, human and computer pairs beat both grandmasters and the best programs in blind matches. On the Elo scale the best humans top out around 2,900, the best programs reach about 3,500, and the best centaurs went higher. At the very top it gets fuzzy: some years the engines win.
A centaur is a system where a human and an AI work together, each doing what it does best. Human skills are creativity, empathy, flexibility, imagination, discernment and social sense. Machine qualities are arithmetic, consistency, precision, scale and speed.
Netflix made a centaur decision when it bought House of Cards in 2011. It used viewing data to choose the director, the lead actress and even the colour palette, a $100 million bet. Its head of original content described the mix as 70 percent gut and 30 percent data.
Other fields work the same way. A Formula One driver races with a team reading live data from every part of the car. Football added video referees to put more eyes on each decision. Researchers who built a model to restore missing text in ancient inscriptions found that historians alone got about 25 percent right, the model 62 percent, and historians working with the model 72 percent.
The most useful finding comes from a 2023 study by BCG with researchers from Harvard, Wharton and MIT, the same paper that named the jagged frontier. Consultants with below-average skills gained about 43 percent from using GPT-4. Those above average gained about 17 percent. Everyone gets a boost, but it is largest where you are weakest.
It also has a limit. In 2023 Amazon sold mushroom foraging guides that were written by AI. You still need someone who knows what is edible. Nature has many kinds of symbiosis, from mutualism to predation, and we get to choose which kind of centaur we build.
Try this
Don't ask AI to do what you already do well. It will probably be worse than you. List three things you are mediocre at, and use AI for those this week. That is where the gain is.
Since 2024
I spent 2025 writing down how I work as a centaur, from thinking to building. The result is Field Notes from a Centaur, seven chapters on daily practice.
Sources
- Garry Kasparov, Don't fear intelligent machines. Work with them (TED, 2017)
- Advanced chess
- Dell'Acqua et al., Navigating the Jagged Technological Frontier (2023)
- Assael et al., Restoring and attributing ancient texts using deep neural networks (Nature, 2022)
- The Guardian, Mushroom pickers urged to avoid foraging books on Amazon that appear to be written by AI (2023)