What if machines had taste?
The Taste-Bench Project is an independent research initiative focused on understanding, measuring, and improving the taste-adjacent capabilities of AI.
The next frontier
Machines acquire new capabilities in waves. First they surpassed us at physical labour. Then at calculation. Then at complex tasks like coding. Soon AI will solve most well-defined technical problem.
We believe the next frontier will be “taste”: the ability to judge what is good, choose what matters, and create something that is truly novel and worth caring about.
Why it matters
More trustworthy chatbots.Current models agree too easily, praise things that don't deserve it, and confidently make poor suggestions while sounding very confident. It feels less like working with a trusted partner and more like being gaslit by someone who'd rather sound helpful than actually be helpful. Tasteful machines would care more about helping you make the right choices, even in subjective domains.
More autonomous agents.The industry is racing toward AI that acts on your behalf: booking, building, researching, deciding. But autonomy without taste is just fast and reckless. An agent that can execute a thousand tasks but can't tell which ones matter, or sense when something feels wrong, isn't useful. It's dangerous. Real autonomy requires knowing when to push forward, when to stop, and when the brief itself is wrong.
More creativity and innovation.Taste is what takes AI from productivity tool (automating tasks faster) to creativity tool (helping humans invent greater things). New art forms no single person could have conceived. Scientific breakthroughs in fields stuck for decades. Tools that don't just generate content but help people think more originally, rescue half-formed ideas before they die, and surface extraordinary talent the gatekeepers would never see.
What we do
We're starting with what feels most urgent: having conversations with people who know what taste looks like in practice (see our podcast), and building rigorous benchmarks to measure it in machines (see benchmarks). Then in the long term, we hope our lessons will inform the design of the next generation of AI systems.
Contact
If any of this resonates, write to us at hello@taste-bench.com.