AI language models killed the Turing test: do we even need a replacement? (www.nature.com)

At a packed Royal Society event marking 75 years since Alan Turing’s imitation game, researchers argued that modern language models have essentially "killed" the Turing test — they can convincingly mimic human text but still lack the deep understanding the test was never designed to measure. Speakers including Anil Seth and Gary Marcus urged discarding the Turing benchmark in favor of targeted evaluations: tests that measure safety, robustness, grounded understanding and specific useful capabilities rather than whether a model can pass as human. The meeting stressed that the pursuit of a vague “AGI” goal narrows thinking about the kinds of systems society actually needs or should avoid. Technically, the panel highlighted concrete model failures that reveal gaps behind fluent output — inability to label elephant parts correctly, draw clock hands outside learned positions, or generalize beyond narrow training regimes — and pointed to the success of highly specialized systems like AlphaFold as a model for purpose-built evaluation. The takeaway for AI/ML practitioners is to shift toward domain-specific benchmarks, adversarial and out-of-distribution tests, and safety-focused metrics (alignment, interpretability, failure modes) that better capture real-world reliability and societal impact than the original Turing setup.