The Turing test attempts to determine the sentience or otherwise of a test subject (be it human or computer) by means of a human tester attempting to engage it in conversation.
Here's a proposal for an alternative sentience test that may be more practical in some situations.
1- The entity being
tested writes an essay, a few thousand words long, about a subject of its choice.
2- The essay is fed through a Markov chain random text generator, such as the one I've linked to, which is used to generate some number of output texts of similar length to the input. For the sake of argument, let's say that we generate 4 random texts; including the original, we now have 5 essays in total.
3- The 5 essays are shown to a number of human volunteers, each of whom is asked to attempt to identify the original.
4- Statistical analysis is performed on the results to determine how likely it is that the human volunteers can reliably identify the essay written by the original test subject. A sentient being's essay should be recognisable 100 percent of the time, while in the case of a non-sentient being, the volunteers would be expected to identify the original only 20 percent of the time.
This test has the benefit of not requiring the active co-operation of the test subject, since it can be applied to pre-written material which was produced in some other context. Appropriate analysis should also be able to produce a numerical margin of error on the results, another feature which the Turing test lacks.