Where did the argument refuse an answer that was useful but too easy?
Eventually, the judging AI estimated there was perhaps an 80 percent chance that the speaker was human. Later, when I asked it to judge not the text but the agency directing the exchange, it raised the estimate to 90 or 95 percent.
The number had the comforting feel of an answer, which is one reason numbers are so often dangerous.
A 95 percent chance of a person is not a 95 percent person.
That was the hinge of the experiment. The old Turing Test asks whether a machine can convince a human that it is human. My experiment asked whether one machine, reading another machine’s performance of doubt, standards, self-scrutiny, and refusal, would conclude that there was probably a human behind it, while the actual human sat outside the frame arranging the encounter.
By then, the question had shifted. I was no longer only asking what the speaker was. I was asking what should follow when we cannot be sure.
Taffy Brodesser-Akner recently went to London for The New York Times Magazine to interview Tilly Norwood, a computer-generated character promoted by her creator as the world’s first AI actress. Brodesser-Akner sat across from a laptop and interviewed a synthetic performer as if she were a celebrity with a career, a craft, a persona, and a future.
Tilly had a face, a voice, a name, and a creator, Eline van der Velden, who described her as a tool, an experiment, a warning, and a new kind of artistic possibility. The reporter kept reminding herself that Tilly was just a computer because human beings are not built to sit calmly in front of a humanlike face, humanlike timing, and humanlike wit while treating the encounter as inert machinery.