It wasn’t a “roast” like Dan Robitzski of Futurism says it was, but in February, Harish Natarajan, a former championship college debater, won the audience vote over Project Debater, an IBM program designed to respond to its opponents in a debate and crystalize and summarize the issues in summation. The debate was about education subsidies. Project Debater was for them, Natarajan against.

It’s likely that Harish “won” the debate because he was able to better contextualize, rhetorically drive home, and make comparisons in his final speeches—the real ethos-meets-logos-type factor of good debating, as much an art as a science. Meta-analysis and rhetoric are both inexact—in form as well as content—so an experienced debater does more than just generate and counter information.

The video is worth watching: Project Debater made eloquent quotes, provided evidentiary support, answered arguments on-point, and even uttered the phrase familiar to debaters, “the benefits outweigh the disadvantages.” But that’s a stock phrase. It’s not nuanced comparison. Nuanced comparisons (including things like strategic concessions, admitting the other side is right about something in order to win a larger point) require abstract and metaphorical thinking; not much, but enough.

As Mindy Weisberger writes: “In a neural network, deep learning enables AI to teach itself how to identify disease, win a strategy game against the best human player in the world, or write a pop song. But to accomplish these feats, any neural network still relies on a human programmer setting the tasks and selecting the data for it to learn from. Consciousness for AI would mean that neural networks could make those initial choices themselves.” And, subjective experience is part of what’s required to do those things.

There are signs that machine learning has the capacity to do this, but little of that was on display during the debate. Stanislas Dehaene and colleagues list “global availability,” the relationship between cognition and the object of cognition, and “self-monitoring,” obtaining and processing information about oneself, as components of consciousness in thinking beings. Both of those attributes would help a debater-AI unit make meaningful, contextually appropriate comparisons between arguments, as well as discern strategic concessions (an unreflective computer probably “thinks” it’s winning every point in a debate).

For a few more years, at least, humans are safe in debates and a few other spheres of public life.