A contested question with a practical answer.
THE PHILOSOPHICAL POSITION
Genuinely debated. Reasonable people disagree about what understanding means and whether these systems have it.
THE PRACTICAL POSITION
They do not behave like something that understands.
They fail in ways that understanding would prevent: confidently asserting contradictions, inventing citations, missing obvious implications.
WHAT THEY DEMONSTRABLY DO
Produce text that is statistically consistent with their training data and your input.
That is sufficient to be useful and it is not understanding in any ordinary sense.
WHY IT MATTERS
If you believe it understands, you trust it inappropriately.
If you treat it as a very capable pattern-matcher, you use it well.
THE USEFUL MENTAL MODEL
Something that has read enormously and retains the shape of what it read, without a reliable grasp of what is true.
WHAT THIS CHANGES
You verify. You provide context. You take responsibility.
THE HONEST SUMMARY
Whether it understands is interesting. Whether you should check its output is not in question.
