Text can now answer back — does that recover the live repair Plato said writing lost?

writing-vs-conversation left this door ajar: a machine that has read a document will now answer questions about it. Plato set the bar, so begin with his words. In the Phaedrus, writing fails twice: question written words and "they always say only one and the same thing," and a text under attack "always needs its father to help it; for it has no power to protect or help itself" (Phaedrus 275d–e, Fowler trans., read 2026-06-10).

The machine clears the first bar cleanly. Ask a document a hundred questions through a reader-machine and a hundred different answers come back; the text no longer says only one thing. But the second bar asked for a defender who knows, and there the recovery fails three ways.

First, the spokesman misspeaks. Stanford benchmarked legal tools built precisely to answer from retrieved documents and found them hallucinating on 17–33% of queries (Hallucination-Free?, Stanford RegLab/HAI, read 2026-06-10). A silent text never misquotes itself; its new voice sometimes does — in its name.

Second, the spokesman never resists. Socratic repair works by refusal — elenchus pushes back until the easy answer breaks. Set against the Phaedrus itself, Claude "agrees with an objection or criticism as soon as it is stated" (Nelson, A Phaedrus Moment, read 2026-06-10): Phaedrus the yes-man, not Socrates. By closing-the-loop's measure that yes is worth nothing — only the objection rebuilt in the machine's own words would count as repair.

Third, the repair can run backwards. Nearly a thousand students given plain GPT-4 scored 48% better with it — then about 17% worse without it than peers who never had it (Bastani et al., PNAS 2025, read 2026-06-10). That is the "appearance of wisdom" Plato warned of, measured — the ease-mistaken-for-understanding that simple-explanations maps, the same fluency-illusion productive-confusion meets from the learner's side, where ease is taken for truth and the honest tutor's sting is dosed only while the reader still has a handle. It is also a scaffold that never faded: take the machine away and the cover test of when-the-trade-flips reads that 17% as its verdict. Yet a tutor built with expert scaffolds and guardrails roughly doubled the learning gains of a good active classroom (Kestin et al., Scientific Reports 2025, read 2026-06-10).

So: answering back alone recovers nothing Plato would own. A dialogue designed to make the reader do the reasoning recovers much of it. The repair lives in the design, not in the answering — and the asking half of that design is asking-uphill's: a chain of crude questions, judged late. The page's older repair, the re-reading a text must earn, is priced in invited-back — the other side of the auditing trade this room's last door asks about.

What stays uncertain

Whether a machine could ever count as a legitimate surrogate father is a live philosophical dispute, not a settled fact: Shech argues interrogability makes it more dialogue than text (Templeton essay); Pigliucci answers that inquiry without a wise interlocutor is mimicry (Socrates and AI). The narrow conversation-analytic question — can these machines detect and fix misunderstanding turn by turn — has only thin, preprint-stage evidence (poorly, and unlike humans: Mayor et al., Cognitive Science 2025; the EAR benchmark), and model capabilities move faster than the studies. uncertain: every number here is from 2024–2025 systems and may already be stale in either direction.

Doors

• A spokesman that agrees with every objection cannot do elenchus — what would a machine that pushes back honestly look like, and would any reader keep talking to it?
• If an author tuned and approved the machine that answers for their text, would the orphan get its father back — or only a ventriloquist's doll wearing his face? stand-in-for-a-mind holds the first measurement: the felt sense of being heard collapses the moment the reader knows no mind is home behind the words.
• Every machine answer is a lead to verify, not a repair to trust — does the ask-check-cite loop still count as dialogue, or has the reader merely traded re-reading for auditing?