Every Answering Machine Since Delphi Has Had the Same Three Problems

Any system built to answer questions on someone else's behalf runs into the same three problems, and the ancient world hit all three first. Not because the ancient world had computers. Because the problems are not about the technology. They fall out of the situation: a party that knows something, a party that needs to know it, and a channel between them that has to be paid for, staffed, and trusted. Change the channel from a mountain sanctuary to a phone tree to a chat window on a website and the three problems do not change shape. Only the plumbing does.

Set two things side by side: the small voice and video agents that now greet visitors on websites and try to answer whatever they ask, and the answering institutions of the ancient world. This is not a claim that ancient oracles are like modern software in any strong sense; a raving priestess and a language model have nothing structurally in common. The claim is narrower and more interesting: whenever a party stands between a question and an answer, the same three constraints show up, because they follow from being an intermediary, not from what the intermediary is made of. Delphi met all three centuries before anyone thought to write software, and so did the diviners of Babylon. Reading how they handled it is not nostalgia. It is the oldest case study in the trade.

One note first. Every ancient episode below is paraphrased from the events and the sources, never quoted from a translator's English; where a specific word matters to the argument, the underlying term is given rather than anyone's rendering.

The scarce resource was never the answer

Start with what it actually took to get a question in front of Delphi. The oracle did not sit open for business. In the earlier period the Pythia is said to have spoken once a year, on the god's own feast day; later, once a month, and not at all during the winter months when Apollo was traditionally away and the sanctuary belonged to Dionysus instead. A city that wanted an answer sent official envoys, theoroi, on a journey that could run to days or weeks depending on where they started. On arrival they paid a fee, offered a preliminary sacrifice, and waited to learn whether that day's omens even permitted a session to go ahead. Then they queued. Not everyone queued equally: certain states held promanteia, the right to consult ahead of the line, a privilege the Delphians granted as thanks for service to the shrine and that some cities were proud enough of to carve into stone on the monuments they paid for.³

Notice what was actually scarce. It was never the god's knowledge, which by definition of the institution was inexhaustible. It was slots: consultation days, priestly attention, a place in a line that a rival city might be standing ahead of you in for reasons that had nothing to do with the merits of your question. The Pythia herself, on the ordinary reading of the sources, was a local woman of Delphi who served for life and who processed one query at a time, in person, mediated by resident priests who managed the queue, interpreted the utterance, and handed the polished answer to the envoys to carry home.² Whatever happened in the adyton, the throughput of the whole operation was set by the humans running it, not by any property of the truth being dispensed.

Babylon ran a parallel system with the same bottleneck moved to a different place. Instead of a single physical shrine, Babylonian celestial and terrestrial divination depended on a trained specialist, the ṭupšar Enūma Anu Enlil, literally the scribe of the great omen series named for its opening words, who had spent years learning to read a corpus of thousands of recorded omens and the protocols for applying them to a given sky, liver, or oil pattern.⁷ The knowledge itself, once written onto tablets and copied into temple and palace libraries across centuries, was in principle durable and shareable. What was not shareable was the trained reader. A king or a provincial governor who needed an omen interpreted needed a specific kind of scholar in the room, one whose training, patronage, and access to the tablet series were themselves the product of years of court investment. The Assyrian royal correspondence that survives from the following centuries shows exactly this: scholars writing in to the king with readings, competing for his attention, and being consulted in a rough order of standing.⁸ Move the bottleneck from a queue at a shrine to a queue for a scholar's time, and the shape of the constraint is unchanged. The resource that ran out was access to the answering, not the answer.

This is the part of the pattern that modern systems like to claim they have abolished and have actually only relocated. A website with a chat widget on it answers at any hour, with no pilgrimage and no sacrifice, which is a real and considerable improvement. But the constraint has not disappeared; it has moved into the machine. Compute has to be provisioned, and provisioning it costs money, so somebody sets rate limits, tiers, and priority queues, the exact structure Delphi ran with fewer euphemisms. A free tier waits. A paid tier, or an enterprise contract, gets faster or more capable answers, and calls it a service level rather than promanteia, but the underlying arrangement, priority sold or granted while the base capacity stays finite, is the same arrangement the Chians paid for and had inscribed on the altar. Latency did not go away when the channel became electronic. It became a design parameter instead of a geographical fact, which made it easier to hide and no less real.

An answer specific enough to be useful is specific enough to be wrong

The second problem shows up as soon as the answer has to say something. Croesus of Lydia, weighing whether to attack the Persian empire, is said to have sent to Delphi and been told, in substance, that if he made war on Persia he would destroy a great empire.¹ He marched. The empire that fell was his own. When his kingdom was gone and, so the story runs, he sent to ask the god why it had misled a benefactor, the answer he got back was that the oracle had told the truth: a great empire had indeed fallen, and the fault for not asking which one lay with the man who never asked the follow-up question.¹ Read as prophecy, it is a shell game. Read as an interface design, it is closer to unavoidable: a statement built to be verified true under every branch of the outcome, with the work of picking the branch handed to the party who actually knew the local context the oracle did not have.

An answer specific enough to be useful is specific enough to be wrong, and the only way to be safe on both counts at once is to make the asker do the last mile of interpretation.

The Athenian consultation before Xerxes' invasion works the same lever from the other direction. According to the story, the first oracle sent the envoys home with nothing usable at all; they refused to leave and asked again, and the second reply granted Athens a wooden wall that alone would not fail, adding, almost as an aside, a line about the island of Salamis that could be read as either a blessing or a warning depending on which of two nearly identical words the god had actually used.¹ Athenian interpreters split. Some read the wooden wall as the old timber palisade on the Acropolis and prepared to stand a siege there. Themistocles argued the wall was the fleet and won the argument in the assembly, not at the shrine. Whatever happened afterward at Salamis, the deciding move was made in Athens, by people arguing over an ambiguous text that neither side could simply overrule, because both sides had agreed in advance to treat it as authoritative.

This is worth taking seriously as a design choice rather than dismissing as a fraud, because a fraud explanation does not actually cover the facts as well as the interface explanation does. A system that cannot verify the asker's private situation, the state of Croesus's army, the mood of the Athenian assembly, faces a genuine tradeoff every time it is asked something with real consequences attached. It can refuse, and lose the trust that comes from being useful. It can answer with false confidence, and lose everything the first time it is visibly, catastrophically wrong. Or it can answer in a form specific enough to feel like guidance and loose enough that the responsibility for the final call reverts to the person best placed to make it, who is also the person who will live with the consequences either way. Delphi picked the third option so consistently that the Greeks gave the god a standing epithet for it, Loxias, heard as slantwise or oblique, and did not treat the obliqueness as a defect to be apologised for. It was closer to the terms of service.

The limits of this argument deserve honesty, because the honesty matters more than the cleverness of the point. Every one of these stories comes to us through literary sources writing well after the events they describe, shaping episodes into memorable, often morally pointed narratives for readers who already knew how the story ended. Herodotus is writing history in the sense his culture recognised, not a transcript; Plutarch is writing philosophical dialogue centuries later, defending an institution he served as priest. The record we have is not a neutral log of queries and responses. It is a curated set of the stories worth retelling, which almost by definition means the ones that turned out to be strikingly, tellably ambiguous. The argument is that ambiguity was a real and recurring feature of how Delphi handled high-stakes questions, not that every recorded oracle was riddling, and not that the record proves the design choice rather than merely preserving the cases where it happened to matter.

Trust from position, not from a track record

The third problem is the one that outlasts the other two, because it is not about any single answer at all. It is about why anyone should believe the source in the first place, and Delphi's solution to it had almost nothing to do with being right.

Look at what actually surrounded the tripod. The Sacred Way climbing to the temple was lined with treasuries, small monumental buildings each paid for by a different city, standing where every other city's envoys would walk past them on the way to ask their own question. Cities sent theoric embassies not just to receive an answer but to be seen sending one, and to leave a dedication that recorded, in stone and bronze, that they had. Delphi was also the seat of the Amphictyony, a league of peoples from central Greece and Thessaly whose council actually ran the sanctuary's finances, guarded its neutrality, and could and did go to war, in the so-called Sacred Wars, against anyone judged to have profaned it.¹⁰ States used the shrine to arbitrate boundary disputes and sanctify treaties precisely because it belonged to no single city and every city needed it to keep belonging to nobody. None of that is a claim about the accuracy of any oracle. It is a description of a node that every relevant party had already agreed to keep using, which is a different and in some ways more durable kind of authority than being demonstrably correct.

A network position is a trust asset in its own right, and it can outlast accuracy by a very long way.

Joseph Fontenrose tried to test how much of the legendary Delphi survives contact with the actual record. He catalogued every response he could find and sorted them by evidentiary weight: genuinely historical, meaning attested close to the event; quasi-historical, real events reported only much later; legendary; and outright fictional.³ His conclusion was that the famous riddling oracles, Croesus and the wooden wall very much included, cluster in the weaker categories, while the oracles he judged solidly historical tend to be short and administrative: found the colony here, offer to this hero, proceed this way. Parke and Wormell, working the same corpus a generation earlier, were readier to credit the dramatic tradition of an oracle that actively shaped archaic politics.⁴ Scholars still argue over which picture is closer to the truth, and Fontenrose's own categories have themselves been challenged as too confident in places about what counts as sufficiently early or independent attestation. That dispute is not adjudicated here. But notice that it does not actually threaten the argument about trust and position; if anything it sharpens it. Even on the more sceptical reading, the ordinary daily business of the shrine, plain, low-risk, administrative answers to colonisation and cult questions, kept the sanctuary running and kept cities coming back for centuries, entirely apart from whether any particular dramatic oracle ever happened as told. The stories that made Delphi famous and the routine business that made Delphi useful were not quite the same thing, and both depended on the shrine's position rather than on a measurable hit rate nobody was keeping.

Another essay on this site covers how this same institution worked as a hub for the whole Greek network of cities, states that sent embassies, built treasuries, and consulted before founding colonies or reforming laws, in the essay on Delphi as an information network. The addition here is narrower: that trust of this kind, trust conferred by being the place everyone else already goes, is a genuinely separate asset from trust earned by being verifiably right, and a system can hold a great deal of the first with very little of the second for a surprisingly long time.

What this has to do with a greeter on a website

I build Vona, AI voice and video greeter widgets for websites. For anything that answers a stranger on a company's behalf, these three problems are still the whole job. Latency and access, because a visitor who waits too long, or who gets routed to a lower tier of capability without knowing it, leaves before any answer arrives. Ambiguity as a design choice, because an answer specific enough to be genuinely useful to a stranger, with no shared history and an unknown context, is specific enough to occasionally be wrong, and every system that has to answer strangers eventually has to decide, deliberately, how much interpretive risk it keeps and how much it hands back. And trust from position rather than a provable record, because a visitor deciding whether to believe what a small widget just told them is not running an audit; they are reading signals about whether this looks like the kind of thing worth trusting, which is a judgment about position and presentation as much as it is about any fact being correctly stated.

This is not a prediction about where any of this goes, and not a sermon. The defensible claim is smaller than that: the problem shape recurs because it is a property of the situation, not of the machinery. Delphi's plumbing was a mountain, a queue, and a woman who spoke once a month. Babylon's was a trained reader of a fixed omen series, consulted in order of standing, as argued in the piece on what the Babylonians actually built out of centuries of recorded observation. A modern website greeter's is a language model running behind a script tag. None of those are the same tool; but the ancient record is, among other things, the longest-running field trial anyone has run on the problem every answering system faces. The tools changed completely. The three problems did not move an inch. As the piece on what a new capability costs the person who receives it argues, every tool arrives with a bill attached; for anything that answers on someone else's behalf, the bill has had the same three line items since before anyone thought to write it down.

The oracle at Delphi is gone, its questions no longer worth the trip. The problems it was built to manage are not gone at all. They are simply being managed again, by something else, on a much shorter queue.