Some persist to this day. Chiliad entered Chapter 7 liquidation in 2014. In 2026, a live public-sector procurement schedule still presents it as software you can buy.

The rabbit hole that brought me to the story is an epic worthy of Homer… and ironically involves Homer, Achilles, the Iliad, and a series of search errata. But that’s another story. Let’s get to what we found. AI’s errors are widely known and highly categorized if not well understood but the thing that we neglect to discuss often when discussing AI errors is the errors the previous systems and humans make as well, and we recently saw a ghost of one in the system. Chiliad Inc. (an intelligence company, founded by Ghislaine Maxwell’s sister, a whole *other* story) filed for Chapter 7 bankruptcy on August 6, 2014. Its assets were subsequently sold, and the bankruptcy case closed in 2016.

In January 2026, however, Chiliad (a word that means a group of 1000) was apparently still available for purchase, according to a US procurement company.

Carahsoft’s current pricing schedule for a national public-sector purchasing contract lists “Chiliad Inc.” as a software manufacturer. Buyers are promised a discount of 1.02 per cent from Chiliad’s list price.

The contract is active through December 2027. The pricing schedule was updated on January 27, 2026.

There is just one problem: the company entered liquidation more than eleven years earlier.

This is not evidence of a secret software operation. It is evidence of something more ordinary—and potentially more consequential for artificial intelligence.

A record retained its authority after the reality it described had disappeared.

This is what a hallucination looked like before neural networks.

The dead vendor

Chiliad was not an imaginary company.

Founded in the late 1990s, it developed technology for searching structured and unstructured information across distributed databases. Its customers and collaborators reportedly included the FBI and US intelligence agencies.

It was the kind of company that could reasonably appear beside major data-integration and analytics vendors in a government technology catalogue.

But Chiliad did not simply become inactive or stop updating its website. It filed a voluntary Chapter 7 petition in the Eastern District of Virginia, a form of bankruptcy intended for liquidation. The bankruptcy docket records the filing, the sale of assets and the eventual closure of the case.

Some of Chiliad’s intellectual property may have survived. A 2021 biography of co-founder Christine Maxwell said that another company, Techtonic Insight, owned the IP.

But that is not the company named in Carahsoft’s schedule.

The schedule says “Chiliad Inc.”

The live contract

Carahsoft is a major US government technology distributor and contract aggregator. Its contract with Region 4 Education Service Center, made available nationally through OMNIA Partners, allows public agencies to acquire technology products under pre-negotiated terms.

The current contract, R240303, began on January 1, 2025. Its initial term runs until December 31, 2027, with renewal options extending through 2029.

In the manufacturer discount schedule updated in January 2026, Chiliad Inc. appears with a 1.02 per cent discount.

This is not merely an archived webpage or an abandoned vendor profile.

The contractual documentation says the listed prices establish which manufacturers’ product lines, services and warranties are available from Carahsoft. In its 2024 response to the procurement solicitation, Carahsoft also said it could fill orders for the proposed solutions 100 per cent of the time because software and services do not present conventional inventory problems.

That makes Chiliad’s inclusion more than a harmless historical footnote. It appears in a current contractual exhibit purporting to describe products available for purchase.

The paper trail

The error did not suddenly appear in 2026. It has been carried forward for years.

The identical discount is revealing.

It suggests that nobody reconsidered whether Chiliad existed, had a current product catalogue or could fulfil an order. A vendor record was simply preserved and repeatedly emitted into new documents.

The document changed. The knowledge inside it did not.

This is not evidence of a sale

A pricing schedule is not a transaction ledger. Chiliad’s presence does not prove that any government agency recently bought its software. It does not prove that Carahsoft knowingly tried to sell a product it could not supply.

It could be the result of a spreadsheet copied from an earlier contract, a forgotten CRM record, an automated document generator, an ETL pipeline or a master supplier database without an effective process for expiring dead entities.

There is also no public evidence that generative AI created the schedule.

That is precisely the point.

The behaviour we now associate with AI hallucination existed long before ChatGPT. Bureaucratic and commercial information systems have always been capable of producing authoritative-looking statements that no longer correspond to reality.

The error is not fabrication. It is persistence.

A hallucination without a neural network

In discussions about AI, hallucination usually means that a system produces a fluent, confident statement unsupported by the facts.

The Chiliad record represents a closely related failure. Every field is plausible. The company name is real. The percentage is formatted correctly. The document has an official contract number, an update date and an institutional source.

It is syntactically correct, internally consistent and temporally false.

The underlying system has not invented Chiliad. It has failed to stop believing in it.

Call it an institutional hallucination, a database ghost or stale master data. The operational result is the same: an information system presents a proposition in the present tense after the proposition has ceased to be true.

A document can be perfectly sourced and still be false.

Why this matters for AI

The conventional prescription for reducing AI hallucinations is grounding: connect the model to search engines, databases and authoritative documents instead of relying entirely on its training data.

But consider what happens when an AI system is asked whether Chiliad still operates.

It finds a current public-sector contract. The document was updated in 2026. It comes from a prominent government technology distributor and is hosted with the official contract materials. It explicitly lists Chiliad as an available manufacturer.

On the surface, that is excellent evidence.

An AI following standard retrieval practices could confidently conclude that Chiliad remains active. It would be wrong not because it invented a fact, but because its authoritative source had hallucinated first.

This exposes a weakness in the idea that retrieval alone solves AI reliability. Provenance is not enough. Systems also need temporal validation, corporate entity resolution and mechanisms for detecting when apparently current documents contain inherited historical claims.

“Published in 2026” does not mean “true in 2026.”

The same problem extends far beyond procurement. Corporate websites list former executives. government directories preserve defunct programs. Customer databases retain dissolved companies. Research papers repeat corrected findings. Compliance systems copy yesterday’s classifications into tomorrow’s reports.

Generative AI did not create these errors. It gives them reach, speed and a persuasive voice.

The database ghost

Chiliad is not a competitor to Palantir today. It is not evidence of a forgotten government software giant still operating in the shadows.

It is a database ghost.

Somewhere inside a procurement information chain, a historical vendor record continues to be treated as current. That record has survived a liquidation, an asset sale, the closure of a bankruptcy case and multiple generations of contracting documents.

The result looks uncannily like modern AI hallucination because both failures share the same deeper cause: a system produces a plausible answer without adequately checking whether its internal representation still corresponds to the world.

Hallucination is older than generative AI.

AI merely industrialized it.

Before demanding that machines stop hallucinating, institutions may need to examine the supposedly authoritative information being used to ground them. Otherwise, even a perfectly obedient AI will reproduce fiction—complete with a contract number, an official update date and a 1.02 per cent discount.