Entry 10 of 10 · 1896, with a question for today

Phantom cotton and the AI watchdog

A Victorian mill inflated its stock. Would an AI assistant expose the discrepancy—or write a persuasive explanation for it?

More than half a million pounds of cotton and yarn is a substantial quantity to conjure out of bookkeeping.

A reconstruction discussed in the Kingston Cotton Mill litigation indicated that the stock journal overstated the quantity on hand in 1891 by 583,935 lb. The calculation allowed for manufacturing waste. Increasing that allowance would have made the discrepancy larger.^{Note 1}

Most auditors remember the case for a different reason. In 1896, Lord Justice Lopes supplied the profession with its most persistent animal comparison: “He is a watch-dog, but not a bloodhound.”^{Note 2}

The cotton deserves some of the dog’s publicity. It has a particularly useful question for anyone experimenting with artificial intelligence.

A signature where a question might have been

The mill’s manager, Jackson, had inflated stock quantities and values. The auditors accepted his figures, and the balance sheets referred to the manager’s certificate. In Kay LJ’s account, that “certificate” was the manager’s signature in the stock journal. The Court of Appeal nevertheless reversed the decision against the auditors, accepting their reliance on him in the circumstances.^{Note 1}

The decision belongs to its own legal and professional setting. It is not a statement of the duties governing a present-day audit.

The practical difficulty is easier to carry across time. A representation can acquire authority because of who supplied it. A record can appear credible because it agrees with another record. Neither relationship guarantees contact with reality.

Give the mill an AI assistant

Now imagine giving the records to a modern team equipped with an AI tool. This is a thought experiment, not a claim that the technology would have detected the historical fraud.

The tool might help extract quantities from invoices and stock records, identify dates and arrange the results for a reconciliation. Opening stock plus purchases, adjusted for output and waste, could then be compared with the recorded closing quantity.

The arithmetic does not require artificial intelligence. The potential benefit is reducing the effort of assembling scattered information so that a reproducible calculation becomes practical.

Before relying on it, the team would still need to check the extraction. Pounds of cotton must not become pounds sterling. Duplicate invoices must not become extra purchases. Missing records must not quietly disappear from the definition of the population.

Suppose those questions have been addressed and a substantial unexplained difference remains. The tool has helped surface a specific problem. That is useful work.

The most dangerous output might be a paragraph

The next prompt is tempting: explain the discrepancy.

The assistant might suggest unusual production losses, transactions around the year end or inconsistencies in units. Those are possible avenues for investigation. They are not findings about the mill, and their plausibility does not establish that any of them occurred.

This is where a useful analytical tool can become a source of misplaced comfort. A fluent explanation is easy to paste into a working paper. It can sound like the end of an inquiry while remaining only a list of ways the figures might be defensible.

The auditor must still establish what happened. Were the units wrong? Is there independent evidence for the loss rate? Does the proposed timing difference reconcile to transactions in the relevant periods? What evidence would contradict the explanation?

There is also a quieter risk when the machine finds nothing. A result without exceptions is informative only if the team understands the procedure’s coverage and limitations. An incomplete population can produce a wonderfully peaceful dashboard.

Keep the calculation, investigate the answer

The historical comparison suggests a useful discipline for AI-assisted work: preserve the path from the question to the records, from the records to the calculation, and from each exception to the evidence resolving it.

That lets a reviewer examine whether the tool helped obtain evidence or merely made the file more persuasive. It also makes errors easier to locate. An extraction mistake, a flawed assumption and an unsupported explanation are different problems requiring different remedies.

Kingston Cotton Mill offers no proof that a machine would have succeeded where the auditors did not. It offers a warning about transferring trust too quickly from a person’s signature to a system’s answer.

A watchdog with a wider field of vision could be useful. Someone still has to investigate the barking.

Continue with the reading guide, or return to the ledger.

Sources

1. In re Kingston Cotton Mill Company (No. 2) [1896] 2 Ch 279, especially Kay LJ’s judgment, available in a transcript of the judgments. The 583,935 lb figure is a reconstruction discussed in the litigation, not a calculation performed by the original auditors. The AI workflow is a modern thought experiment. ↩ ↩
2. Lopes LJ’s judgment in the same case. The project also holds the Lindley and Lopes judgments in Lawrence R. Dicksee, Auditing: A Practical Manual for Auditors, fifth edition (Gee & Co., 1902), appendix B, pp. 611–615. This article adapts the project essay Phantom cotton and the AI watchdog; it makes no claim about the contents or current status of modern AI regulation. ↩