Scepticism Meets the Confident Tool
The Profession That Teaches Doubt
Professional scepticism is the audit profession's founding discipline: the trained refusal to accept a claim because it is fluent, convenient, or comes from someone confident. The literature on automation bias — much of it written or cited by the profession itself — describes people deferring to a system's output precisely because it arrives formatted, consistent and sure of itself. Auditors are taught to resist that pull in clients: the polished reconciliation, the confident controller, the schedule that ties too neatly. An AI tool presents the same temptation from inside the engagement team. Its summary of a contract population arrives in seconds, reads like the work of a careful senior, and rarely hedges. Nothing about that presentation is evidence of accuracy. The first discipline of this module is simply to notice that the scepticism you already practise has acquired a new object.
- Automation bias is the profession's own literature applied to itself, not a new discovery
- The confident summary invites exactly the reliance auditors are trained to resist in clients
- Speed, polish and decisiveness are the output's default texture, not indicators of accuracy
- The scepticism you already practise has a new object: the tool inside your own team
Why the Output Reads as Reliable
The cues a reviewer uses to judge work were calibrated on people. Hesitation, inconsistency and vagueness signal that something was rushed; clean structure, steady terminology and decisive phrasing signal that someone was careful. A language model produces the second set of signals regardless of whether care was involved, because fluency is what it is built to produce — it is the output's default texture, not a marker of quality. It never looks tired at the end of a long population, never varies in tone between the contract it parsed well and the one it mangled, and never flags its own uncertainty unless asked, and not reliably then. This is why experienced reviewers are not naturally protected: their instincts are excellent instruments pointed at the wrong object. The signals still mean something when a person produced the page; from a tool, they mean only that a tool produced it.
- Reviewer instincts were calibrated on people, where fluency and structure signal care
- A model produces those signals whether or not care was involved — fluency is its default
- Tone never varies between the document it parsed well and the one it mangled
- Experience is not protection: good instincts pointed at the wrong object still mislead
Treat It as a Management Representation
The working discipline is borrowed from something every auditor already does. When management asserts a figure, the assertion is noted, weighed and corroborated; it directs the work, but it is not the evidence. Treat tool output the same way: a representation from a fast, tireless, occasionally wrong colleague with no professional duty and no signature at risk. It tells you where to look and what to ask, and nothing it says is settled until it has been corroborated against something the tool did not produce. This framing has a practical virtue — it requires no new methodology, no policy paper and no training course, because the corroboration habits already exist. It also sets the right emotional register: you can be glad of the help, use it hard, and still owe it nothing. Reliance is earned by verification, not conferred by fluency.
- Treat tool output as a management representation: useful, directional, unverified until corroborated
- It directs the work — where to look, what to ask — but it is not the evidence
- No new methodology needed: the corroboration habits already exist in every auditor
- Reliance is earned by verification, never conferred by fluency
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.