Garbage In, Garbage Detected
Two objections to AI in PSM compliance are circulating among experienced auditors, and both deserve to be stated fairly, because both are built on real findings.
The first: facility data is so inaccurate that AI doesn't stand a chance of producing anything accurate. Garbage in, garbage out.
The second: decades of software tools came before AI, and audits still find gaping holes. The SOP copied from another plant with the valve numbers never changed. Management certifying it true and accurate. KPIs built on documentation being 100% complete. So AI has a long way to go.
The objections are true of many facilities, but not all of them. There are plants running sharp programs where the SOPs match the plant and the meetings actually happen, and blanket statements about the industry roll right over them. Still, the findings are real. Anyone who has audited in this industry has seen the copy-paste SOP. Most have seen it certified. The objections come from people who have spent careers opening files and finding the gap between what the paper says and what the plant is. They've earned the skepticism.
Every failure in that list was produced under the current regime. The copy-paste SOP was written by a person, reviewed by people, and signed by management. The completeness KPI was designed by people and reported to people. Software held the documents; humans read them; a signature certified them. AI wasn't in the room. The evidence offered against AI is a case file on the system we already have. And that system's verdict is in. It certified the wrong valve numbers as true and accurate.
So the comparison the objections assume — accurate human programs on one side, risky AI on the other — doesn't exist. The real baseline is broken programs already certified as complete. Whatever comes next gets measured against that, not against a standard the current regime never met.
Garbage in, garbage out is true of any system that reads a document and believes it. Feed a trusting reader a bad SOP and it will summarize a bad SOP fluently. If that's the tool, the skeptics are right about it.
A system built to cross-reference is a different machine. Check the SOP against the P&IDs. Check the valve numbers against the equipment list. Check the written program against the records of work performed. That system doesn't inherit garbage. It detects it.
The copy-paste SOP is the cleanest example. It survived human review for years because humans read documents one at a time and extend trust to what's already signed. Its defining feature — that it describes a different plant — is invisible to serial reading and fatal under cross-referencing. The very inaccuracy that supposedly defeats AI is the easiest thing for a cross-referencing system to catch.
The check extends past documents agreeing with each other. An employee participation plan says the facility meets four times a year. The records show one meeting in the last twelve months. No judgment call required: the program made a promise and the file doesn't contain the evidence. A human auditor can run that check too, and does, once a cycle, for the elements there's time to pull. Software runs it across every element, every commitment the program makes, continuously. The question shifts from what the program says to whether the work is being done.
Which inverts the objection entirely. The dirtier the data, the more valuable the system that flags dirt instead of reading past it. Clean programs don't need a detector. The industry the skeptics describe, inaccurate data in the vast majority of facilities, is the strongest argument for building one.
Cross-referencing catches contradiction. It does not catch a consistent falsehood. If every document is wrong the same way, copied from the same bad source, nothing cross-checks. That blind spot is real. It's also the same one the human auditor has. Nobody catches a consistent lie by reading. The claim isn't omniscience. The claim is that most garbage in these programs lives in the inconsistencies: the valve number that matches nothing, the meeting that never happened. And inconsistency is exactly what verification is good at.
Goodhart's law: when a measure becomes a target, it stops being a good measure. "Documentation 100% complete" started as a proxy for a healthy program. The moment management was judged on it, facilities produced completeness: every SOP exists, every box checked. The number hit 100% while the program underneath it rotted. Nobody exactly lied. They satisfied the target. The certified copy-paste SOP is Goodhart's law wearing a hard hat.
And no tool escapes it, including this kind. If cross-referencing becomes the standard programs are judged by, programs will optimize to pass cross-referencing. Gaming doesn't die. What changes is the price.
Faking complete is cheap. The document only has to exist. Faking consistent is expensive. A forged SOP that survives checks against the P&IDs, the equipment list, and the maintenance records has to match the actual plant. A document that matches the actual plant is most of the way to being accurate. Same with execution: backfilling a year of participation meetings that never happened is ongoing fabrication, not a copy-paste. Raise the cost of gaming far enough and it converges with the cost of doing the work. That's the most a measurement system can offer, and it's more than completeness ever offered.
Two things stay true after all of this. An AI that only reads the file and trusts it inherits every inaccuracy in the file. The objections stand against that architecture, fully. And detection isn't repair. A flagged SOP still needs someone with authority and budget to fix it. The finding that surfaces on day one can still sit unclosed on day 400. Verification moves the problem into daylight. Daylight isn't the fix. It's the precondition for one.
The certification signature starts meaning something again. Today a manager signs "true and accurate" over more documentation than anyone has time to check line by line. The signature is trust in the process that produced the stack. When every document has been checked against every other document and against the records of work performed before it reaches the desk, the signature becomes what the standard always intended: an informed attestation. That protects the person signing more than anyone.
Audits change shape. The auditor who spends day one discovering the wrong valve numbers spends it instead on what discovery can't automate: the judgment calls, the field walk, the questions a check can't ask. The finding list stops being an archaeology report and starts being a work list.
And the standard for evaluating tools gets simple. Ask one question of anything that wants a role in judging whether a facility's program is real: does it verify the documents against each other and against the work, or does it read them and trust? A trusting reader automates the exact failure the skeptics have spent careers documenting. A verifying one is the first tool that's ever been on their side of the argument.