Picture the board room. A director leans forward and asks the question every CDO and CIO now has to answer: Do you trust what AI is telling us? It's not a rhetorical question anymore. It's the question our clients are facing today, and it's the one that separates organizations with real AI governance from organizations that just have an AI policy.
Wrong Is Not the Same as Fabricated
We've always known AI can be wrong. That's precisely why a human owns the decision. But "wrong" implies a traceable error: a stale figure, a bad input, a value that came from somewhere and can therefore be found and fixed.
Generative and agentic AI introduce something categorically different: fabrication. A citation for a case that was never filed. A drug interaction that was never documented. A research study, a quote, or a figure with no source at all. We call this synthetic truth: output that looks true but isn't.
The distinction matters because it breaks the governance models most organizations already have. Bad input is a real value that happens to be wrong; you can trace it back to its origin. A fabrication came from nowhere. It arrives in the same confident, fluent voice as everything true, so nothing about it looks off. Every control built to check decisions still trusts the underlying input and inference to be real — and that's exactly the assumption synthetic truth breaks.
Are your employees knowledgeable on how to identify fabricated outputs?
The Courts Are Already Setting the Precedent
This isn't a hypothetical risk. It's already showing up in case law, and the pattern is consistent: courts aren't punishing the error. They're punishing the unowned assertion.
In Mata v. Avianca (2023), an attorney filed a brief citing several court decisions that ChatGPT had invented outright. The court's sanction wasn't really about using AI — it was about failing to verify the output and then standing behind it once questioned.
In Johnson v. Dunn (2025), a large, well-regarded firm filed fabricated citations, and the court signaled plainly that monetary fines alone are no longer an adequate deterrent.
And in Fletcher v. Experian (2026), the Fifth Circuit sanctioned appellate counsel over 16 fabricated quotations in a single brief. The person who signed the filing owned the claim — where the fabrication came from didn't matter to the outcome.
The throughline across all three: accountability doesn't dissolve because the source was a model instead of a person.
Not Every Fabrication Carries the Same Risk
Before your organization can govern synthetic truth, you need a way to triage it. We use a simple lens built on two axes: consequence and clarity.
Ask three questions of any AI-assisted decision:
- Is it reversible?
- Is it defensible — can you attribute the outcome to its inputs?
- Is it evident — how would a negative outcome even surface?
A breakdown to consider when assessing the risks would be the following:
- High consequence, high clarity → known risk, manageable with existing frameworks and close monitoring.
- High consequence, low clarity → your highest-priority governance gap. Refine the problem, add oversight, proceed with caution.
- Low consequence, high clarity → standard governance is sufficient.
- Low consequence, low clarity → the quiet danger zone. Risk can escalate without anyone noticing until it's a high-consequence problem.
Most organizations over-invest in the first quadrant and under-invest in the second and fourth. That's where synthetic truth does the most damage, because nobody was watching closely enough to catch it early.
Identifying problematic AI is half of the solution. Knowing how to prevent issues is equally important.
Stop Promising Trust, Start Promising a Trail
You cannot certify that every AI output is true. No governance framework can make that guarantee, and any vendor who claims otherwise is selling you a trust score, not a control. What you can do is move the locus of trust from the decision itself to the process that produced it. That process needs to be grounded, traced, and owned.
Take a bank using AI to draft adverse-action letters — the notices explaining to a customer why they were declined credit. Governing every sentence the model generates is impossible. Governing the process flow around that one consequential decision is entirely attainable: ground the output in verified data, trace which inputs produced which claim, and assign a named owner accountable for the outcome.
This is where our Decision Stewardship model earns its place. The decision-making that happens inside the model can't be governed directly, it has to be attributed. The pod around that process owns the evidence of adherence to the flow, works the exceptions, and watches for the real signal: an attribution gap. No source, no owner, no pass.
Accountability Beats a Trust Score
A defensible record will always outperform a confidence metric. Accountability means every consequential claim has a named owner and a traceable origin. When something goes wrong (and something eventually will) you catch it, you know where it came from, and you fix it. Nothing fails silently.
No system can guarantee truth. What good governance can guarantee is that errors are visible and answerable instead of hidden, and that the judgment stays human.
Where to Begin
You cannot govern every output an agentic system produces. You can govern one consequential process flow, end to end. Pick the decision where a silent fabrication would hurt the most — the one that's hard to reverse, hard to check, and slow to surface. Make it grounded, traced, and owned, with a named steward and a pod built around it.
That's how you walk back into the board room with an answer that actually holds up: "I can't certify that every answer is true. I can show you that every consequential decision is grounded, traced, and owned — and walk you through one."
That's not a lower bar than "trust the AI." It's a higher one. And it's the only one that scales.
Want to go deeper on Decision Stewardship and building governance pods for agentic AI? Reach out to our team to talk through where synthetic truth risk is highest in your organization.
