When KPMG recently pulled its own published report on AI usage after the document appeared to contain hallucinated content, the story was treated as a curiosity — an embarrassing footnote in the ongoing conversation about generative AI’s limitations. Most coverage moved on within a news cycle. But the more I sit with what actually happened, the more I think we are looking at the wrong problem. The embarrassment is not the cost. The cost is something quieter, harder to see, and accumulating every day across organizations that are not watching for it.
We talk endlessly about AI’s productivity gains. We measure token costs, time saved, headcount ratios. What we are not measuring — not seriously — is the expense of eroded institutional credibility, the compounding drain of verification work, and the slow corrosion of trust that occurs when authoritative sources publish things that turn out to be fabricated. KPMG pulling a report is simply the version of this problem that became visible. Most of it never does.
The Credibility Tax Is Real, Even When It Goes Unpaid
Consider what it actually costs an organization to publish flawed information under its name. There is the direct cost of retraction — communication, internal review, public acknowledgment. But the larger cost is the credibility discount that every future publication now carries, often invisibly. Readers who noticed the error will apply a skepticism they did not have before. Clients who trusted the firm’s AI research now hold that trust at arm’s length.
For a firm like KPMG, whose advisory value is fundamentally built on rigor and reliability, that discount is not trivial. And yet it will not appear on any balance sheet. It will not show up in a quarterly report. It will register only gradually, in deals that take slightly longer to close, in client conversations where a question lingers a beat too long. The credibility tax is real, but because it is diffuse and delayed, organizations rarely calculate it before publishing AI-assisted work.
Verification Work Has Become a Hidden Line Item
Here is something I keep seeing in my work with clients: the more AI-generated content flows through an organization, the more time knowledge workers spend checking it. This is the hidden labor cost that productivity dashboards consistently miss. An analyst who uses an AI tool to draft a 20-page research report in two hours has not saved two hours of work — they have shifted the work. Now someone must verify the claims, cross-reference the sources, and interrogate the conclusions that the model arrived at with complete confidence and no citation trail.
When that verification step is skipped or rushed — as it clearly was somewhere in the KPMG process — the cost does not disappear. It simply gets paid later, in public, at a much higher rate. The organizations that are genuinely absorbing AI’s efficiency gains are the ones that have built rigorous review processes alongside their AI workflows. Those processes cost time and people. If you are not accounting for them, you are not actually measuring what AI costs you.
Hallucination Normalizes a Dangerous Standard
There is a subtler cost that concerns me more than any single retracted report. Every time a hallucination reaches publication — and especially every time it is caught and quietly corrected — we collectively shift our expectations about what accuracy means. We begin to accept a lower standard. Research that might once have required independent source verification now gets published with a note that “AI was used in preparation.” The caveat becomes a liability shield rather than a quality signal.
The practical danger is that this normalization moves fastest in exactly the places where accuracy matters most: regulatory analysis, financial guidance, policy research. If the organizations that society trusts to get these things right — auditors, consultancies, research institutions — begin treating AI-generated errors as an acceptable cost of efficiency, the downstream consequences for decision-makers who rely on their outputs are severe and almost entirely unmeasured.
What Organizations Should Be Counting
Before deploying AI in any content that carries institutional authority, the questions worth asking are straightforward. What is the cost if this information is wrong and we publish it? Who is responsible for verification, and is that time factored into the efficiency math? Do we have a process that is genuinely catching errors, or one that just creates the appearance of review?
The KPMG story is useful precisely because it made visible a dynamic that usually stays hidden. The real question is not how one firm ended up publishing a flawed AI report. The question is how many organizations will read that story, nod knowingly, and return to the same workflow the next morning without changing a thing.