When trusted firms ship AI hallucinations

In the space of a single year, KPMG, Deloitte and EY each had to pull or refund a report after someone discovered its citations were fabricated by AI. Not a rushed intern's memo. The flagship, paid deliverable of firms whose entire product is being trusted.

The reflex response is "they should have used better AI" or "they should have checked their work." Both miss what actually broke. These were not failures of model quality. They were failures of accountability: the AI produced confident claims with no durable link to a real source, and nobody inside the firm could tell what was real until an outside investigator checked, after publication.

A pattern, not an accident

Four of the most trusted names in professional services, plus a top law firm, in twelve months:

GPTZero's CEO warned that error-riddled papers from trusted firms "poison the well of information" and breed second-hand hallucinations, fabrications laundered into the next generation of research. When the source of trust becomes the source of error, the damage compounds.

Why "use better AI" misses the point

Hallucination is not a bug that a better model quietly retires. It is a property of how these systems work: they produce fluent, plausible text whether or not the underlying claim is real. You cannot make a model stop being confidently wrong.

What you can control is whether anything ships that the AI cannot prove. In every case above, the missing thing was the same: no evidence trail binding each claim to the source it supposedly came from. The output floated free of its provenance, so a fabricated citation looked exactly like a real one until a human, or an outside firm, checked each one by hand.

You cannot stop a model from being confidently wrong. You can refuse to ship anything it cannot prove.

The layer that was missing

This is the gap AskLedger is built to close. Not by making the model smarter, but by making its output accountable. Three things turn a floating claim into a provable one:

• Bind every claim to its source. Each AI output carries a signed receipt whose evidence references point, by cryptographic hash, to the actual documents it drew on. A citation with no verifiable source reference is a red flag before publication, not a scandal after.
• Record whether it was checked. If a verification step ran, did the cited source exist, did the quote match, the receipt records that it ran and what it found. "Unverified" stops being invisible.
• Require a human to sign off. Sensitive deliverables pass through an approval step, and the receipt records who approved it and when. The human is in the loop on the record, not just in the room.

Captured at the moment the output is produced, the receipt looks like this:

Now a fabricated citation is catchable at the gate: a claim whose evidence reference points to nothing, or fails its check, does not get a clean receipt, and does not ship. And months later, when a journalist or a client asks, you can show exactly what every claim was based on, verifiable by anyone with a public key.

Trust you can hand someone

The firms that keep their credibility through this era will not be the ones that stopped using AI. They will be the ones that can show where every claim came from, before it goes out the door. On a consulting report, an audit, a regulatory filing, or a board deck, trust is no longer something you assert on your letterhead. It is something you can hand someone and let them verify for themselves.

This article summarizes publicly reported incidents, linked to their sources, and is general information, not legal or audit advice. Claims about specific firms reflect the cited reporting.