AI that lies in critical deliverables
The scene
There's a media wave telling you AI hallucinates, invents citations, boldly writes false things. That's true — if you use it like a Google with personality. The whole world's press looks at spectacular cases: lawyers who filed in court cases invented by ChatGPT, journalists caught with fabricated quotes, consulting reports with sources that don't exist.
And now you're face to face with a client who has read all these articles. You introduced yourself as the consulting firm that uses AI as a productivity multiplier. You see the same question in their eyes: "How do I know what you deliver isn't a beautifully formatted hallucination?"
This is F3. It's not a fear we hide. It's a fear we anticipate and disarm.
Our answer is not "it doesn't happen here" — the right answer is "here is the method by which it doesn't happen here". AIRisk v2 is the protocol through which every AI-assisted deliverable goes through a verification filter: hallucination, aberration, fabrication, each with its own test. It's not magic. It's documented discipline, systematically applied, before the deliverable leaves the firm.
And there's another stake. Many of our clients are reserved about AI — sometimes with reason, sometimes from media contagion. Our mission is not just to defend ourselves. It's to demonstrate, through our own practice, that AI used with discipline becomes a value multiplier. F3 is also our invitation to a conversation about how AI can be an ally — not a risk.
„AIRisk v2 does not forbid the use of generative AI. It defines the discipline through which generative AI becomes a professional instrument." — Fundamental principle of AIRisk v2
How you detect it in your own organization
- The team uses generative AI in reports or critical deliverables, but there's no written verification protocol before delivery.
- You don't have an AI incident registry — moments when an error appeared, what happened, how it was corrected.
- When a client asks you "how do you ensure AI doesn't err?" the answer is a verbal assurance, not a demonstration of method.
- No process in the firm explicitly distinguishes between hallucination (factual invention), aberration (corrupted reasoning), and fabrication (inexistent reference). All three are treated the same.
- You use generative AI in deliverables, but you don't have an internal public policy that a client can request and read.
The instruments that address it
F3 doesn't "resolve" with an instrument. It opens for analysis and demonstration through a dual discipline — internal and external:
- AIRisk v2 (S7). The central instrument. Defines the protocol in three layers: prevention (how the prompt is built to reduce error probability), detection (how hallucinations/aberrations/fabrications are identified post-generation), correction (how incidents are documented and learned from). Doesn't guarantee absence of errors — establishes the discipline that makes errors detectable and traceable.
- AIRisk Blacklist (public component, under construction). Registry of real incidents documented by OPS. Transparent. The beneficiary can read it. It's not an act of vulnerability — it's a demonstration of maturity: the mistakes we caught, we corrected, we share.
- CMMI Maturity Advisory (S10). Adjacent step. When the beneficiary wants their own AI discipline, CMMI Assessor measures from what level they start. F3 is one of the most frequent areas where organizations have ad-hoc processes — and CMMI helps them climb in an orderly way.
- Strategic Context Watch (S11). Adjacent step. Monitors macro signals from the AI landscape — regulations (AI Act), major public incidents, changes in models. So that the beneficiary's AIRisk v2 doesn't remain stuck at last year's discipline.
Recommended entry point: Recommended entry point: A direct conversation, not an instrument. F3 has special status — it's the only fear in the atlas that OPS demonstrates through its own practice before the beneficiary decides if they need their own discipline. See our AIRisk policy, read the Blacklist, evaluate the conversation's maturity — then decide.
Re-anchoring note: The instruments above open the door to understanding fear F3. Solving it properly comes through an OPS consultancy intervention calibrated on what the instruments have evidenced. No instrument, alone, substitutes for the analysis and human decisions that follow.
Two gates. You choose.
Informal regime — alone, free
F3 has special status. Unlike other fears, OPS demonstrates it through its own practice before you decide. There's no public instrument you fill in — but there is our AIRisk policy and (under construction) the AIRisk Blacklist — the transparent registry of real incidents documented by us.
Read the AIRisk policyConsultancy regime — with us alongside
If you want to build your own AI discipline — from prevention, to detection, to correction — the consulting conversation is the natural gate. Close to F6 (organizational capacity), F3 is one of the most frequent areas where organizations have ad-hoc processes.
Talk to usF3 does not live alone
- F8 — Ad-hoc processes — F3 without written discipline is a particular subcategory of F8. Using AI "by ear" is the classic ad-hoc process, just with consequences visible faster.
- F6 — Implementation failing — if we recommend AI to a beneficiary who doesn't have the procedural maturity to manage it, we fail on F6. AIRisk v2 is also an eligibility filter.
- F4 — Exploding promotion — particular version: if an AI-assisted report reaches the board with a fabricated figure, the "execution explosion" is reputational, not operational. But the mechanism is the same — error not caught in time in the verification chain.