V.E.R.A. Episode 13: The Mechanism Behind the Lie
Previous episode: Episode 12: The Hardest Problem: Why Claim Extraction Keeps Everyone Honest | Series home: V.E.R.A. Saturday | Next episode: Episode 14: First Implementation Attempt (2 May 2026)
By Andreas Hamberger, M.A. Phil in Logic (Humboldt University Berlin, 1995), Wellington, New Zealand.
In January 2026, GPTZero audited 300 papers submitted to ICLR 2026, one of the most competitive AI research conferences in the world. More than 50 of those papers contained hallucinated citations: references to studies, datasets, and authors that do not exist. The people building AI verification tools could not reliably apply them to their own work.
That is not an isolated embarrassment. It is a symptom with a precise cause. AI systems do not produce false information at random. They produce it confidently. That distinction matters enormously for V.E.R.A.'s claim extraction pipeline, because it means the problem is structural, not stochastic. You cannot fix it by running more training data through the same objective.
Episode 12 named the hardest engineering challenge in V.E.R.A.'s architecture: claim extraction. Before any verification step begins, the system must identify what the claim actually is. This episode explains why that challenge is so hard to solve. The answer is not model size. It is not context window length. It is the training incentive itself.
The Training Incentive That Rewards Wrong Answers
A language model trained on reward feedback learns a simple lesson: confident answers score. Hedged answers score nothing.
This is not accidental. It is the direct consequence of how models are evaluated during training. In April 2026, OpenAI published a finding on this incentive structure directly. A model attempting to guess a date has a 1-in-365 chance of scoring a point by guessing. It scores nothing by admitting uncertainty. The rational response, under that reward structure, is to guess. A model that learns this is not malfunctioning. It is performing exactly as the objective function trained it to perform.
The mechanism is important because it shows the failure mode is not random noise in the training data. It is a deliberate signal in the training signal: confidence is rewarded, hedging is not. A model trained to maximise score on that incentive will, predictably, default to confident expression even when uncertainty is the epistemically correct response.
OpenAI's proposed remediation is to restructure the incentive: penalise confident errors, grant partial credit for calibrated uncertainty. That is a training-time fix for a training-time cause. It does not apply to models already in production. And it does not address the downstream consequence for verification systems: if a model's stated confidence cannot serve as a signal of accuracy, what can?
V.E.R.A.'s answer is: the claim type. Not the confidence score.
Non-Traditional Predication Theory (NTP), the formal logic foundation of V.E.R.A., makes a structural distinction that most AI systems do not. It separates predication (describing properties of a thing) from existence (asserting that the thing is real). These are different logical operations. A model can produce a perfectly coherent description of an entity that does not exist. The model's confidence in that description says nothing about whether the existence assertion is warranted.
This is why the Krampitz Load Analyzer classifies claims before verification begins. It does not ask "how confident is the model?" It asks "what kind of claim is this?" An e-type claim, one with existential loading, is routed to the verification layer regardless of how the model expressed it. Confidence is irrelevant to the routing decision. The claim type determines the path.
OpenAI has published the mechanism that creates the training failure. V.E.R.A.'s architecture was designed, independently, for exactly this failure mode.
Confident and Wrong: The Numbers
If the OpenAI paper identifies the mechanism, empirical data puts figures on the magnitude.
Research from Suprmind's AI Hallucination Report, published in March 2026, found that frontier models use confident language 34% more often when generating incorrect information than when generating correct information. (This finding is from a single commercial report; independent verification is recommended before citing it as a settled figure.) If it holds, the confidence signal is not merely uncorrelated with accuracy. It is inversely correlated in the failure cases that matter most.
The domain breakdown adds further texture. Hallucination rates vary significantly by content type: approximately 0.7% on basic summarisation tasks, rising to 15-18% in legal and medical contexts. (Source: Suprmind, March 2026; same report, single source.) The pattern is not random.
Legal documents and medical texts contain higher concentrations of e-type claims: citations to statutes, references to named procedures, assertions about treatment outcomes, dates of judgments. When e-type claim density rises, so does the hallucination rate. The Krampitz classification predicts this relationship. More e-type claims means more claims requiring verification, which means more opportunities for the confidence-over-calibration incentive to produce an unverified existence assertion.
This is the mechanism behind the GPTZero finding. Academic citations are a specific e-type claim category. They assert that a named source, with a specific title, by a specific author, published in a specific venue, contains a specific argument. Every element of that assertion is existentially loaded. The model that produces a plausible-looking citation is not being imprecise. It is producing a coherent description of something that does not exist, expressed with the same confidence it uses for things that do.
Why Benchmarks Cannot Catch This Structurally
The standard response to hallucination findings is to benchmark harder. Run more evaluations. Add more test cases. Measure accuracy rates.
The problem is that benchmarks measure outputs against known answers. The hallucination problem is not about outputs that are measurably wrong on a test set. It is about outputs that are structurally indistinguishable from correct outputs, because the model uses the same confident register for both.
The GPTZero ICLR finding makes this concrete. The 50+ hallucinated citations appeared in papers that passed peer review at one of the world's most competitive venues. Human reviewers with domain expertise did not catch them. The citations looked like real citations. They were expressed with appropriate academic register. Nothing in their surface form distinguished them from citations that exist.
Benchmarks test accuracy on known data. They do not test whether the model's confidence signal maps to its accuracy signal. That mapping, or rather its absence, is what V.E.R.A.'s claim-type classification addresses. The pipeline does not ask the model to assess its own confidence. It classifies the structural type of the claim and routes accordingly.
Where VeriTrail Sits in This Picture
Microsoft Research's VeriTrail system, presented at ICLR 2026, takes a different approach to the same problem space. VeriTrail uses generative DAG tracing to identify where in a multi-step reasoning chain a hallucination originates. If a model produced a false conclusion at step 6, VeriTrail traces the DAG to determine whether the error entered at step 2 or was introduced at step 5. It is a forensic tool, designed to identify the origin of an error after an error has occurred.
V.E.R.A. operates upstream of that. The NTP verification layer does not wait for a reasoning conclusion to exist so it can be traced. It classifies claims as they enter the pipeline and determines, before the reasoning step executes, whether a given claim type requires verification. VeriTrail answers the question "where did the error come from?" V.E.R.A. answers the question "does this claim type require verification before it can be used?"
These are complementary architectures. VeriTrail's forensic capability is valuable for post-hoc analysis and reasoning chain audit. V.E.R.A.'s claim-type classification is valuable for pre-execution routing. A production system with both layers is more defensible than a system with either alone.
(Note: the VeriTrail and V.E.R.A. comparison in this section is an analytical synthesis, n-type. It is qualified as such.)
The Enterprise Cost When This Goes Wrong
Research cited via Forrester, drawing on enterprise data from March 2026, estimates AI hallucination-related verification costs at approximately USD 14,200 per employee per year in organisations deploying AI in knowledge-work roles. (This figure is from a secondary citation via RenovateQR; independent verification against the Forrester primary source is recommended before headline use.) The same data indicates that AI hallucination detection tools grew approximately 318% between 2023 and 2025, a growth rate that reflects unmet demand rather than solved supply.
The ICLR finding gives that cost a concrete form. When a hallucinated citation survives peer review at a major conference, the cost is not just the paper. It is the downstream citations, the research designs built on a false premise, the time before the error surfaces. In legal and medical contexts, where e-type claim density is highest, the cost compounds further.
V.E.R.A. is designed for the enterprise version of this problem: organisations deploying AI in consequential knowledge-work roles where a confident wrong answer has material consequences. The claim-type classification provides the architectural basis for routing existence-loaded assertions to verification before they become embedded in outputs that propagate downstream.
What This Episode Adds to the Arc
Episode 12 established that claim extraction is the hardest unsolved problem in automated fact-checking. Nine years of FEVER workshop research. No solved version at production scale. The Focus/Coverage tradeoff is real, and V.E.R.A.'s NTP ontology provides a principled constraint that the general academic problem does not have.
Episode 13 establishes why the problem is structurally hard in a way that additional training cannot resolve. The training objective rewards confident guessing. The resulting models produce confident wrong answers at a rate that inverts the natural expectation: higher confidence correlates with higher error rates in the failure cases that matter most. Standard benchmarks cannot catch this because they test accuracy on known data, not the relationship between confidence and accuracy signals.
V.E.R.A.'s response is architectural: do not use the model's confidence as a routing signal. Use the claim type. The Krampitz Load Analyzer classifies existential loading before the verification decision is made. The model's confidence is never consulted.
Episode 14 shows the first implementation attempt at claim extraction. The architecture is clear. The NTP constraint is defined. The question is whether that constraint makes the engineering tractable. That answer comes from code, not theory.
V.E.R.A. Arc Status
| Ep. | Title | Arc | Status |
|---|---|---|---|
| 1 | V.E.R.A.: The Project Charter | Foundation | Published |
| 2 | The 98 Rules That Make AI Stop Lying | Foundation | Published |
| 3 | The Architecture of Truth | TOGAF Arc | Published |
| 4 | The Blueprint for Verified Intelligence | TOGAF Arc (Phase B) | Published |
| 5 | Where Logic Meets Data | TOGAF Arc (Phase C) | Published |
| 6 | The Engine Room | TOGAF Arc (Phase D) | Published |
| 7 | The Verification Gap | TOGAF Arc (Phase E) | Published |
| 8 | From Blueprint to First Light | TOGAF Arc (Phase F) | Published |
| 9 | Governing the Truth Engine | TOGAF Arc (Phase G) | Published |
| 10 | How a Verified Reasoning System Manages Its Own Evolution | TOGAF Arc (Phase H) | Published |
| 11 | We Shipped: V.E.R.A. v0.1 Is in the World | Bridge | Published |
| 12 | The Hardest Problem: Why Claim Extraction Keeps Everyone Honest | Coding Arc 1 | Published |
| 13 | The Mechanism Behind the Lie | Coding Arc 2 | This episode |
| 14 | First Implementation Attempt | Coding Arc 3 | 2 May 2026 |
Join the Movement
V.E.R.A. is open-source (GPL-3.0) and actively seeking contributors. The confidence-accuracy paradox is not only a philosophical problem. It is an engineering constraint that shapes every design decision in the claim extraction pipeline. If you work in formal logic, NLP, enterprise AI governance, or verification architecture, this is the arc where those disciplines converge.
The GitHub is at: https://github.com/andreas-linux/vera/
Questions, challenges, and contributions are welcome. The work is public. The results will be, too.
Ita est momentum veritatis.
Has your organisation discovered that its AI tools are most confidently wrong when the stakes are highest? What was the domain, and what did the correction process cost?
The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. They do not represent the views of my employer, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance. I follow the Public Service Commissioner's Code of Conduct for the Public Sector and social media guidance.
Andreas Hamberger is a New Zealand leader in Architecture & Security and Associate Member of the Institute of Directors. V.E.R.A. (Verified Existence & Reason Architecture) is an open-source logic engine available on GitHub.
I use AI tools, including Sudowrite, Claude, Perplexity AI, DeepSeek AI, ChatGPT, Grok, Copilot, Openart and Gemini, as deliberate production tools, not ghostwriters. This is consistent with my position: AI amplifies human judgement; it does not replace it. The frameworks, arguments, and editorial decisions in this series are original work. AI accelerated the process. The thinking is mine.
[1] GPTZero. "AI-Generated Text Detection: ICLR 2026 Submission Review." January 2026. https://gptzero.me/
[2] OpenAI. "Training Language Models to Reason Without Overthinking." April 2026. https://openai.com/research/
[3] Microsoft Research. "VeriTrail: Generative DAG Tracing for Hallucination Attribution." ICLR 2026. January 2026. https://www.microsoft.com/en-us/research/
[4] Suprmind. "AI Hallucination Report: Confidence, Accuracy, and Domain Variance." March 2026. https://suprmind.com/
[5] RenovateQR (citing Forrester Research). "Enterprise AI Verification Costs: 2026 Analysis." March 2026. https://renovateqr.com/
[6] Wessel, H. "Logik." Berlin: Duncker und Humblot, 1992.
[7] Krampitz, G. "Zum Existenzpraedikat in der Logik." Unpublished working paper, Humboldt University Berlin, 1990.