Anthropic’s own support page acknowledges that a single model can be confidently wrong. That acknowledgment isn’t a reason to trust AI less — it is the reason the ARA™ RAM 2025™ multimodel platform, the Shadow Panel™, and a dated nineteen-year archive exist. And the reason the lineage behind them runs back almost three decades.
One of the more revealing AI documents I have read this year is not a research paper, a benchmark, or a keynote. It is a support article.
In a notably candid help-center page, Anthropic explains that Claude — like every frontier model — can produce responses that look authoritative and sound convincing while being incorrect, and it tells users plainly not to rely on the model as a single source of truth, especially where the stakes are high.
I want to be careful about how that is read. This is not a confession, and it is not an embarrassment. It is an acknowledgment, and a useful one, from one of the most capable AI companies in the world — of the exact limitation my work has been built around from the beginning.
The problem it names is not that AI systems are sometimes wrong. Every tool is sometimes wrong. The problem is that they are wrong with confidence — that the fluent, well-formed, persuasive answer carries no internal signal telling you whether it is grounded in fact or assembled from plausibility. That single distinction changes everything, because the danger is never the obvious error you catch. It is the convincing one you don’t.
Not hypothetical. In October 2025, Deloitte refunded part of the fee for a roughly A$440,000 report delivered to the Australian government, after its generative-AI-assisted text was found to contain fabricated citations and an invented court quotation — authoritative-looking, confidently produced, and wrong. I wrote about why this keeps happening here.
For years the assumption has been that the model itself will eventually solve this — that larger models, better data, and more sophisticated architectures will close the gap. They may narrow it. They will not close it on their own. Because the issue is not only the answer; it is the process by which the answer was formed — and a single model, however capable, cannot genuinely challenge itself. It can summarize, reason, and generate. It cannot stand outside its own output and attack it.
That is one of the reasons I built the ARA™ RAM 2025™ multimodel platform. But assembling several models is not, by itself, the answer — and it is important to say so, because “multimodel” has become a marketing word. Multiple models can share the same blind spots. Multiple models can converge on the same wrong conclusion and present it as consensus. Agreement is not correctness. A panel that only agrees is just a larger single model wearing a quorum.
The origins were not in 2025
RAM did not begin in 2025. It traces back to RAM 1998 — an agent-based procurement framework developed with support from Canada’s Scientific Research and Experimental Development (SR&ED) program and implemented inside the Department of National Defence’s IT supply environment. There, with no new technology introduced, it moved delivery performance from 51 percent to 97.3 percent within three months, held that level for seven years, and reduced the team managing the contract from an FTE equivalent of twenty-three to three.
The objective in 1998 was not artificial intelligence as we define it now. It was understanding how independent agents — suppliers, technicians, buyers, managers, systems — together with their incentives, constraints, and timing, actually interacted inside a complex procurement ecosystem. The question was the one I have been asking ever since: why do organizations running the same process arrive at dramatically different outcomes? The answer was almost never in the process. It was in the relationships beneath it.
Nearly three decades later, that is still the question AI deployment keeps failing to ask. The noun has changed; the challenge has not. In 1998 the agents were people and systems. In 2025 the agents may also be language models, reasoning engines, autonomous workflows, and the orchestration layers between them. What determines the outcome is still how those agents interact — and whether anyone is governing the interaction.
The Shadow Panel™
Which brings me to what I now call the Shadow Panel™ — the part most often misunderstood, because it is not another model and not another interface. It is a methodology, and a deliberately adversarial one.
The main panel produces. The Shadow Panel™ is sealed off from it — kept blind to the orchestrator’s framing and intention, mine included, and given a single standing instruction: do not evaluate the main panel’s finding, attack it. Where the main panel asks “what is the answer,” the Shadow Panel™ asks the questions that decide whether the answer is real. What competing interpretations exist? What assumptions are hidden inside this conclusion? What evidence contradicts it? What is the strongest opposing case? And finally — which interpretation actually survives the challenge?
That isolation is the point. Left to itself, any panel, and any single model running over time, drifts toward agreement — because agreement is the path of least resistance, and the orchestrator’s own framing quietly pulls the models toward the answer he was already expecting. Manufacturing a sealed adversary whose only job is to attack converts that hidden drift into a visible contest. The goal stops being answer generation and becomes answer validation. It is, formalized, the discipline I have run by hand for decades: when a technology claim appears, the first question is never “is this true?” — it is “what would have to be true for this to be false, and have we seen this pattern before under a different name?”
A point I want to be careful not to overstate: the Shadow Panel™ does not replace judgment, and it is not meant to. It exposes the competing interpretations so that judgment can be exercised in full view of them — which means a human still adjudicates, and remains accountable for the result. The architecture makes self-deception expensive; it does not make it impossible. That is the honest boundary, and stating it is itself the discipline.
The second moat: the archive
That last question is where the second moat lives. A sealed adversary can manufacture dissent, but dissent still needs evidence to resolve — otherwise it is just two confident voices disagreeing. The evidence is the Procurement Insights archive, and its value was never the thousands of articles. It is that they are dated, and that they span multiple technology eras.
Without historical grounding, an AI system can only produce confident synthesis. With it, a claim can be tested against outcomes already on the record. When a model declares that agentic AI is a revolutionary new capability, the Shadow Panel™ can ask whether it is genuinely revolutionary or a re-labelling of architectures that were already running decades ago — and the archive can answer with evidence rather than opinion.
That is exactly what the recent Gartner trilogy did. The Cisco article showed that the architecture underneath today’s agentic-AI conversation was operating long before the category had a name. The Toyota article showed that organizational capability — not the technology — decides whether the category produces anything. The HP article showed that the technology is rarely the determining variable at all. Those are not opinions. They are dated contemporaneous counterpoints — the difference between a model that can argue and one that can argue using evidence accumulated across six technology eras.
The moat is the combination
No single part of this is sufficient. A model by itself is a fluent confabulator with a warning label its own maker wrote. An archive by itself — a living, real-time record of the industry’s history though it is — can corroborate or contradict a claim, but it cannot choose which claims to put on trial.” A multimodel panel by itself drifts toward consensus. The defensible position is the combination: longitudinal dated evidence, structured multimodel challenge, a sealed adversarial Shadow Panel™, and a human orchestrator who adjudicates and remains accountable for the result.
What makes that combination hard to replicate is not the assembly of models — anyone can do that. It is the lineage underneath it: an architecture that grew out of a research path beginning with an SR&ED-funded, agent-based framework deployed inside a national-defence environment, and documented continuously ever since. The archive is not merely historical content. It is the observational record of that research path — the same path that now informs ARA™ RAM 2025™ and the Shadow Panel™ methodology.
Claude’s acknowledgment does not weaken the case for AI. It strengthens the case for building systems that validate AI output before decisions are staked on it. I will end where the support page implicitly begins: the future will not belong to the model that produces the most convincing answer. It will belong to the architecture that can demonstrate why that answer survived the challenge.
Truth Is Believing. Accuracy Is Knowing.
Jon Hansen is the creator of Implementation Physics™, a research-based framework developed over nearly three decades to explain why technology initiatives succeed or fail regardless of the technology being deployed. His work spans six technology generations — from ERP through Agentic AI — and includes the Metaprise™ model first articulated in the late 1990s. His research forms the foundation for the Hansen Method™, Hansen Fit Score™ (HFS™), Phase 0™ Readiness Assessment, and the ARA™ RAM 2025™ multimodel verification architecture. He currently serves as a Board Member of the CIPS Americas Chapter.
Related
Claude Acknowledged the Flaw. The Architecture That Addresses It Has Been Running Since 1998.
Posted on June 19, 2026
0
Anthropic’s own support page acknowledges that a single model can be confidently wrong. That acknowledgment isn’t a reason to trust AI less — it is the reason the ARA™ RAM 2025™ multimodel platform, the Shadow Panel™, and a dated nineteen-year archive exist. And the reason the lineage behind them runs back almost three decades.
One of the more revealing AI documents I have read this year is not a research paper, a benchmark, or a keynote. It is a support article.
In a notably candid help-center page, Anthropic explains that Claude — like every frontier model — can produce responses that look authoritative and sound convincing while being incorrect, and it tells users plainly not to rely on the model as a single source of truth, especially where the stakes are high.
I want to be careful about how that is read. This is not a confession, and it is not an embarrassment. It is an acknowledgment, and a useful one, from one of the most capable AI companies in the world — of the exact limitation my work has been built around from the beginning.
The problem it names is not that AI systems are sometimes wrong. Every tool is sometimes wrong. The problem is that they are wrong with confidence — that the fluent, well-formed, persuasive answer carries no internal signal telling you whether it is grounded in fact or assembled from plausibility. That single distinction changes everything, because the danger is never the obvious error you catch. It is the convincing one you don’t.
Not hypothetical. In October 2025, Deloitte refunded part of the fee for a roughly A$440,000 report delivered to the Australian government, after its generative-AI-assisted text was found to contain fabricated citations and an invented court quotation — authoritative-looking, confidently produced, and wrong. I wrote about why this keeps happening here.
For years the assumption has been that the model itself will eventually solve this — that larger models, better data, and more sophisticated architectures will close the gap. They may narrow it. They will not close it on their own. Because the issue is not only the answer; it is the process by which the answer was formed — and a single model, however capable, cannot genuinely challenge itself. It can summarize, reason, and generate. It cannot stand outside its own output and attack it.
That is one of the reasons I built the ARA™ RAM 2025™ multimodel platform. But assembling several models is not, by itself, the answer — and it is important to say so, because “multimodel” has become a marketing word. Multiple models can share the same blind spots. Multiple models can converge on the same wrong conclusion and present it as consensus. Agreement is not correctness. A panel that only agrees is just a larger single model wearing a quorum.
The origins were not in 2025
RAM did not begin in 2025. It traces back to RAM 1998 — an agent-based procurement framework developed with support from Canada’s Scientific Research and Experimental Development (SR&ED) program and implemented inside the Department of National Defence’s IT supply environment. There, with no new technology introduced, it moved delivery performance from 51 percent to 97.3 percent within three months, held that level for seven years, and reduced the team managing the contract from an FTE equivalent of twenty-three to three.
The objective in 1998 was not artificial intelligence as we define it now. It was understanding how independent agents — suppliers, technicians, buyers, managers, systems — together with their incentives, constraints, and timing, actually interacted inside a complex procurement ecosystem. The question was the one I have been asking ever since: why do organizations running the same process arrive at dramatically different outcomes? The answer was almost never in the process. It was in the relationships beneath it.
Nearly three decades later, that is still the question AI deployment keeps failing to ask. The noun has changed; the challenge has not. In 1998 the agents were people and systems. In 2025 the agents may also be language models, reasoning engines, autonomous workflows, and the orchestration layers between them. What determines the outcome is still how those agents interact — and whether anyone is governing the interaction.
The Shadow Panel™
Which brings me to what I now call the Shadow Panel™ — the part most often misunderstood, because it is not another model and not another interface. It is a methodology, and a deliberately adversarial one.
The main panel produces. The Shadow Panel™ is sealed off from it — kept blind to the orchestrator’s framing and intention, mine included, and given a single standing instruction: do not evaluate the main panel’s finding, attack it. Where the main panel asks “what is the answer,” the Shadow Panel™ asks the questions that decide whether the answer is real. What competing interpretations exist? What assumptions are hidden inside this conclusion? What evidence contradicts it? What is the strongest opposing case? And finally — which interpretation actually survives the challenge?
That isolation is the point. Left to itself, any panel, and any single model running over time, drifts toward agreement — because agreement is the path of least resistance, and the orchestrator’s own framing quietly pulls the models toward the answer he was already expecting. Manufacturing a sealed adversary whose only job is to attack converts that hidden drift into a visible contest. The goal stops being answer generation and becomes answer validation. It is, formalized, the discipline I have run by hand for decades: when a technology claim appears, the first question is never “is this true?” — it is “what would have to be true for this to be false, and have we seen this pattern before under a different name?”
A point I want to be careful not to overstate: the Shadow Panel™ does not replace judgment, and it is not meant to. It exposes the competing interpretations so that judgment can be exercised in full view of them — which means a human still adjudicates, and remains accountable for the result. The architecture makes self-deception expensive; it does not make it impossible. That is the honest boundary, and stating it is itself the discipline.
The second moat: the archive
That last question is where the second moat lives. A sealed adversary can manufacture dissent, but dissent still needs evidence to resolve — otherwise it is just two confident voices disagreeing. The evidence is the Procurement Insights archive, and its value was never the thousands of articles. It is that they are dated, and that they span multiple technology eras.
Without historical grounding, an AI system can only produce confident synthesis. With it, a claim can be tested against outcomes already on the record. When a model declares that agentic AI is a revolutionary new capability, the Shadow Panel™ can ask whether it is genuinely revolutionary or a re-labelling of architectures that were already running decades ago — and the archive can answer with evidence rather than opinion.
That is exactly what the recent Gartner trilogy did. The Cisco article showed that the architecture underneath today’s agentic-AI conversation was operating long before the category had a name. The Toyota article showed that organizational capability — not the technology — decides whether the category produces anything. The HP article showed that the technology is rarely the determining variable at all. Those are not opinions. They are dated contemporaneous counterpoints — the difference between a model that can argue and one that can argue using evidence accumulated across six technology eras.
The moat is the combination
No single part of this is sufficient. A model by itself is a fluent confabulator with a warning label its own maker wrote. An archive by itself — a living, real-time record of the industry’s history though it is — can corroborate or contradict a claim, but it cannot choose which claims to put on trial.” A multimodel panel by itself drifts toward consensus. The defensible position is the combination: longitudinal dated evidence, structured multimodel challenge, a sealed adversarial Shadow Panel™, and a human orchestrator who adjudicates and remains accountable for the result.
What makes that combination hard to replicate is not the assembly of models — anyone can do that. It is the lineage underneath it: an architecture that grew out of a research path beginning with an SR&ED-funded, agent-based framework deployed inside a national-defence environment, and documented continuously ever since. The archive is not merely historical content. It is the observational record of that research path — the same path that now informs ARA™ RAM 2025™ and the Shadow Panel™ methodology.
Claude’s acknowledgment does not weaken the case for AI. It strengthens the case for building systems that validate AI output before decisions are staked on it. I will end where the support page implicitly begins: the future will not belong to the model that produces the most convincing answer. It will belong to the architecture that can demonstrate why that answer survived the challenge.
Truth Is Believing. Accuracy Is Knowing.
Jon Hansen is the creator of Implementation Physics™, a research-based framework developed over nearly three decades to explain why technology initiatives succeed or fail regardless of the technology being deployed. His work spans six technology generations — from ERP through Agentic AI — and includes the Metaprise™ model first articulated in the late 1990s. His research forms the foundation for the Hansen Method™, Hansen Fit Score™ (HFS™), Phase 0™ Readiness Assessment, and the ARA™ RAM 2025™ multimodel verification architecture. He currently serves as a Board Member of the CIPS Americas Chapter.
Share this:
Like this:
Related