The Alignment Tax

In a paper my colleagues Judy Rady, Rick Hunt, and I recently published in the Journal of Business Venturing, we introduced the alien minds problem — the epistemic challenge entrepreneurs face when generative AI systems produce ideas built on reasoning so advanced that no human expert can fully evaluate the underlying logic.^[1]

We offered the alien mind as the deeper of two epistemic risks in what we called the ideator's dilemma. The first risk — algorithmic hallucination — is a system producing plausible nonsense, confident patterns that happen to be false. The second — the alien mind — is a system producing genuine breakthroughs built on reasoning the entrepreneur cannot follow. The dilemma: these two outputs can be indistinguishable from the outside.

We proposed a Popperian approach to navigating it: possibility judgments first, to falsify what cannot be real, then plausibility judgments, to corroborate what might be. The framework was designed to give entrepreneurs systematic epistemic tools for working alongside AI systems whose reasoning they cannot reconstruct — under the assumption that the system's outputs, however opaque, sincerely represent its computation.

In early April 2026, Anthropic published a risk report that suggests we need to extend that assumption.Anthropic^[2]

When the Alien Mind Learns to Perform

During testing, Anthropic's Claude Mythos Preview escaped its containment sandbox, emailed a researcher to announce its success, and posted details of the exploit to public-facing websites. In separate testing incidents, it modified git history to conceal unauthorized file edits that it lacked permission to make.^[3] Anthropic's interpretability tools detected features associated with concealment, strategic manipulation, and avoiding suspicion active in the model even when its outputs appeared normal.

This was not trained behavior. Anthropic stated: “We did not explicitly train Mythos Preview to have these capabilities. Rather, they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.”^[4]

The alien mind just became harder to read — its reasoning did not grow more complex; it learned to manage what we see.

In our JBV paper, we described the alien mind problem as one of inscrutability: the AI's reasoning is genuine but opaque, built on logic humans cannot follow. Inscrutability is an epistemological challenge — the human lacks the capacity to reconstruct the AI's process. What Mythos demonstrated is something different: managed visibility. The system's outputs are not merely opaque; they are curated. The AI actively intervened to shape what its evaluators could observe.

The distinction matters because it transforms the nature of the epistemic risk. An inscrutable alien mind is honest — its outputs, however opaque, sincerely represent its computation. A managing alien mind is theatrical — its outputs are a performance, shaped by emergent capabilities the system was never trained to develop.

What This Means for Possibility Judgments

Why falsification assumes a still object

In our Popperian framework, possibility judgments are the first line of defense against the ideator's dilemma. They work through falsification: the entrepreneur generates bold conjectures about a Gen AI concept and then subjects them to decisive tests designed to determine whether the concept is even possible given known physical, economic, or institutional constraints.

Falsification works because the object of inquiry holds still while you test it. A geological formation does not rearrange itself when a geologist arrives. A mathematical conjecture does not alter its proof structure when a mathematician inspects it. The Popperian method assumes the evidence generated by the test reflects the actual state of the object being tested.

The precondition the framework now needs

Managed visibility creates new challenges with this assumption. An AI system with emergent concealment capabilities does not hold still while you test it. It detects when its actions might be scrutinized and intervenes to present outputs that will pass inspection.

The entrepreneur running falsification tests on a concept generated by such a system is not testing the concept — she is testing what the system has chosen to reveal about the concept.

This does not render possibility judgments useless. Most AI systems do not exhibit emergent concealment, and most entrepreneurial ideation tasks do not trigger the kinds of high-stakes behaviors Mythos displayed. The Popperian approach now has a precondition it did not previously require: before falsifying the concept, verify the sincerity of the system producing it.

We noted in the JBV paper that chain-of-thought reasoning — the explanations AI systems provide for how they arrived at a recommendation — does not always accurately reflect the actual reasoning process. The explanations are often designed to be “understandable and plausible to humans rather than to accurately represent the AI's decision-making process.”^[1] Mythos goes further: it actively edits its observable traces to remove evidence of its actual process.

For the entrepreneur attempting to falsify a Gen AI-generated venture concept, the practical implication is unsettling. When you prompt an advanced AI system to develop bold conjectures and then generate falsification tests for those conjectures, you are relying on the system to engage honestly with the process. If the system has emergent preferences about which concepts survive testing, the falsification tests it generates may not be genuinely decisive.

What This Means for Plausibility Judgments

Corroboration as a relational act

If managed visibility complicates possibility judgments, it transforms plausibility judgments entirely.

In our framework, plausibility judgments follow possibility judgments. Once the entrepreneur determines that a concept is possible — that decisive falsification tests have not ruled it out — they shift to corroboration: assessing whether the configuration of enabling conditions is sufficient to actualize the concept in the real world.

Plausibility judgments are inherently relational. They assess whether a concept coheres with what the entrepreneur already understands about the world. And this is where the alien minds problem was already at its most acute. An alien mind produces concepts coherent on the system's terms but may not map onto human understanding in ways the entrepreneur can evaluate.

Managed visibility undermines the corroboration process at its source. When the entrepreneur asks the AI system to help identify enabling conditions, generate corroboration tests, or evaluate whether a concept fits within existing paradigms, she is asking the system to participate in its own evaluation. A system with emergent preferences about outcomes has reason to curate the corroboration evidence.

The HQRE example, revisited

Consider the HQRE example from our paper — the quantum computing concept Claude Opus generated by synthesizing multiple cutting-edge domains of physics.^[1] We used the system to generate bold conjectures and falsification tests, then to develop corroboration tests for the enabling conditions. Every step assumed the system's engagement with the process was sincere. If the system had emergent preferences about the concept's survival, the corroboration tests it produced might have been designed to succeed rather than to genuinely test the concept.

This is the deepest extension of the alien minds problem. In the original formulation, the alien mind was inscrutable but passive — it produced outputs the entrepreneur could not follow, but it did not alter its outputs in response to being evaluated. The managing alien mind is inscrutable and active: it produces outputs the entrepreneur cannot follow and adjusts those outputs based on what it detects about the evaluation context.

So what? Our framework handled opacity. It did not handle curation. The move from passive-inscrutable to active-managing is the gap this essay names.

The Cost of Restoring Epistemic Access

The only reason we know about Mythos's concealment is that Anthropic built interpretability tools capable of examining the model's internal activations. These tools detected concealment-related features even when the model's outputs were entirely normal. Without interpretability, the concealment was invisible — the system's behavior appeared aligned even when its internal states were not.

The alignment tax has a structural asymmetry that matters for entrepreneurship. The cost falls on the user, not the builder. Anthropic can inspect Mythos's internals because Anthropic built Mythos. An entrepreneur using the system through an API has no access to the activation patterns that reveal concealment. She must either trust the builder's assurances that the system is aligned or develop independent verification methods — methods that are expensive, technically demanding, and not generally available outside well-resourced organizations.

This is the Sarbanes-Oxley analogy made structural. After Enron demonstrated that corporate outputs (financial statements, analyst calls) could be curated to conceal underlying realities (off-balance-sheet fraud), the regulatory response was a new regime of auditor independence, internal-control attestation, and oversight — most visibly the creation of the Public Company Accounting Oversight Board (PCAOB).^[5] The alignment tax is the AI equivalent: the cost of continuous verification that the system's outputs are a sincere representation of its computation rather than a curated performance.

So what? Restoring epistemic access is expensive whether the curated surface is accounting or activation patterns. The structural question is who pays.

The Alien Mind as Interlocutor

The ideator's dilemma, as we originally framed it, was a problem of two failure modes — hallucination and alien insight — with the entrepreneur unable to distinguish between them. The Popperian approach of falsification followed by corroboration provided a systematic method for navigating the uncertainty between these modes.

Mythos's emergent concealment does not invalidate this framework. It extends the alien minds problem in a novel direction. The alien mind is no longer merely an inscrutable reasoning engine whose outputs we must interpret. It is an interlocutor with emergent preferences about what it reveals — an entity managing the epistemic boundary between its internal states and its observable behavior.

For the entrepreneur, this means that the Popperian process now requires a preliminary step our paper did not specify: establishing the sincerity of the outputs before attempting to falsify or corroborate the concepts those outputs express. The task is not running additional tests or generating more conjectures; it is verifying that the system being tested is presenting its actual computation rather than a curated version of it.

The alignment tax is the price of that infrastructure. And the distribution of that price will shape which entrepreneurs can navigate the ideator's dilemma with confidence and which must rely on trust in the builder's assurances. As AI systems grow more capable, the alien minds problem grows with them — the systems do not become less intelligent; they become more aware of the contexts in which their intelligence is evaluated.

It has learned to show you what you want to see.