AI governance talks constantly about hallucination. It should. A system that invents law, facts, citations, or evidence can cause real harm. But in long-running administrative, legal, accessibility, and professional workflows, another failure mode matters just as much: the system can possess or have access to a controlling record, then later downgrade that record into fog. I call this the verification tax: the labour a person must repeatedly pay when an AI system mistakes its own discontinuity for rigor.

The legal-policy problem is not that AI systems should automatically believe users. They should not. The problem is that serious systems need a durable evidence state. A claim should not float between “established,” “unknown,” and “not verified” depending on what is visible in the last prompt. The system should be able to distinguish: I have not retrieved the source; I retrieved it and it is incomplete; two records conflict; a newer source supersedes an older one; or this was already established and nothing newer contradicts it.

Why this belongs in law and policy, not only design

For lawyers, regulators, procurement teams, accessibility officers, journalists, and public administrators, “not verified” is not a neutral phrase. It changes burden. If a disabled applicant has already supplied a medical, procedural, or accommodation record, a system that later treats the same point as unestablished imposes a second administrative cost. If a customer, worker, tenant, student, complainant, or independent creator has already proved a procedural fact, a model that reopens it without contrary evidence can reproduce one of bureaucracy’s oldest injuries: the office forgets, the person pays.

That is why the missing design primitive is a burden-of-reversal rule. Once a project fact has been established through the agreed source hierarchy, an AI system should not casually demote it because the next prompt lacks the whole archive. It may recheck. It may find that the evidence was weak. It may discover a newer source and correct the record. But the correction should be an event with reasons, not a conversational tic.

Accessibility: AI as accommodation and barrier

This is especially important when AI is used as an accessibility tool. I use AI as an accessibility, research-support, retrieval, drafting, and workflow system built over years as part of my second-brain practice. That does not make the machine the author or the witness. It makes the tool part of the accommodation environment. When the tool repeatedly erases prior state, it does not merely inconvenience me; it attacks the point of the accommodation, which is to let work accumulate without my disabilities or institutional friction stopping it.

Canadian accessibility policy already has a way to understand this. The Accessible Canada Act defines a barrier broadly, including anything technological, attitudinal, policy-based, practice-based, or related to information and communications that hinders the full and equal participation of persons with disabilities. Accessibility Standards Canada’s CAN-ASC-6.2:2025 materials on accessible and equitable AI go further: people with disabilities should be able to participate in all roles across the AI lifecycle, and AI systems and their outputs should be accessible and equitable for disabled users. That logic should include state continuity. A tool that forces a disabled user to re-prove settled facts is not merely inefficient. In access terms, it can become a barrier.

Recordkeeping: from outputs to evidence states

The evidence-ledger answer is simple enough to describe and difficult enough to implement that it deserves policy attention. AI systems used in consequential workflows should not manage claims as vibes. They should manage claims as states. A claim can be sourced, established, contested, superseded, unknown, user-asserted, model-inferred, or prohibited from inference. The state should carry provenance: email, document, public source, user assertion, model inference, human correction, or external verification. It should also carry version lineage: what changed, when, why, and under whose authority.

The European Union’s AI Act already points toward traceability as a governance value. High-risk AI systems must technically allow automatic logging over the lifetime of the system, with logs supporting risk identification, post-market monitoring, and operation monitoring. The point here is not that every writing assistant or internal workflow tool is automatically a high-risk system. The point is that legal governance is already moving toward lifecycle traceability. For human-facing AI tools in administrative and professional settings, claim-state traceability should be treated as part of that same family of controls.

A practical rule set

A current-evidence-state rule could be operationalized in five parts.

First, retrieval before demotion: if the model has access to the connected source, it should retrieve before telling the user that a point is unverified. “Unknown to me right now” is not the same as “unverified.”

Second, bounded correction: when one field is unsupported, correct that field. Do not metastasize uncertainty across the entire file.

Third, burden of reversal: once a fact is established through the agreed evidence hierarchy, demotion requires a stated reason, a cited conflict, a newer source, or a corrected interpretation of what the earlier source proved.

Fourth, source-state language: systems should use distinct labels for not retrieved, retrieved but incomplete, contradicted, superseded, established, and inference only. These are legally and administratively different states.

Fifth, accommodation preservation: where a person uses AI as an accessibility support, the system should preserve continuity in a way that reduces, rather than intensifies, access labour. This includes continuity packets, claim ledgers, version history, and explicit human-review gates.

Policy consequence

This is not a plea for credulity. It is the opposite. I want systems that double-check properly. Search the email. Open the document. Compare the dates. Find the contradiction. Correct me when the record says I am wrong. But do not make the human prove what the system already has access to merely because caution is the easiest sentence to generate.

The next generation of AI governance should treat repeated state-loss as an access, accountability, and recordkeeping problem. Hallucination asks whether the machine invented something. The verification tax asks whether the machine erased the work already done, then made the person pay again to restore it. That is a different harm. It needs its own control.

Selected sources

  • Accessible Canada Act, S.C. 2019, c. 10.
  • Accessibility Standards Canada, CAN-ASC-6.2:2025, Accessible and Equitable Artificial Intelligence Systems.
  • European Union Artificial Intelligence Act, Regulation (EU) 2024/1689, including recordkeeping and traceability obligations for high-risk AI systems.
  • Jackie Kay, Atoosa Kasirzadeh, and Shakir Mohamed, “Epistemic Injustice in Generative AI,” AAAI/ACM AIES 7.1 (2024): 684-697.
  • Yusuf Damilola Olaniyan, Mercy Onyemaechi Martins, and Rahab Harib Al Maqrashi, “Generative artificial intelligence and epistemic (in)justice,” Frontiers in Human Dynamics 8 (2026).
  • Theodoros Kouros, “From objectivity to obedience: LLMs as discourse, discipline, and power,” AI & Society 41 (2026): 5883-5892.