Semantic Integrity and Public Communication Assurance
A research programme examining when official language permits materially misleading interpretations; who is responsible for prevention and correction; and how AI and human review can make public information clearer without erasing legitimate uncertainty.
Conceptual research framework · 5 November 2026 · Source-grounded case studies distinguished from constructed examples and untested proposals
Core principle: clarity is a duty of publication, not only reading
A public communication should not pass assurance merely because its author can defend one intended reading. It should also be assessed for reasonably foreseeable, materially false impressions among the populations whose rights, duties, safety, finances or decisions the communication can affect. Readers can misunderstand clear text, but publishers also have a duty to anticipate plausible harmful interpretations and correct avoidable ambiguity.
Ambiguity is not identical to factual falsehood; materially misleading effect is not identical to proven deliberate deception. A rare fanciful reading should not automatically disqualify a sentence. The relevant test is whether a significant misunderstanding is plausible in context, consequential, reasonably foreseeable and remediable through proportionate changes. The original record and correction history must be retained.
AI responsibility: a drafting or review model may contribute to errors or omissions, but is not a substitute for a named human content owner, legal reviewer and accountable institution. A model's confident suggestion is a hypothesis for validation, not independent evidence that the public is misled. Government's own AI Playbook discusses hallucination, human oversight and organisational accountability.
New research journal: frontier studies and engineering gaps
The companion 2025–26 Semantic Integrity Frontier Research Register examines fifteen source-specific research programmes, their reported advances and limitations, multilingual ambiguity, pre-publication prevention, the previously proposed Centres of Truth evidence network and bounded next experiments. These are research observations and proposed tests, not validated Manifold Grace implementations.
Taxonomy of consequential wording failures
| Failure mode | Example mechanism | Assurance question |
|---|---|---|
| Lexical polysemy | “Service”, “entitlement” or “priority” has multiple senses | Which meaning is operative and could another lead to a wrong action? |
| Missing referent | “You are not eligible” without naming what for | Who or what does each term refer to? |
| Normative vs descriptive meaning | A recommendation appears mandatory | Is this law, policy, advice, prediction or aspiration? |
| Scope and quantified claims | “All”, “usually”, “may” or “only” is vague | To whom, where and when does it apply? |
| Agency and responsibility inversion | The institution's duty becomes the user's supposed fault | Which party owes which duty under which authority? |
| Temporal ambiguity | Unclear start date, superseded requirements | What was true as of the relevant event or decision? |
| Jurisdictional conflation | UK-wide voice for England-only rule | Which nation or authority applies? |
| Omitted conditions | A true statement masks decisive exceptions | Could missing information materially alter a decision? |
| Misleading visual hierarchy | Warning headline dominates qualifying footnote | What impression does the complete interface create? |
| AI-generated assertion | Invented source, false confidence, shifted meaning | Is each operative claim grounded and independently checked? |
| Accessible-language mismatch | Correct terminology unusable for real readers | Can intended users accurately act on the information? |
| Strategic reframing or quotation | Context-free extract reverses subject or responsibility | Would readers who see a snippet infer something materially different? |
Eight sourced case studies and institutional examples
Evidence labels matter: some entries are historical cases with independent criticism or inquiry evidence; others are an analysis of current text or official risk and prevention guidance. No constructed harm is represented as an observed incident.
1. Ministerial Code: public service, privilege and entitlement
Documented basis: The June 2026 Ministerial Code foreword frames service in office as a privilege, contrasted with politicians treating service as an entitlement; clause 1.5 specifies ministerial office.
Potential mechanism: The noun phrase “public service” can refer to holding office, delivering services, or services received. A detached excerpt may blur whose entitlement is in question.
Consequential risk and evidence limit: A reader might wrongly infer that lawful service users should feel undeserving; that reading is not established as prevalent and is weaker in the original surrounding context.
Possible clearer wording / control: Holding ministerial office is a privilege carrying duties to the public. This does not diminish the legal entitlements of people eligible for public services.
Research test: Compare full-context and detached-excerpt comprehension; separately ask what the statement means for ministers and for service users.
2. Pandemic slogans: “Stay home” versus “Stay alert”
Documented basis: A 2020 King's College London analysis reported that only around 30% in an initial poll felt the new Stay Alert message clearly communicated what they should do.
Potential mechanism: Replacing an executable instruction with an open-ended behavioural metaphor makes the action criterion uncertain.
Consequential risk and evidence limit: Different readers can choose different behaviours while believing they complied; downstream health impact cannot be quantified from a slogan alone.
Possible clearer wording / control: State the actual required or recommended actions, the date and the jurisdiction; distinguish legal rules from advice.
Research test: Run task-based comprehension and behavioural-choice tests across reading levels and languages.
3. Four nations: cross-jurisdictional restrictions
Documented basis: A joint Commons committee report describes divergent messages across UK nations during the May 2020 easing, and evidence reporting confusing or apparently contradictory statements.
Potential mechanism: A UK-wide political or media frame can obscure which jurisdiction's instructions apply.
Consequential risk and evidence limit: Readers may apply rules from another nation; assessment must distinguish genuine contradiction from different valid local policies.
Possible clearer wording / control: For England only, from [date], [rule]. Different requirements may apply in Scotland, Wales and Northern Ireland; consult the relevant administration.
Research test: Use mixed-location comprehension tasks and verify users identify jurisdiction, effective date and legal status.
4. Guidance presented as law
Documented basis: The House of Lords Constitution Committee identified examples where public-health advice, guidance and regulations were conflated or guidance was enforced as law.
Potential mechanism: Imperatives such as must, should, required and advised are applied without naming the legal instrument.
Consequential risk and evidence limit: People may comply under a mistaken view of legal penalties or challenge actual duties incorrectly.
Possible clearer wording / control: Specify whether a requirement is law or nonbinding guidance, the legal source, exceptions, enforcement body and commencement date.
Research test: Compare legal-status classification accuracy for original versus redrafted notices.
5. A government AI chatbot giving false service advice
Documented basis: The 2025 UK AI Playbook explicitly warns that an LLM-powered public chatbot could hallucinate and mislead a person about a government service, policy or legal point; it says organisational responsibility remains.
Potential mechanism: Fluent generation can insert invented eligibility conditions, omit exceptions or turn a probabilistic inference into asserted fact.
Consequential risk and evidence limit: Users may miss entitlements, deadlines, appeals or needed support, especially if they reasonably trust an official interface.
Possible clearer wording / control: Use retrieval from current authoritative rules, disclose uncertainty, provide official source and human correction paths; do not present unverified advice as a determination.
Research test: Test with known eligibility edge cases, conflicting policy versions, out-of-scope questions and appeals.
6. Home Office word choice: “abroad” and “overseas”
Documented basis: The Home Office content guide specifically notes that abroad and overseas can confuse and suggests clearer contextual alternatives.
Potential mechanism: A relative term may fail to specify the relevant legal geographic boundary, place of application or residence.
Consequential risk and evidence limit: A person can select the wrong application route or misunderstand whether their circumstances qualify.
Possible clearer wording / control: Apply from outside the UK; or work in another country, as appropriate to the rule.
Research test: Give representative users realistic application scenarios and compare route-selection errors.
7. Government AI accountability and validation
Documented basis: The AI Playbook calls for human control at relevant stages, testing of outputs and traceable accountability across the lifecycle.
Potential mechanism: AI-origin text may be published without verifying interpretation, temporal accuracy, legal provenance or the user impact of wrong advice.
Consequential risk and evidence limit: The institutional publisher can inadvertently amplify automated inaccuracies at scale.
Possible clearer wording / control: Require named content owner, source citations, risk-scoped user testing, audit trace and documented corrections before publication.
Research test: Measure AI and human review detection rates against independent expert labels.
8. ONS stakeholder check for misleading or missing material
Documented basis: The ONS content guide asks stakeholders to identify content that is incorrect, misleading or missing, while peer reviewers assess clarity and structure.
Potential mechanism: A factual sentence can mislead through omitted qualification, inconsistent headings, missing context or implied scope.
Consequential risk and evidence limit: Literal correctness may pass a narrow fact check while the whole user journey conveys a false operational impression.
Possible clearer wording / control: Add a separate material-misinterpretation check alongside factual verification and accessibility testing.
Research test: Use blinded reviewers to identify omissions, then track error reduction in the published user journey.
Twelve constructed test scenarios across public services
These examples are drafted for testing, not allegations that any named authority issued the wording. The point is to reveal the sorts of misunderstandings that content reviewers and AI systems could detect before publication.
1. Benefits decision notice
Illustrative wording: “You are no longer entitled.”
Potential ambiguity: Unspecified benefit, date, legal basis or appeal rights
Revision principle: Name the exact entitlement, decision period, reasons and review route
Intersections: Eligibility law; administrative justice; accessibility
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
2. University assessment notice
Illustrative wording: “Your attempt was unsuccessful.”
Potential ambiguity: Could mean fail, invalid upload, capped result or missing work
Revision principle: Identify the recorded submission, mark, decision rule and permitted remedy
Intersections: Education policy; record systems; disability law
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
3. Healthcare appointment
Illustrative wording: “Your referral is not urgent.”
Potential ambiguity: Could be perceived as denial of medical need
Revision principle: Explain triage category, clinical next step, time horizon and escalation route
Intersections: Clinical communication; health literacy; safety
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
4. Housing allocation
Illustrative wording: “You do not qualify for priority.”
Potential ambiguity: Priority scheme, period and appeal rights unstated
Revision principle: State criteria, evidence evaluated and challenge mechanism
Intersections: Housing law; social policy; public administration
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
5. Tax notification
Illustrative wording: “You may owe additional tax.”
Potential ambiguity: Unclear whether liability is confirmed or speculative
Revision principle: State calculation status, relevant tax period and required action
Intersections: Fiscal administration; legal informatics
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
6. Immigration status
Illustrative wording: “Further action may be taken.”
Potential ambiguity: Could imply imminent removal or merely further review
Revision principle: Specify authority, possible actions, safeguards and timeframe where lawful
Intersections: Immigration law; rights; procedural fairness
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
7. AI risk dashboard
Illustrative wording: “Low risk.”
Potential ambiguity: Metric, group, model version and timeframe unknown
Revision principle: Name evaluated risk, observed limits, affected cohorts and review date
Intersections: Model evaluation; statistics; safety engineering
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
8. Public consultation
Illustrative wording: “Your views will be considered.”
Potential ambiguity: May suggest individual reply or binding influence
Revision principle: Explain process, deadline, synthesis and limits on decision influence
Intersections: Democratic participation; deliberation
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
9. Police/public order guidance
Illustrative wording: “You should cooperate.”
Potential ambiguity: May blur legal duties, voluntary requests and rights
Revision principle: Separate legal obligation from request and cite rights and authority
Intersections: Criminal procedure; institutional communication
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
10. School admissions
Illustrative wording: “Places are available.”
Potential ambiguity: May imply automatic offer despite selection rules
Revision principle: Clarify number, cohort, criteria, dates and actual decision process
Intersections: Education access; service design
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
11. Employment recruitment
Illustrative wording: “Equivalent experience accepted.”
Potential ambiguity: No specification of equivalence, evaluation or evidence
Revision principle: Describe acceptable routes, assessor criteria and appeal channels
Intersections: Labour law; HR; recognition of prior learning
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
12. Public funding notice
Illustrative wording: “Funding is available to eligible projects.”
Potential ambiguity: Unclear status of funding pool or applicant conditions
Revision principle: Specify confirmed budget, restrictions, assessment route and allocation timing
Intersections: Public finance; grants governance
Evidence status: No claim is made that a particular department issued this sentence or caused a documented harm.
Twenty-eight research intersections: disciplines, methods and operational roles
The field spans interpretive humanities, empirical social sciences, law, assurance engineering and public-service practice. Each intersection describes a contribution to the research problem, not a claim that all are mastered by a single researcher or already implemented on this website.
| No. | Research intersection | Question or method contributed |
|---|---|---|
| 1 | Semantics and lexicography | Polysemy, scope, reference, vagueness, modality and dictionary senses |
| 2 | Pragmatics and discourse analysis | Implicature, framing, speaker intention, audience interpretation and omission |
| 3 | Psycholinguistics and comprehension | How people actually understand text under time, stress and literacy constraints |
| 4 | Cognitive science and behavioural economics | Framing effects, reliance, defaults, authority cues and interpretation |
| 5 | Public law and statutory interpretation | Difference between guidance, binding rules, rights, duties and delegated powers |
| 6 | Administrative justice and ombuds practice | Decision notices, reasons, redress and procedural fairness |
| 7 | Human rights and equality law | Foreseeable differential consequences for protected and vulnerable groups |
| 8 | Disability studies and accessibility | Plain language, screen readers, cognitive accessibility and supported comprehension |
| 9 | Sociolinguistics and multilingual communication | Dialect, translation, legal vocabulary and cultural interpretations |
| 10 | Government communication and public administration | Approval chains, public-interest duty and content-design controls |
| 11 | Public policy and evaluation | Counterfactual harms, distributional impacts and policy effectiveness |
| 12 | Health communication and risk science | Clarity of protective action and distinction between advice and instruction |
| 13 | Education and assessment | Validity of evaluation decisions and student-facing procedural language |
| 14 | Labour markets and recruitment | Ambiguity in entry criteria, equivalence and professional recognition |
| 15 | Human-computer interaction and UX research | Task completion, error rates, navigation and situated user testing |
| 16 | NLP and linguistic annotation | Ambiguity detection, entity resolution, scope and modality classification |
| 17 | LLM evaluation and calibration | Hallucination, uncertainty, factual grounding and multiple plausible readings |
| 18 | Knowledge graphs and provenance | Mapping every communication claim to its authoritative source and valid date |
| 19 | Causal inference and field experimentation | Showing whether wording caused differences in actions, not merely correlation |
| 20 | Statistics and measurement theory | Operationalising materially misleading interpretation and uncertainty intervals |
| 21 | Cybersecurity and information integrity | Prompt injection, deceptive AI outputs and trusted-publication pipelines |
| 22 | Organisational governance and QA | Roles, handoffs, approval gates, incident learning and version histories |
| 23 | Design ethics and philosophy of language | Responsibility for foreseeable meaning, rights and epistemic duty |
| 24 | Democratic theory and political accountability | Citizen understanding, legitimacy, scrutiny, non-partisan standards |
| 25 | Computational law and legal informatics | Rule-linked eligibility language and changes in statutes over time |
| 26 | Archival science and open records | As-of-date wording, retractions, corrections and evidence preservation |
| 27 | AI assurance and safety engineering | Risk thresholds, red-teaming, rollback and independent certification |
| 28 | International comparative governance | Cross-jurisdictional differences and portability limitations |
Proposed publication assurance pipeline
- Source control: store document, author, approver, jurisdiction, purpose, audience, policy source, effective dates and prior wording.
- Claim extraction: identify operative statements about rights, eligibility, duties, safety, sanctions, deadlines, costs or access.
- Lexical and semantic audit: use dictionaries, legal definitions and domain terminology to test alternate meanings; inspect syntax, reference, scope, modality and omissions.
- Contextual adversarial reading: test the whole page, snippet, notification, letter and translation; distinguish plausible from fanciful interpretations.
- Risk triage: identify material impact, affected users, probability, reversibility and any severe credible downside; green = sufficiently supported, amber = unresolved, red = foreseeable serious harm requiring correction.
- Independent human validation: content designers, source owners, legal and accessibility specialists assess AI-generated flags, retaining dissent and uncertainty.
- Audience comprehension: test with realistic tasks and representative users, including those with accessibility needs and low prior knowledge.
- Constrained redrafting: use precise descriptive wording without falsely simplifying legal complexity or changing a policy's meaning.
- Publication gate: publish when material issues are resolved or appropriately disclosed; retain named accountability and evidence of sign-off.
- Post-publication correction: preserve previous versions, report discovered harm, correct promptly, monitor recurrence and provide redress where appropriate.
Automation boundary: an AI flag is not a legal judgement, and automated replacement should not overwrite the authoritative original without human approval. Information security and confidential draft handling are part of the assurance design.
Proposed experiments and measurable outcomes
| Study | Design | Metrics and failure conditions |
|---|---|---|
| E1. Alternate interpretation extraction | Blind expert annotation of 100+ independently selected public communications, including control text with no material ambiguity. | Precision/recall of plausible harmful readings; false positives on clear text; annotation agreement. |
| E2. Wording and behaviour | Randomised A/B comprehension tests comparing authentic text with independently approved clearer versions. | Correct action selection, legal-status understanding, time to task completion, confidence calibration, accessibility breakdown. |
| E3. AI vs human review | Compare reviewers alone, AI alone and AI-supported reviewers on blinded materials. | Detection of material errors, time, cost, false reassurance, hallucinated criticism and reviewer dependence. |
| E4. Law and time | Versioned notices linked to laws in force on historical dates, then changed. | As-of-date source accuracy, version fidelity, outdated advice rate and correction latency. |
| E5. Accountability reversal | Evaluate wording that changes whether the public or institution appears to bear a duty. | Attribution of legal responsibility and perceived access rights, separately from political agreement. |
| E6. Equity and language | Recruit varied readers and test translation, literacy and assistive formats with safeguards. | Subgroup error rates, false-denial perceptions and uncertainty; do not infer universal access from one cohort. |
| E7. Real-world pilot | Partner department or independent organisation tests a review gate prospectively. | Confirmed correction rates, avoidable complaints, user task failures and ongoing operating burden; require consent and permissions. |
No trials or validation outcomes are claimed here. Pre-register selection and scoring rules, include both serious counterexamples and clear-writing controls, publish failures, and distinguish AI-detected risk from measured reader misunderstanding and evidenced consequential harm.
Responsibility and remedy
Reader: may need to seek clarification where text is clearly incomplete; this does not eliminate publisher obligations. Author/editor: should provide the degree of precision reasonably required for the context and audience. Publisher and approving organisation: own the publication process, its controls and timely corrections. AI developer/deployer: may contribute design, training, testing or deployment failures; legal responsibility depends on the applicable facts and law, while public organisations remain responsible for their communications.
Do not confuse accidental ambiguity, objectively misleading effect, negligent review and intentional deception. Only the latter requires evidence of intent; all can motivate proportionate prevention or correction. A reader's unexpected interpretation is a signal to investigate—not automatic proof of publisher misconduct.
Existing standards, research lineage and sources
Government standards already provide a foundation: GOV.UK design principles, GOV.UK writing for interfaces, ONS review guidance, Government Communication Service ethics and the government AI Playbook. This proposal extends these with a specifically measurable material-misinterpretation and authorial accountability research architecture.
Manifold Grace intersections: Information Science; Cross-sector Evidence Verification; AI Compliance and Governance; AI Inventory; Interdisciplinary Atlas; Research Gaps; Leadership and Overreach; Life Philosophy; and Green–Amber–Red economic risk research.
Originality and evidence note: The integrated project framing was developed through user-led Manifold Grace discussions, with AI assistance in research and drafting. It does not claim to invent plain-language guidance, discourse analysis, red-teaming or responsible AI. Case-study descriptions paraphrase cited sources and preserve distinctions between observed evidence, source commentary, hypothetical risks and proposed experiments. No government approval, independent validation or measured reduction in misleading communications is claimed.