− news and insights

Insights

− Independent reading

Why AI claims need governance

Reporting, research, rulings and guidance from Switzerland and around the world. Each source shows a real problem. The Why Validix note explains the control that addresses it and links to the relevant walkthrough.

− Before generation

01. What can AI claim?

Authoritative evidence defines which claims AI can make and the conditions that must travel with them.

SRF Switzerland · Reporting · 2025

When AI misreads a real Swiss news source

An AI answer cited SRF’s reporting on Swiss exports and added an import claim the source did not support.

Why Validix
Validix checks the direction of the assertion against the evidence. A passage about exports cannot support a claim about imports.

Min et al. · EMNLP Global · Research · 2023

FActScore and the case for checking individual facts

FActScore breaks generated text into individual facts because one passage can mix supported and unsupported assertions.

Why Validix
Validix checks assertions individually and keeps qualifications attached, so a supported sentence cannot carry an unsupported claim through review.

Gao et al. · EMNLP · Research · 2023

Do citations support what the model actually says?

The ALCE benchmark evaluates citation quality alongside answer correctness.
Why Validix Validix checks whether the cited passage actually supports the assertion. A real citation can still be the wrong evidence for that wording.
Niu et al. · ACL Global · Research · 2024

Hallucinations in retrieval-augmented generation

RAGTruth documents unsupported or contradictory text even when a model has retrieved relevant material.

Why Validix
Validix uses governed evidence to constrain drafting and separately checks what the model actually writes against that evidence.

Columbia Journalism Review · Research · 2025

AI search has a citation problem

Tests of eight AI search tools surfaced errors in identifying and citing news.

Why Validix
Validix ties a governed assertion to a specific source passage so a reviewer can inspect what the reference actually supports.

OpenAI United States · Research · 2025

Why language models hallucinate

The authors argue that evaluation can reward models for guessing over admitting uncertainty.
Why Validix When the available evidence cannot support an assertion, Validix can withhold it and route the draft for review.
Cambridge · AVeriTeC · Research · 2024

Claim verification requires a verdict and usable evidence

The shared task scores both the verdict and the quality of the evidence used to verify a claim.
Why Validix Validix keeps the decision and its supporting passage together. Reviewers can inspect the basis for a claim instead of reconstructing it from a verdict alone.
Song et al. · EMNLP Global · Research · 2024

VeriScore focuses on claims that can be checked

VeriScore distinguishes verifiable assertions from other material in generated text.

Why Validix
Validix checks the assertions a draft makes and the conditions that govern their use. Reviewers can focus on what the text actually claims.

Swissmedic Switzerland · Regulatory

How Swissmedic checks medicinal product advertising

Swissmedic checks promotional material against the approved information for the specific medicinal product.
Why Validix Validix can scope drafting to evidence for the applicable product and market, then flag wording that widens a claim beyond those conditions.
U.S. SEC United States · Regulatory

Substantiating factual claims in investment adviser advertising

The SEC guide explains the need for a reasonable basis to substantiate a material factual statement in covered advertising.
Why Validix Validix preserves the source behind a specific assertion so reviewers can retrieve the support for the wording they are asked to review.
U.S. FDA United States · Regulatory

Truthful and balanced prescription drug promotion

The FDA describes its review of prescription drug promotion for accuracy and balance.

Why Validix
Validix can check whether a generated claim preserves material qualifications from its source. The applicable human review still determines release.

Dhuliawala et al. · ACL Global · 2024

Testing a separate verification pass

Chain-of-Verification studies whether a separate verification pass can reduce hallucinations in model drafts.


Why Validix
Validix separates drafting from verification and checks the output against governed evidence and permissions.

− Before RELEASE

02. What actually gets released?

The wording AI actually produces needs its own check against the evidence and the rules for using it.

MIT Sloan Management Review · 2026

The real question to ask about AI governance

Joseph Wallace asks who has authority to intervene when AI goes wrong.


Why Validix
Validix gives teams a point to hold a specific unsupported assertion before release and a record reviewers can use to make the next decision.

FINMA Switzerland · Regulatory · 2024

AI governance and risk management in Swiss finance

FINMA identifies model correctness and explainability as risks for supervised firms to address.

Why Validix
Validix gives a reviewer the wording, its supporting evidence and the reason for the release decision. Those details make the decision easier to reconstruct.

MIT CISR United States · Analysis · 2026

Establishing guardrails on the AI roller coaster

MIT CISR examines how organisations can make safeguards reusable as AI initiatives proliferate.
Why Validix Validix lets teams reuse governed claims across content and channels while checking each output against the conditions for its use.
NIST United States · Standards · 2024

Generative AI risk profile

NIST’s generative AI risk profile describes confidently wrong content and fabricated citations.
Why Validix Validix checks what a draft actually asserts against the evidence that was permitted for that use. Unsupported wording can be held before it reaches a user.
Stanford Law School · Research · 2025

Testing the reliability of AI legal research tools

The researchers found hallucinations in the legal research tools they evaluated, including tools using retrieval.

Why Validix
In a legal workflow, Validix’s approach would check whether the cited authority supports the precise proposition, including relevant jurisdiction and scope.

U.S. District Court · S.D.N.Y. · 2023

Mata v. Avianca and fabricated legal authorities

The court sanctioned lawyers after fake judicial opinions appeared in filings.
Why Validix In a filing workflow, Validix’s approach would require a claim to trace to a governed source before it could move forward for review.
High Court of England and Wales · 2025

Ayinde and Al-Haroun on fictitious authorities

The judgment examines citations that were fictitious or failed to support the proposition.
Why Validix Validix’s approach checks two things separately in this example. Is the authority in the governed evidence, and does the passage support the proposition?
Global News Canada · Reporting · 2024

When a chatbot promised the wrong fare terms

A customer relied on incorrect information from Air Canada’s chatbot about bereavement fares.

Why Validix
Validix can compare a customer-facing answer with the applicable policy and hold wording that promises an exception the policy does not support.

Associated Press Australia · 2025

Incorrect citations in a government-commissioned report

AP reported fabricated references and a partial refund after Deloitte corrected its report.
Why Validix Validix ties a draft assertion to the exact supporting passage. A convincing footnote with no supporting passage leaves the assertion open for review.
The Markup · THE CITY US · Reporting · 2024

Incorrect advice from New York City’s business chatbot

Reporters found New York City’s business chatbot misstating rules for business owners.

Why Validix
Validix can check whether the rule applies to the case and whether the answer preserves its exceptions before the wording is used.

European Broadcasting Union · BBC · 2025

News integrity in AI assistants

Public broadcasters tested four AI assistants on current affairs and examined problems with accuracy, sourcing and context.

Why Validix
Validix checks the assertion together with its source and qualifications, so a summary can be held if it changes who, what or when the source describes.

Swiss Federal Chancellery · Regulatory · 2024

Large language models in the Swiss Federal Administration

The factsheet warns that generated text can contain factual errors and calls for quality assurance.

Why Validix
Validix gives reviewers a concrete task before publication. Compare the generated assertion with its evidence, inspect any mismatch, and decide how to proceed.

NAIC United States · Regulatory · 2023

Model bulletin on the use of AI by insurers

The model bulletin sets out governance expectations for insurers’ use of AI.

Why Validix
For an insurance communication, Validix can check a coverage statement against the applicable policy version and confirm that the answer carries the exclusion.

European Broadcasting Union · BBC · 2025

Governance of large multimodal models in health

WHO discusses risks from false, inaccurate or incomplete generated statements in health.

Why Validix
Validix can check whether the wording preserves the studied population and the endpoint before the draft goes to its reviewers.

Council of Europe Europe · Regulatory · 2024

Framework Convention on Artificial Intelligence

The convention addresses transparency, oversight and AI risk management.

Why Validix
Validix gives a reviewer a trace from the output to the applicable evidence and decision. That makes it possible to inspect why a specific assertion was held.

Swiss Federal Council · Regulatory · 2026

Switzerland’s sector-specific work on AI regulation

The federal update describes sector-specific work on AI regulation, including in health, transport and media.

Why Validix
Validix can use sector- and audience-specific permissions to scope drafting, then check the wording against the conditions that apply to that use.

Financial Stability Board · Analysis · 2024

The financial stability implications of AI

The FSB examines financial-sector vulnerabilities including model risk and data quality.

Why Validix
In a financial communication, Validix can check whether a draft turned a conditional result into an unconditional guarantee before it reaches a customer.

− AFTER RELEASE

03. What happens when the evidence changes?

When a source is revised, dependent claims and tracked content can be flagged for reassessment.
Swissmedic Switzerland · Regulatory

Therapeutic product safety updates

Swissmedic publishes new therapeutic product safety information for healthcare professionals.

Why Validix
When a new notice changes the evidence for a claim, Validix’s dependency map shows which claims and tracked materials should be reassessed.

NICE United Kingdom · Guidance

How NICE checks whether published guidance needs updating

NICE assesses new evidence and changes in the health system after publishing guidance.

Why Validix
Validix retains the link between a claim and the source version it used. Reviewers can use that link to find affected claims when guidance is updated.

Vu et al. · ACL Global · Research · 2024

Testing models on facts that change over time

FreshQA tests how well AI handles questions that depend on current knowledge.

Why Validix
Validix uses evidence versions and effective dates to determine which claims are in scope at the time of use.

EMNLP Global · Research · 2023

When models repeat outdated facts

This study explores how models can avoid repeating facts that have become outdated.

Why Validix
Validix can flag governed claims that still depend on an old source version and send their tracked uses back for review.

European Commission EU · Regulatory

Post-market monitoring for high-risk AI systems

Article 72 describes post-market monitoring for providers of high-risk AI systems.

Why Validix
At the claim level, Validix is designed to connect post-release changes to the claims and tracked materials affected by them.

NIST United States · Standards

AI Risk Management Framework Playbook

The AI RMF Playbook includes activities for AI operation and monitoring.

Why Validix
Validix’s dependency links help teams narrow a review to the claims and tracked outputs tied to the source that changed.

Australian Government Australia · 2025

A revised assurance report replaces the published original

The department published a revised assurance report and identified the version it replaced.

Why Validix
When a revised report supersedes an earlier version, Validix can point reviewers to tracked summaries built on the earlier version.

NHMRC Australia · Guidance

Updating guidance when benefits and harms change

The council describes why guidelines may need updating, including changes in evidence on benefits and harms.

Why Validix
Validix’s claim dependencies help reviewers find which statements about a benefit or harm relied on the evidence that changed.

CISA United States · Advisory · 2026

New evidence changes which vulnerabilities are prioritised

CISA adds vulnerabilities to its catalogue based on evidence of active exploitation.

Why Validix
In a security workflow, Validix’s approach would flag tracked summaries that depend on the affected catalogue entry when it is updated.

Shekelle et al. · JAMA US · Research · 2001

How quickly can clinical guidance become outdated?

The study examined when previously issued clinical guidance needed revision.

Why Validix
Validix is designed to preserve the link from a claim to its evidence version after release. Teams can use that record when earlier guidance is superseded.

− Continue exploring

Our Edge → 

Inspect a decision and see why a claim is blocked.

Manifesto → 

Follow the effects of a source change.