Teaching students to question sources offers a lesson for AI in public service
An AI helpdesk can cite a real government page and still give a resident the wrong guidance. A teacher’s classroom habit of questioning sources points to a four-field review card for public sector teams.
-1794444189740.jpg)
Digital responsibility is when students learn to ask what a source actually establishes, not simply accept a confident-looking citation. Image: Canva
An artificial intelligence (AI) assistant can give a resident a fluent answer, cite an official website and still provide unsafe guidance.
The link may be genuine, but the page could concern a different service, a superseded procedure or people outside the relevant jurisdiction.
In public service, the question is not simply whether a cited document exists, but whether that document supports this particular answer for this particular person and at this particular time.
That distinction is familiar in education.
In my work as a computer science teacher, source evaluation is part of how I approach digital responsibility.
Digital responsibility is when students learn to ask what a source actually establishes, not simply accept a confident-looking citation.
The same habit matters in public administration, where an apparently minor mismatch can send a resident to the wrong office or cause an unnecessary application.
A plausible answer can have the wrong evidence
Consider a hypothetical council helpdesk. A resident asks which documents are needed to access a public service. An AI system presents a tidy checklist and links to an authentic government page.
But the page applies to a different group of applicants. The words may sound reasonable; the citation may even resolve correctly. Neither observation establishes that the checklist is applicable.
A staff member needs a review process that works without a specialised AI laboratory, a new procurement or a claim that software can replace administrative judgement.
Start with one high-volume, low-risk category of enquiry and use a compact four-field review card.
The four-field review card
1. The claim. Copy the exact statement that the AI proposes to send to the resident.
Break multi-part guidance into individual claims where necessary. A generic answer is too broad to verify in a single pass.
2. The supporting passage. Record the exact words in the primary source that supposedly justify the claim, plus its URL or document identifier.
If the link opens but the passage does not support the claim, mark the evidence as insufficient rather than assuming the answer is valid.
3. Applicability and version. Check the date, jurisdiction, eligibility group, exceptions and whether the source has been superseded.
A genuine source about a neighbouring jurisdiction or an older rule cannot automatically support current local guidance.
4. Human accountability. Name the officer or team authorised to approve the advice, correct it or escalate the question. The model should not be treated as the responsible decision-maker.
Each field should be completed in ordinary language. Where a decisive field is missing, do not issue the AI-generated guidance as a settled answer.
Route the query to a competent person and communicate the uncertainty clearly to the resident.
What a manager could try tomorrow
A service manager could pilot the card with a small sample of anonymised, previously answered questions.
For each question, compare the proposed AI response with the applicable primary guidance and
document the result.
Keep identifiable resident information out of any experimental AI tool unless appropriate data-protection controls and authorisations are already in place.
Record three practical measures: how often the citation directly supports the claim, how often the source is current and applicable, and how often a human reviewer must correct or escalate the answer.
Review disagreements with the relevant policy owner. The aim is not to manufacture a confidence score, but to reveal recurring failure patterns and improve staff guidance.
Consider adding a 'stop condition' for enquiries involving eligibility, legal deadlines, financial entitlements or other consequential decisions.
An incomplete source, conflicting documents or unclear jurisdiction should trigger human review before the information reaches the citizen.
No automated checklist eliminates the need for lawful procedures and established accountability.
From classroom literacy to public sector assurance
The classroom parallel is useful because the lesson transfers: 'Can you show me the passage that supports that conclusion?' is a stronger question than 'Did the chatbot give me a source?'
Educators can cultivate this habit, while public-service leaders can turn it into a repeatable review step and auditable record.
Frameworks such as the US National Institute of Standards and Technology's AI Risk Management Framework and its Generative AI Profile provide broader risk-management context.
The four-field card is my own practical proposal; it should not be represented as a mandatory NIST procedure.
A trustworthy public service does more than provide a plausible answer and a convincing link.
It can explain which passage supports a decision, why that source applies, and who remains accountable when the evidence is incomplete.
That is a small, testable improvement public-sector teams can begin exploring without surrendering judgement to the machine.
AI-assisted drafting and language editing were used in preparing this guest column.
------------------------------------------------
The author is a computer science teacher and independent researcher based in Italy, with a degree in Business Administration.
His work explores AI literacy, source evaluation and experimental approaches to checking the relationship between AI-generated claims and their supporting evidence. He can be reached at michelevaldelli@gmail.com