Using AI at Work and School Without Outsourcing Your Judgement

The standard advice on using AI is "check the output". It is useless advice. Nobody has time to verify everything, and the parts most in need of checking are precisely the parts that look most convincing.

What follows is a working method instead — the three points in your workflow where checks actually belong, what to decide at each, and how to tell when you have crossed from using the tool to being used by it.

The paper that was never written

Ask an AI for a source and it will give you one. Author, title, journal, year, page number. Everything correct except the part where the paper exists.

This is not a bug awaiting a patch. It is a by-product of how the technology works. A model will sometimes tell you it is unsure — but where it lacks a reliable answer, it may instead produce something with the shape of one, because shape is what it was trained to generate. Fluency and accuracy are separate qualities, and these systems are far better at the first than the second.

The clearest illustration came in 2025, when Deloitte Australia delivered a report to the federal Department of Employment and Workplace Relations under a contract worth A$440,000. A University of Sydney researcher reading it found references to academic papers that had never been written — several attributed to colleagues whose work he knew well — plus a quote from a Federal Court judge that appeared nowhere in the judgment.

Deloitte published a revised version, disclosed that a generative AI tool had been used, and repaid the final instalment of its fee. The department said the substance and recommendations were unchanged.

The model failed predictably. Nobody caught it.

Nobody involved intended to submit fiction. That is what makes this worth your attention rather than your contempt. The output looked competent, so it was treated as competent, and the review step that existed on the org chart produced nothing in practice.

A hallucination is a model failure you should expect. Letting it into finished work is a control failure you own.

The model's failure was the predictable one. The consequential failure was human: treating a draft as a finding. And that decision is not rare or exotic. You make it every time an answer arrives looking finished, the deadline is close, and opening the source would take twenty minutes you do not have.

Singapore adopted AI faster than it supervised it

We are not casual users here. IMDA's Digital Economy Report found roughly three in four surveyed workers already using AI tools at work, most of them several times a week or daily. Separately, Microsoft placed Singapore second globally on its AI diffusion index.

Microsoft's 2026 Work Trend Index supplies the uncomfortable half. Around 88 per cent of Singapore AI users said they remain responsible for the thinking. Only about a quarter said their leadership was clearly and consistently aligned on AI. Individuals sprinted; institutions are still lacing up.

That gap is a personal risk before it is a corporate one. Where no policy exists, the operative rule becomes whatever you happened to do — and you will be judged against it afterwards, by someone with hindsight and a printout.

It has already reached our courts. In March 2026, the Singapore High Court ordered two lawyers to pay personal costs after their submissions cited cases that did not exist. In the same week, the Ministry of Law launched its Guide for Using Generative AI in the Legal Sector, whose first principle is blunt: professionals remain ultimately responsible for their work product. Substitute "student", "analyst" or "engineer" and the sentence holds.

The schools have been equally explicit. MOE told Parliament in May 2026 that its approach rests on four ideas: students should learn about AI, learn to use AI, learn with AI, and — the emphasis is MOE's own — learn beyond AI. Primary 1 to 3 pupils are assigned no work requiring AI. National examinations stay proctored with AI prohibited. Secondary students permitted to use it must declare it and cite sources, and passing off AI output as their own is academic dishonesty. The named concern is cognitive offloading: the thinking you never developed because something else did it for you.

At the other end of the pipeline, NUS made a foundational generative AI module compulsory for every incoming undergraduate from AY2026/27 — covering prompting and evaluation — and opened ChatGPT Edu across the university from 31 August 2026, while its law faculty has moved assessments back toward closed-book conditions.

Unlimited access, verified competence. Nobody is banning the tool. They are removing the excuse.

Before you prompt: decide what kind of task this is

Not every task deserves the same scrutiny, and pretending otherwise is how people abandon verification altogether. Sort the work first.

Low stakes. Brainstorming, rephrasing your own writing, generating counterarguments to your own position, explaining a concept you will test elsewhere. If the output is wrong, you notice and move on. Use freely.

Medium stakes. Drafting emails, summarising documents you have already read, turning notes into prose, code you will run and test. The AI accelerates; you own every line before it ships.

High stakes. Anything asserting a fact to a third party, anything that becomes a record, anything a regulator, examiner, client or court could later hold you to. AI can still help early here — surfacing questions you had not considered, pointing at gaps, suggesting where to look. What it must not do is originate a conclusion you then adopt without independent analysis. Treat everything it gives you as a lead, and do the reasoning yourself.

The second decision is a one-time setup you have probably never done. Most consumer AI tools use your conversations to improve future models unless you tell them not to. Fixing that takes two minutes:

• ChatGPT (Free, Plus, Pro): Settings → Data Controls → switch off "Improve the model for everyone".

• Claude (Free, Pro, Max): Settings → Privacy → switch off "Help improve Claude".

• Gemini: Settings → Activity → turn off Keep Activity.

• 
  

Do it today. Then understand exactly how little it buys you, because this is the point at which people get comfortable and stop thinking.

It is forward-looking only. Nothing already absorbed into a trained model can be pulled back out. The models you are using now were trained on data gathered before you found the setting.

It governs training, not necessarily retention. Turning it off does not make every copy of a conversation disappear. Providers may keep data for service delivery, security, abuse prevention or legal obligations, and the specifics vary by provider, account type and feature — sometimes considerably. Opting out means your prompt should not shape a future model. It does not mean the text is gone. Check the current policy for the tool you actually use rather than treating the toggle as a confidentiality guarantee.

Which brings us to the third decision, and the only one that actually protects you: what you are willing to type. Ask one question — would I be comfortable if this text appeared in an email to someone outside my organisation? If not, do not paste it. Client names, NRIC numbers, salary data, unreleased results, proprietary code, someone else's personal information — none of it belongs in a consumer account.

Anonymise before you prompt. Replace names with placeholders, strip identifiers, describe the shape of the problem rather than its particulars. You will be surprised how rarely the model needed the real details. And where your employer provides a sanctioned enterprise tool, use it — not because it is better, but because somebody has actually read its data-handling terms.

While you work: protect the skill you are being paid to have

There is a difference between using a calculator and never learning arithmetic. The first is efficiency. The second is a career risk that arrives quietly, five years later, in a meeting where you cannot answer the follow-up question.

Notice which capability you are trading away. If you are a junior analyst and AI writes all your first drafts, you never build judgement about what belongs in a draft — the thing that would have made you a senior analyst. If you are a student and AI structures every essay, you have practised prompting, not argument.

The correction is small and specific: do the hard part first, then bring the AI in. Write your own outline before asking for one, so you can see what it changed and why. Attempt the problem before requesting the solution. Form your view before asking for counterarguments.

You keep the productivity gain without the atrophy. More usefully, you become able to tell the difference between AI output that is genuinely better than yours and output that is merely smoother.

Before it leaves your hands: verify, declare, defend

AI can point you to a source. It cannot be one. That single distinction is the whole of the Deloitte lesson.

A model can tell you that a paper, case, statute or statistic probably exists and roughly what it says. That is a lead. It becomes a fact only when you have opened the original — the MAS circular, the SingStat table, the journal article, the actual judgment — and read the relevant line yourself. If a claim carries a number, name, date, citation, or a legal or regulatory consequence, verify it at source before it leaves your hands.

Be most suspicious when the output is specific. Vague AI text is obviously vague. A confidently precise fabrication with a plausible page number is the one that gets through.

Then declare it. Disclosure norms are hardening across Singapore institutions — MOE requires it of secondary students where AI use is permitted, universities require it in assessed work, and MinLaw's guide includes a sample disclosure clause for client communications. Assume the expectation applies to you before someone tells you it does.

Finally, apply the test that covers everything above. Before you submit: if someone questions any sentence in this, can I explain where it came from and why it is right — without going back to the AI?

If yes, you used the tool. If no, the tool used you.

What to do now

If you are a student or an employee:

• Verify every fact, figure, name, date and citation at the primary source. No exceptions for time pressure — that is precisely when this fails.

• Switch off model training in your AI settings today — then keep behaving as though you had not, because it governs training, not retention.

• Anonymise before you prompt. Nothing confidential goes into a personal account.

• Find out what your school or employer's AI policy actually says. If none exists, write your own one-paragraph rule and follow it. A documented habit is a defence; an ad-hoc one is not.

• Do the difficult thinking first, then use AI to check it. Never the reverse on anything that matters.

• Keep a light record of what you used AI for on significant work. It costs nothing until the day it costs everything.

• If a sentence in your draft contains a fact you did not know before you started, it needs a source you have personally opened.

If you run a team, a firm, or a classroom:

• Publish a policy, even a short one. Silence is not neutrality — it is an unwritten policy everyone interprets differently and nobody can be held to.

• Provide a sanctioned tool with reviewed data terms. Bans without alternatives push usage onto personal accounts, where you lose the last of your visibility. Note that training exclusion on business tiers is contractual; on a reimbursed personal subscription it is a toggle you are trusting a colleague to have found.

• Make verification a named, owned step. Deloitte had review. What it lacked was anyone whose actual job was to open the footnotes.

• Train on limitations, not just prompting. Staff who understand why models fabricate check differently from staff taught only to prompt well.

• Sample AI-assisted work in your existing quality assurance. If nobody ever tests the output, you do not have a control — you have a hope.

• Never let AI-drafted material go out under a signature without the signatory having read it. That is the whole lesson, and the alternative is learning it in public.

The technology is not the hazard. The hazard is the moment a fluent paragraph feels like a verified one — and Singapore's institutions have now put a price on the difference.