Skip to main content
Keeping Humans in Control of AI: A Practical Playbook for Families, Teams, and Small Businesses
Home / Blog / Human Intelligence & AI, The AI-Augmented Mind
Human Intelligence & AI, The AI-Augmented Mind

Keeping Humans in Control of AI: A Practical Playbook for Families, Teams, and Small Businesses

Control is a practice, not a petition.

In 2026, two campaigns asked governments to keep humans in charge of AI. One, Team Human, is a creator-led call for an international slowdown. The other, the Pro-Human AI Declaration, sets out 33 principles. Whether or not you sign either, a more immediate question sits closer to home: who is in control of the AI that you, your family, and your team use today?

This is the third article in a series. The first read both campaigns and fact-checked their headline claims. The second turned the Declaration’s principles into checks for a small business. This one is the playbook: five habits of control, seven practical routines, a family agreement, a one-page charter for teams, and a 30-day plan. It also looks at how to take part in the wider debate without being steered by it.

We are not asking you to adopt a position on any campaign or bill. Everything below is useful whichever way that debate ends, because it is about the part of AI you can actually govern: your own use of it.

Key Takeaways

  • Control comes down to five verbs: know, decide, verify, label, and stop.
  • If nobody owns an AI tool, nobody is in control of it. Naming an owner for each tool is the cheapest, highest-value step.
  • An off-switch you have never tested is a hope, not a control. Practise it.
  • At home, the biggest levers are attempt-first learning, protecting personal data, and keeping human relationships central.
  • You can join the public debate without signing anything. Read primary sources, check disclosures, compare polls, and use public consultations.

What Does It Mean to Keep Humans in Control of AI?

Keeping humans in control of AI means that people decide what AI may do, can see and check what it did, tell others when it is involved, and can stop it quickly. In practice that is five habits: know, decide, verify, label, and stop. The same five habits apply to a family choosing a homework tool, a shop owner using an AI assistant, and a company running an agent that can send payments. Only the stakes change.

The Control Ladder: Five Rungs

Rung The question At home At work
Know What AI is in use? List the apps, devices, and games that use AI Inventory every tool, including free ones staff use on their own
Decide Who decides what it may do? Agree family rules for what AI may be used for Write down which decisions need a person
Verify How do we check its work? Check facts in a second source Review outputs; keep logs; test before launch
Label Do people know AI is involved? Say when AI helped with schoolwork or content Disclose AI in chat, voice, email, and images
Stop Can we turn it off? Know how to disable features and delete accounts Test the off-switch and revoke access quickly

Rung 1: Know

You cannot control what you cannot see. Most households and many businesses underestimate how much AI they already use. It sits inside search engines, phone keyboards, photo apps, customer support widgets, meeting tools, and browser extensions. At work, staff often add tools on their own, which is the pattern we described in our Shadow AI article.

The routine is simple. Once a quarter, write a list: the tool, who uses it, what data it touches, and who owns it. At home, ask each family member to show you the AI features they use most. Do not treat it as an audit of behaviour. Treat it as a map.

Rung 2: Decide

Deciding means setting the limits before the tool is in daily use. Which decisions may an AI make alone? Which need a person to approve? Our guide to where to put the human in the loop offers four useful triggers: actions that are high-stakes, actions that cannot be undone, behaviour that is unusual, and limits the affected person has set. At home, the equivalent is a family conversation: AI may help with brainstorming and explanations, but a person decides what to submit, send, or believe.

Rung 3: Verify

AI tools can state wrong facts and invent references in a confident tone. The habit of checking is the same one we recommend for children in Raising Children for the AI Era: ask for the source, open a second tab, and look for who is behind a claim. At work, verification extends to systems: log what an AI did and test it before changes go live, as in our pre-launch checklist.

Rung 4: Label

People are entitled to know when they are dealing with a machine. The Declaration includes principles on labelling AI content and on AI honestly identifying itself. You do not need to wait for a law to do the same. A short line works: “This reply was drafted with AI and checked by a person.” At home, the equivalent is a norm that children and adults say when AI helped with homework, a message, or a picture.

Rung 5: Stop

Stopping is the control that makes the others credible. For a device or app at home, it means knowing how to switch off a feature, delete an account, and remove data. At work, it means that a named person can pause a workflow, revoke an agent’s credentials, and disable a tool without waiting for a vendor. Our vendor checklist lists what to ask. The key word is practised. A stop procedure that has never been run may not work when it is needed.

Seven Routines for Staying in Control

1. The Call-Back Rule

Deepfakes make voices and faces unreliable as proof of identity. In our article on the global AI safety debate, we described Singapore cases in which a finance director authorised US$499,000 after a video call with fake executives, and a businessman transferred US$3.8 million after a fabricated briefing. The routine that defeats this kind of fraud is boring and effective: any request involving money or sensitive information, however senior or urgent the requester appears, is confirmed through a second, independent channel, such as a call to a number you already hold. Say it out loud to the whole team, and to your family: no payment on a video call alone.

2. Name the Owner

Every AI tool and every AI-made output has a human owner. For a tool, that person is accountable for its settings, access, and results. For an output, such as a quotation or a report, it is the person who sends it. Our piece on agent identity explains how to make this traceable. The principle is easy to remember: if you cannot name who owns it, nobody does.

3. Attempt First

Try the task yourself before asking AI, then use AI to check and improve. In a field experiment published in PNAS in 2025, nearly 1,000 students who used a standard chatbot did better on practice problems but worse than other students once the tool was removed, while a tutor designed to protect learning largely avoided that harm. That is one school and one subject, so treat it as evidence of a mechanism, not a law. The habit transfers to adults: draft first, then compare. It also protects against the quieter problem of skills fading when practice disappears, which we covered in Will Your Staff Forget How to Do Their Jobs?

4. Say When AI Is Involved

Make disclosure a default, not a confession. In email signatures, support widgets, voice greetings, image captions, and school submissions, a short note builds trust and removes the risk of being found out. A good test: if the other person would feel misled on learning an AI was involved, tell them first.

5. Feed It Less

Control also means controlling what goes in. Do not paste customer personal data, contracts, source code, or children’s details into tools you have not approved. Agree a short list of data that never goes into an external AI tool, and distribute it. Our Shadow AI article shows why a written rule beats good intentions.

6. Run the Off-Switch Drill

Once a quarter, practise stopping. Pick one AI tool that acts, such as an agent, an automated email, or a voice assistant, and time how long it takes to pause it, revoke its access, and confirm that it has stopped. Write down the result. If it takes more than a few minutes, or requires a vendor, you have found a gap while the stakes are low.

7. Keep One Skill Human

For each person or team, choose one important skill to keep practising without AI: writing a first draft, reconciling an account, diagnosing a fault, doing mental arithmetic. The Declaration lists avoiding enfeeblement as a principle. The routine is the practical version. The goal is not to reject the tool. It is to keep the ability to check it.

At Home: Children, Companions, and Conversations

Families face a distinct version of the control question, because children meet AI inside games, homework tools, voice assistants, and chat features, often with little explanation.

What the Campaigns Say, and What Polls Show

The Declaration includes principles on child protection, on the right to grow, and against AI that creates emotional attachment or compulsive use. In the survey published alongside it, run by the Future of Life Institute among 1,004 likely US voters in February 2026, 77 percent agreed that companies must not exploit children through AI interactions that create emotional attachment or leverage, and 76 percent agreed that AI companies should not stunt children’s physical, mental, or social growth. These are US voter opinions on statements written by an advocacy group, so they show concern, not settled evidence about harm.

A Family AI Agreement

Whatever the policy debate, a family can set its own rules. A one-page agreement could say:

  1. AI is a tool and not a person. It does not have feelings, and it is not a friend.
  2. We try first, then ask AI to check or coach, not to do the work.
  3. We check facts in a second source before we trust them or share them.
  4. We do not put names, schools, addresses, photos, or other personal details into AI tools.
  5. We say when AI helped with schoolwork, messages, or pictures.
  6. We use only tools that fit the age rules and that a parent has looked at.
  7. Important feelings and problems go to people first.
  8. Anyone can say stop, and we review the rules every few weeks.

Warning Signs and Where to Turn

Watch for changes that suggest a tool is taking too large a place in a child’s life: withdrawing from friends and family, secrecy about chats, distress when a tool is unavailable, or treating an AI as a confidant for serious worries. If you see persistent distress, talk to the school counsellor or a doctor. A family agreement is a start, not a substitute for professional support.

At Work: A One-Page AI Use Charter

A charter turns the five rungs into something staff can read in two minutes. Adapt this template.

  1. Approved tools. We use only the AI tools on the approved list. New tools are requested from [owner].
  2. Owners. Every approved tool has a named owner who is accountable for its settings, access, and results.
  3. Data. We never put customer personal data, contracts, source code, or financial records into unapproved tools.
  4. Decisions. AI may draft, suggest, and summarise. A person approves anything involving money, contracts, hiring, discipline, health, or legal advice.
  5. Checking. We verify facts, figures, and sources in AI output before we use them.
  6. Labelling. We tell customers when they are talking to an AI, and we do not give bots human identities.
  7. Payments. We confirm any payment or sensitive request through a second, independent channel.
  8. Stopping. [Named person] can pause any AI that acts, and we test this every quarter.
  9. Skills. We keep practising core skills without AI so that we can check it.
  10. Incidents. If AI causes or nearly causes harm, we log it and review what to change.

A Monthly Twenty-Minute Review

Control improves when it is rehearsed. Once a month, spend twenty minutes with the team on three questions. What did AI get wrong this month? What did it save us? What would we not want it to do next month? Record the answers. After three months you will have a record of how AI is behaving in your own business, which is more useful than any general claim about it.

When Something Goes Wrong: The First Hour

Even a well-controlled setup will sometimes fail. An assistant gives a customer a wrong price, an agent sends an email it should not have, a staff member pastes the wrong file into a chatbot, or a convincing voice asks for a transfer. What you do in the first hour decides whether it stays a small incident. A short script, agreed in advance, helps people act calmly.

  1. Stop it. Use the off-switch you practised: pause the workflow, revoke the credentials, disable the tool. Do this before you investigate.
  2. Keep the evidence. Save the logs, the messages, the prompts, and the outputs. Do not delete anything in a rush to tidy up.
  3. Tell the owner and one decision-maker. The named owner of the tool and a person with authority to decide should both know within minutes. Avoid long threads with many people.
  4. Contain the harm. If money moved, call your bank immediately on a number you already hold. If personal data left the building, find out which data and whose. If a customer received wrong information, contact them first with a short, plain correction.
  5. Check your obligations. Some incidents, such as a personal data breach, carry notification duties under Singapore law. Check the current PDPC guidance or ask an adviser quickly, because timing can matter. This is general information and not legal advice.
  6. Write it down. Record what happened, when, which system, who owns it, the effect, and what was done. A few lines are enough.
  7. Change something. Within a week, decide what will be different: a new checkpoint, a narrower permission, a revised rule, or a retired tool. Then add the lesson to the monthly review.

The value of this routine is that it removes the question of what to do when everyone is stressed. It also connects to the Declaration’s principle that when AI causes harm, it should be possible to find out why and who is responsible. That is hard to achieve unless logs and owners exist before the incident, which is the point of the agent identity practices we describe elsewhere, and of knowing the five ways an agent can hurt your business in advance.

Three Short Scenes

The five rungs are easier to remember when you see them at work. These scenes are illustrations, not case studies.

The Parent and the Homework Helper

A parent notices that a child’s maths scores on practice sheets have jumped, while test results have not. Rather than banning the tool, the parent asks to see how it is used. It turns out to give full solutions. Together they change the rule: attempt first, ask the tool for a hint, then compare. They also add a line to the family agreement that AI use is mentioned when work is handed in. Rungs used: know, decide, label.

The Shop Owner and the Chat Widget

A small retailer adds an AI chat assistant to answer delivery questions. Before launch the owner writes down ten risky situations, including an angry customer, a request to speak to a person, and a question about a refund the assistant is not allowed to approve. Two tests fail: the assistant promises refunds it cannot give, and it never says it is an AI. The owner fixes both, adds a greeting that says what it is, and sets a rule that refunds always go to a person. Rungs used: verify, label, decide.

The Finance Manager and the Urgent Call

A finance manager receives a video call from someone who looks and sounds like the managing director, asking for an urgent transfer to a new supplier. The manager points out that company policy requires a call-back on a known number for any new payee. The real director, reached on that number, knows nothing about the request, and the attempt ends there. Rungs used: decide, verify, stop.

Joining the Debate Without Being Steered

You may want a view on the campaigns we covered. Here is how to form one without being rushed.

  • Read the primary text. Both Team Human and the Declaration are short enough to read in full, and the Declaration’s principles are itemised. Do this before reading commentary.
  • Check who benefits and who pays. Look at funding, sponsorships, and disclosures. Team Human’s own FAQ states how creator sponsorship is handled.
  • Compare more than one poll. We showed that results change with wording and sample, and that most polling so far is of US voters.
  • State the best counter-argument. If you can describe the strongest objection fairly, you understand the question.
  • Sign what you have read. A signature is a personal choice. We are not recommending for or against any petition.
  • Use legitimate channels where you have a stake. Public consultations and calls for case studies are a way for businesses to be heard. For example, IMDA’s framework for agentic AI invites organisations to submit case studies for future versions, and we describe how an architecture can be mapped against it in One Governance Architecture, Three Autonomy Levels.

Is There Common Ground?

Reading across the sources in this series, supporters and critics of a slowdown tend to overlap on more than the headlines suggest. Both sides generally accept that high-stakes AI should be tested before it is deployed, that someone has to be accountable when it fails, that people should know when they are dealing with a machine, and that there should be a way to stop a system that is misbehaving. Where they differ is on speed, on international enforcement, and on how much to rely on voluntary measures versus law. The practices in this article sit in the area of agreement, which is part of why they are worth adopting now.

A 30-Day Starter Plan

  • Week 1: Know. Make the list of AI tools at home and at work. Add an owner to each work tool.
  • Week 2: Decide and label. Write the family agreement and the one-page charter. Add AI disclosures to customer-facing tools and drafts.
  • Week 3: Verify and protect. Introduce the call-back rule and the data list of what never goes into an unapproved tool. Run a fact-check on one AI answer together.
  • Week 4: Stop and review. Run an off-switch drill on one tool. Hold the first twenty-minute review. Decide which skill each person will keep practising.

Frequently Asked Questions

How do I keep control of the AI my business uses? Start with an inventory and a named owner for each tool, decide which decisions need a person, verify outputs, tell people when AI is involved, and test that you can switch it off.

What is the simplest protection against AI deepfake fraud? Confirm any payment or sensitive request through a second, independent channel, such as calling a number you already hold, regardless of how real a video or voice seems.

Should children use AI tools? That depends on the child, the tool, and its age rules. A sensible approach is supervised use with clear rules: attempt first, check facts, protect personal data, and keep human relationships central. If a child shows persistent distress, speak to a counsellor or doctor.

Do I have to sign Team Human or the Declaration to act on this? No. Everything here is useful whatever you think of either campaign, and signing is a personal decision.

Is a human in the loop enough? Not by itself. A person who cannot understand the output, who is overloaded, or who approves everything on reflex is a weak control. Train approvers, limit the number of approvals, and audit how well oversight works.

Where can a Singapore business start? With the IMDA frameworks and the practical guides in this blog, and with the one-page charter above. Check current law, sector rules, and your professional body’s guidance.

Continue the Series

Sources


Willie is the Managing Director of Mxgsoft Pte Ltd, a Singapore-based digital transformation company specialising in ERP implementations, workflow automation, and AI-powered business solutions.

Next step: Do the five-minute version tonight. Write down every AI tool you or your family use, put a name next to each, and agree one rule you will keep: no payment on a video call alone.