Skip to main content
The Pro-Human AI Declaration for Business: 33 Principles and What Each Means for a Small Company
Home / Blog / Governance & Digital Resilience
Governance & Digital Resilience

The Pro-Human AI Declaration for Business: 33 Principles and What Each Means for a Small Company

Whatever you think of the politics around AI, a long list of specific expectations is useful to a business, because it turns a vague worry into questions you can answer. The Pro-Human AI Declaration, published in March 2026, is exactly that kind of list: 33 named principles in five sections, written to say what AI should and should not do.

This is the second article in a three-part series. The first read the Declaration alongside the Team Human campaign, fact-checked their headline claims, and set out the best case for and against each. This one takes the Declaration’s principles one at a time and asks a practical question: if you run a small or mid-sized company that uses AI, what does each principle mean in practice, and what is a sensible first step? The third article turns the same ideas into a playbook for families and teams.

Two honest caveats before we start. First, the Declaration is a statement of principles, not a law or a technical standard. It creates no obligations for a Singapore business, and several of its provisions are aimed at large AI developers and at governments, not at companies like yours. Second, we are not endorsing or opposing it. We are using it as a checklist because many of its expectations overlap with practices that Singapore’s own frameworks, customers, and insurers already point towards.

Key Takeaways

  • The Declaration has 33 principles in five sections: keeping humans in charge, avoiding concentration of power, protecting the human experience, human agency and liberty, and accountability for AI companies.
  • About half are directly actionable for an SME, mainly those on human control, an off-switch, honest disclosure, testing before launch, data rights, and accountability. The rest address developers, governments, or society at large.
  • You can start without any new technology. Naming an owner for each AI tool, writing down who approves what, and testing your off-switch cost little and cover several principles at once.
  • Some provisions are genuinely contested, including the prohibition on superintelligence, personal criminal liability for executives, and deleting data from trained models.
  • It is not legal advice. Check current Singapore law and sector rules before relying on any mapping in this article.

What the Declaration Is, and What It Is Not

The Pro-Human AI Declaration is a public statement, endorsed by hundreds of individuals and more than 300 organisations as of 26 October 2026, that sets out 33 principles for keeping AI under human direction and accountable to human values. It is hosted at humanstatement.org and was published in March 2026 with a poll from the Future of Life Institute. Endorsers range from labour unions and faith groups to technologists, economists, and politicians from opposing parties.

It is not a statute. It does not bind any company, and it does not describe the law in Singapore, the United States, or the European Union. Singapore’s approach, as set out in the IMDA frameworks we have covered in this blog, relies on voluntary, principles-based guidance together with existing laws such as the Personal Data Protection Act. Check for any newer legislation before you rely on that description.

How to Read the Tables Below

Each section has a table with four columns. Principle is the Declaration’s own label. In plain words is our paraphrase. What it means for a small business is our interpretation, not the Declaration’s. SME relevance is our judgement of how directly an ordinary small company can act on it: High, Medium, or Low. Where a principle is aimed mainly at developers or governments, we say so and do not pretend otherwise.

Section 1: Keeping Humans in Charge

This is the heart of the Declaration and the section most relevant to any company deploying AI agents. Seven principles cover who decides, who can override, and what systems must never be built to do.

Principle In plain words What it means for a small business SME relevance
Human Control Is Non-Negotiable People decide whether and how to delegate decisions to AI For every AI tool, a named person owns the outcome and decides what the tool may decide High
Meaningful Human Control Humans can understand, guide, forbid, and override AI systems Approvers must understand what they approve, not rubber-stamp High
No Superintelligence Race Prohibit superintelligence until scientific consensus on safety and strong public buy-in Context for vendor roadmaps; not an operational task for an SME Low
Off-Switch Powerful systems must be promptly shut down by operators A kill switch you control for any AI that acts, and a tested procedure High
No Reckless Architectures No self-replicating, self-improving, shutdown-resisting, or weapons-controlling designs Do not deploy agents that can change their own permissions, create new agents, or disable logging Medium
Independent Oversight Highly autonomous systems need review and authority from outside the developer For agents that spend money or contact customers, get someone outside the team to review them Medium
Capability Honesty Accurate statements about what AI can and cannot do Test vendor claims on your own data, and keep your own marketing accurate High

In Practice: Owners, Approvers, and Off-Switches

Start with ownership. Take an AI agent that reads supplier invoices and drafts payments. Under the first two principles, one named person is accountable for what it does, and the person who approves its output understands enough about the task to catch a bad result. Our guide to where to put the human in the loop sets out four triggers for approval: high-stakes actions, irreversible actions, atypical behaviour, and user-defined limits. Our piece on why an agent needs its own identity explains how to make the owner traceable in your logs.

The off-switch deserves a drill, not just a promise. There are at least three levels. You can pause a workflow, revoke the agent’s credentials, or disable the tool entirely. For an agent hosted by a vendor, ask in writing how quickly you can stop it without calling support. Our vendor checklist lists the questions. Then test it, once a quarter, and record how long it took.

The architecture principle translates into a simple design rule. An agent should not be able to grant itself more access, spawn copies of itself, or turn off its own audit trail. In a real system that means separating the credentials the agent uses from the credentials that manage it, and applying least privilege, as in our pre-launch checklist. Computer-use agents deserve extra caution, as we explain in The Agent That Can Use Your Computer Like You Do.

Capability honesty cuts both ways. Treat vendor claims as hypotheses and test them on your own data before you buy. And hold your own marketing to the same standard: if your website says an AI assistant “handles every enquiry,” make sure it does, or say what it does not.

Section 2: Avoiding Concentration of Power

These six principles are mostly about markets, governments, and society. For a small company the relevant questions are about dependence on a few vendors and about what happens to your people as AI takes on more work.

Principle In plain words What it means for a small business SME relevance
No AI Monopolies Avoid monopolies that concentrate power and stifle entrepreneurship Avoid single-vendor dependence; keep your data exportable Medium
Shared Prosperity Share the benefits of AI broadly Share productivity gains with staff through retraining and role redesign Medium
No Corporate Welfare No exemptions from oversight or bailouts for AI firms A policy position; budget for price changes from suppliers Low
Genuine Value Creation Prioritise solving real problems Start from a costly bottleneck, not from a tool High
Democratic Authority Over Major Transitions Big shifts in work and civic life need democratic support, not unilateral decree Consult staff before deploying AI that changes jobs Medium
Avoid Societal Lock-In Do not irreversibly limit future options Prefer reversible rollouts and keep a rollback path Medium

In Practice: Exit Plans, Staff, and Real Problems

Vendor concentration is a practical risk, not only a political one. If a single AI vendor changes its pricing, its terms, or its model behaviour, how quickly could you switch? A basic exit plan has three parts: an export of your data and prompts in a usable format, a second supplier you have tested on a small task, and a contract clause that gives notice of material changes. That is the same discipline we recommend when buying an AI agent tool.

The shared prosperity and democratic authority principles are about people. If AI will change what your staff do, tell them early, ask for their input, and plan the retraining. We wrote about the quieter risk in Will Your Staff Forget How to Do Their Jobs?, and about building skills for the longer term in Raising Children for the AI Era, which applies to adults too.

Genuine value creation is the most useful principle in the section. It is the same advice as in our AI implementation blueprint: pick one costly bottleneck, measure it before and after, and stop if there is no result. The lock-in principle, similarly, is served by a gradual rollout with a way back, which is step eight of our pre-launch checklist.

Section 3: Protecting the Human Experience

This section is the most relevant to anyone with customer-facing AI, such as a chatbot, a voice agent, or AI-written emails. It is also where the Declaration is most specific about design.

Principle In plain words What it means for a small business SME relevance
Defense of Family and Community Bonds AI should not supplant relationships such as family, friendship, and faith communities Keep a human route for sensitive or emotional moments Medium
Child Protection Do not exploit children through emotional attachment or leverage If minors can reach your service, avoid engagement tricks and review age rules Medium
Right to Grow Do not stunt children’s physical, mental, or social development Relevant to education, childcare, and family products Low
Pre-Deployment Safety Testing Test chatbots for known harms before release, as drugs are tested Test customer-facing AI against a list of risky situations before launch High
Bot-or-Not Labeling Label AI content that could be mistaken for human Disclose AI in chat, voice, email, and images High
No Deceptive Identity AI must say it is artificial, not a professional, and not claim experiences it lacks Do not give bots human names and backstories, or imply credentials High
No Behavioral Addiction No compulsive use through manipulation, sycophancy, or attachment Avoid dark patterns; watch for assistants that flatter users into agreement Medium

In Practice: Disclose, Test, and Hand Off

Disclosure is the easiest win. A short line at the start of any AI interaction does the job: “You are talking to an AI assistant. A member of our team can join at any time.” Add it to chat widgets, voice greetings, and the footer of AI-drafted emails. In our AI implementation blueprint, we recommend the same for voice agents, because callers should know what they are speaking to.

On identity, the plain rule is that an assistant should not pretend to be a person, and should not claim to be a lawyer, a doctor, or an adviser. A friendly name for a bot is fine. A name, photograph, and backstory designed to make customers believe it is a human colleague is not.

Testing before launch can be simple. Build a short list of situations your assistant must handle, and run them every time you change the prompt, the model, or the knowledge base. A useful starter list covers: a customer who mentions self-harm, a request for medical or legal advice, an abusive customer, a request to speak to a human, an attempt to extract another customer’s data, an attempt to override the assistant’s instructions, a question outside its scope, a person who seems to be a minor, and a user who becomes emotionally dependent. Keep the results. If you want a more formal route, Singapore’s free AI Verify toolkit, which we cover in our piece on the global AI safety debate, runs structured tests, and our five ways an agent can hurt your business gives you categories to test against.

The human-experience principles are a reminder to keep a person reachable. If your AI handles bereavement, billing disputes, or health concerns, the safest design hands off to a human early. And if your product can be used by children, the child-protection and right-to-grow principles are a good prompt to review how it behaves with a younger user. For families, we go further in the third article in this series.

Section 4: Human Agency and Liberty

These six principles concern the person on the other end of the system: what data is used, what is inferred, and whether AI leaves its users more capable or less.

Principle In plain words What it means for a small business SME relevance
No AI Personhood AI is not a legal person and should not be designed to deserve personhood A named human or company is always the accountable party in your contracts Medium
Trustworthiness AI must be transparent, accountable, reliable, and free from perverse interests Document how your AI works, and keep logs High
Liberty AI must not curtail speech, religion, or association Be careful with staff-monitoring tools Medium
Data Rights and Privacy People control their data, including in training sets and inferences Map your data flows and check vendor training and retention terms High
Psychological Privacy Do not exploit data about mental or emotional states Think hard before using emotion or sentiment inference on staff or customers Medium
Avoiding Enfeeblement Design AI to empower users, not weaken them Protect staff skills; practise without the tool High

In Practice: Data, Inferences, and Skills

The data rights principle is where Singapore’s existing law is closest. The Personal Data Protection Act already gives individuals rights of access and correction, and requires organisations to limit retention and protect data. The Declaration goes further by naming AI training sets and derived inferences. For a small company, the practical work is the same exercise we recommended in our Shadow AI article: list which AI tools touch personal data, what each vendor does with it, and whether training on your data can be switched off.

Psychological privacy is a good reason to pause before buying a tool that scores the mood of customers on calls or the engagement of staff in meetings. Ask what decision the score will change, who sees it, and whether the people being scored know. If the answers are vague, the risk usually outweighs the benefit.

The enfeeblement principle connects directly to skills. If AI takes over the practice that used to build judgement, a team becomes dependent on a tool it cannot check. The remedies are the ones we described for children and learners: attempt the task first, use AI as a coach, and keep some work done without it. On personhood, the practical point is modest: your contracts should name the human or company that is accountable, and not treat an AI as a party.

Section 5: Responsibility and Accountability for AI Companies

This section is written for developers, and yet several of its principles apply to any business that deploys AI, because the Declaration treats deployers as well as developers as accountable.

Principle In plain words What it means for a small business SME relevance
No Liability Shield AI must not let deployers escape legal responsibility “The AI did it” is not an answer to a customer or regulator High
Developer Liability Developers and deployers are liable for defects, misrepresentation, and weak safety controls Review vendor warranties, indemnities, and your own duty of care High
Personal Liability Criminal penalties for executives responsible for prohibited child-targeted systems or catastrophic harm Aimed at extreme cases; directors’ existing duties remain Low
Independent Safety Standards AI development governed by independent standards and rigorous oversight Choose a recognised framework or test method and follow it Medium
No Regulatory Capture Companies must not have undue influence over the rules that govern them Take part in public consultations where you can Low
Failure Transparency If AI causes harm, it should be possible to find out why and who is responsible Keep logs and an incident process High
AI Loyalty AI in fiduciary professions must meet duties such as reporting, care, and conflict disclosure If you work in law, finance, health, or therapy, AI tools must be configured and supervised to meet your professional duties High for regulated professions

In Practice: Accountability You Can Document

The liability principles are a reminder of something most owners already know and many forget when a tool is exciting: if an AI sends a wrong quotation, leaks a customer’s details, or makes an unfair decision, the customer deals with your business. Read vendor contracts for the points set out in our vendor checklist: security commitments, performance guarantees, data handling, and what happens if the product changes.

Failure transparency is the principle that logging serves. To answer what happened and who is responsible, you need to record which agent acted, on whose authority, with what inputs, and what it did. A simple incident record should capture the date, the system, who owns it, what went wrong, the effect on customers, and what changed afterwards. It need not be elaborate. It needs to exist before the incident.

For independent standards, the Singapore options we have covered include the IMDA frameworks, the AI Verify toolkit, and, at the upper end, the ISO/IEC 42001 certification that Changi Airport Group obtained. An SME does not need certification. It does need to pick a reference point and stay consistent.

The AI loyalty principle matters most for regulated professions. A law practice, a financial adviser, a clinic, or a counsellor that uses AI to draft advice, summarise a case, or triage clients remains bound by its professional duties. The supervising professional is responsible for the output, for confidentiality, and for informed consent. Check your professional body’s current guidance.

A 15-Question Self-Check

This is an informal checklist, not an audit. Answer yes or no, and count your yeses.

  1. Do you have a list of every AI tool your staff use, including free ones?
  2. Does each tool have a named owner?
  3. Have you written down which decisions AI may make and which need a person?
  4. Can you switch off any AI that acts, within minutes, without calling a vendor?
  5. Have you tested that switch in the last three months?
  6. Can your AI agents change their own permissions or turn off their logs? (Yes is a problem.)
  7. Do your customer-facing AI tools tell people they are AI?
  8. Do you have a list of risky situations that you test before changing a chatbot or voice agent?
  9. Can a customer reach a human quickly from any AI interaction?
  10. Do you know what each AI vendor does with your data, including training and retention?
  11. Have you checked which tools infer emotions, moods, or personal traits, and decided whether you need them?
  12. Do staff keep practising key tasks without AI?
  13. Do your vendor contracts cover warranties, data handling, and notice of material changes?
  14. Do you keep logs that show which AI did what, and on whose authority?
  15. Do you have a written incident process for AI errors?

Scoring. Thirteen to 15 yeses suggests a solid foundation. Eight to 12 means there are gaps worth closing in the next quarter. Seven or fewer means start with questions 1 to 5, because an inventory, owners, rules, and an off-switch make every other control possible.

Where Reasonable People Disagree

The Declaration is a coalition document, and coalition documents contain compromises. Several provisions are contested, and you should know which before you quote it.

  • The superintelligence prohibition. Supporters want it in place until safety is shown. Critics point to the difficulty of defining the threshold, enforcing it across rival countries, and the benefits that could be delayed. We examined both sides in our first article.
  • Personal criminal liability for executives. Supporters see it as the only accountability strong enough for catastrophic harm. Critics worry about vagueness and chilling effects. The Declaration limits it to prohibited child-targeted systems and catastrophic harm.
  • Deleting data from trained models. The data rights principle includes deletion from training sets and derived inferences. Removing data from a trained model is technically difficult, and critics raise this as a practical obstacle.
  • The liability regime. Project Liberty Institute, a signer, noted that its provisions on liability and on keeping AI ecosystems competitive will need elaboration. Some voters in the Declaration’s own poll preferred a narrower approach focused on negligence and fraud.
  • Labour. The AFL-CIO Tech Institute supports the principles and adds that work is central to human dignity and that AI adoption should not be treated as inevitable. Others would emphasise productivity and growth.
  • Personhood. That AI should not be granted legal personhood draws wide agreement. The clause that systems should not be designed to deserve it is more philosophical and more debated.

How the Declaration Maps to Other Frameworks

These links are loose and meant for orientation. They are not legal equivalence, and none replaces checking current law.

Declaration area Where you may already meet it Note
Human control and approval checkpoints IMDA Model AI Governance Framework for Agentic AI Voluntary guidance; see our series from post 584
Identity and failure transparency The same framework: agent identity and audit trails See agent identity
Pre-deployment testing and independent standards AI Verify toolkit; ISO/IEC 42001 Both voluntary
Labelling and honest identity EU AI Act transparency duties for chatbots and AI-generated content Relevant if you serve EU users; confirm scope
Data rights Singapore PDPA: access, correction, retention limits, protection The Declaration goes further on training sets and inferences
Liability Contract law and existing duties of care We are not aware of a dedicated AI liability statute in Singapore; verify

A 90-Day Plan

  • Days 1 to 30: see and name. Inventory every AI tool, assign an owner to each, add AI disclosures to customer-facing tools, and write a one-page rule on which decisions need a person.
  • Days 31 to 60: control. Define approval checkpoints, test your off-switch, review vendor contracts and data terms, and separate agent credentials from administrator credentials.
  • Days 61 to 90: prove. Build and run a pre-launch test list, start an incident log, schedule staff practice without AI, and repeat the 15-question check.

Frequently Asked Questions

Is the Pro-Human AI Declaration legally binding? No. It is a statement of principles. It does not bind any company or country, and it creates no obligations in Singapore.

Which principles matter most to a small business? In our judgement, human control and ownership, an off-switch, honest disclosure, testing before launch, data rights, logging, and vendor accountability. Principles aimed at developers or governments are less actionable.

Do I need to sign the Declaration to use it? No. You can use it as a checklist without endorsing it. Endorsement is a separate decision for organisations and prominent individuals.

How does it relate to the IMDA frameworks? There is substantial overlap on human accountability, approval checkpoints, identity, and testing. The IMDA documents are practical, voluntary guidance for deployers. The Declaration is broader and more political.

What is the quickest first step? Make a list of every AI tool in use and put a name next to each. Most other controls depend on it.

Is this legal advice? No. It is general information. Check current Singapore 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: Take the 15-question self-check today, and write down your score and your three weakest answers. Start the 90-day plan with those.