![\"sovereign](\"https:\/\/himalini.com\/en\/wp-content\/uploads\/2026\/11\/Sovereign-AI_-Nepals-Digital-Key.png\" "\\"Nepal")
<\/p>\n<\/p>\n
\u00a0Suman Tiwari, 4 November<\/strong>.<\/span><\/i><\/p>\n As Nepal prepares for its third National Summit on Artificial Intelligence on 7 November, the debate should move beyond whether the country needs more computing power. A harder question deserves attention: before a government AI system affects people\u2019s lives, who checks it, who approves it and who has the authority to stop it?<\/span><\/p>\n Nepal\u2019s National Artificial Intelligence Policy 2082 sets out an ambition for responsible AI. The 2026\/27 budget announces a Sovereign AI Compute Center at Syuchatar, Kathmandu, and plans for more accessible computing capacity. These are important ambitions, not proof that the proposed centre is already operating. But a country can own servers and still lack control over the systems running on them. Nepal needs a practical AI Security Gate: a risk-based checkpoint before high-impact public-sector AI is deployed.<\/span><\/p>\n A data centre in Kathmandu can still depend on foreign chips, software, cloud services, AI models and suppliers. Such dependencies are not automatically a problem. The problem begins when a public institution cannot inspect them, limit a supplier\u2019s access, investigate an incident or continue essential services if a contract or connection fails. Sovereignty is about control; cybersecurity is about protecting that control.<\/span><\/p>\n The keys to sovereign AI are therefore more than passwords. Who can view confidential records? Who may change a model or its instructions? Who can connect it to another government database? Who approves an upgrade, and who can suspend the system when something goes wrong? Ownership matters, but operational authority matters just as much. A Nepali address cannot answer those questions. Owning the machine is not the same as holding the keys.<\/span><\/p>\n AI can make mistakes faster than people can correct them. It can multiply mistakes as efficiently as it multiplies productivity. A flawed rule or unreliable dataset might cause limited harm in one manual decision; built into an automated public service, the same mistake could affect hundreds or thousands of people before anyone notices. Not every harmful outcome is a cyberattack. Sometimes the system is wrong. Sometimes its information is poor. Sometimes someone has manipulated it. Officials need the evidence and expertise to tell the difference, and affected citizens need a meaningful route to human review.<\/span><\/p>\n The stakes rise when AI can act, not merely answer. An assistant connected to government systems might read sensitive records, update information or trigger another service. Its permissions should match its job. A system that only needs to read a record should not be able to rewrite it; one that drafts a recommendation should not automatically execute the decision. The ability to monitor and halt those actions is not an optional extra. It is part of keeping government in control.<\/span><\/p>\n Kathmandu already has cameras capable of recognising number plates and faces, which the city says support traffic monitoring and investigations. Their existence is not evidence of wrongful automated penalties. It does, however, illustrate why public authorities should think about accountability before technologies become connected to consequential decisions.<\/span><\/p>\n The weakness in Nepal\u2019s AI ambitions is not a lack of principles. It is the distance between principle and deployment. A commitment to ethical and accountable AI does not, by itself, say what evidence a ministry must produce before launch, who can challenge the evidence or who has the final authority to refuse an unsafe system.<\/span><\/p>\n An AI Security Gate could help close that gap. It should be triggered by potential harm, not by the mere presence of an AI label. Systems handling sensitive personal data, changing official records, influencing access to essential services or taking actions in other public systems deserve stronger scrutiny than an internal writing assistant or a research prototype using public information.<\/span><\/p>\n Before deployment, the responsible agency should document what the system is meant to do, the information it uses, its outside suppliers and the permissions it requires. Testing should examine ordinary cybersecurity weaknesses alongside AI-specific risks, such as manipulated inputs, disclosure of confidential data and unauthorised actions through connected tools. The assessment must be proportionate: this is not a demand for an expensive penetration test of every experiment.<\/span><\/p>\n An assessor with relevant cybersecurity and AI-system expertise, and sufficient independence from the supplier, should report the findings. But the assessor should not become the decision-maker. A named official in the deploying institution must accept any remaining risk or refuse deployment. Government can outsource technology and testing; it cannot outsource accountability.<\/span><\/p>\n Imagine a public agency considering an AI tool to help screen applications for a benefit. Its recommendation could influence whether a family receives essential support. During assessment, it emerges that the outside supplier can quietly alter eligibility rules without the agency\u2019s approval. A generic statement that the software is \u201csecure\u201d would be meaningless. The Security Gate should require that access to be restricted, changes to be independently checked and the approval process documented. If the agency cannot demonstrate control, deployment should wait.<\/span><\/p>\n That is the distinction between a checklist and a decision gate. The point is not to promise a flawless model. It is to require enough evidence for someone to make an accountable decision\u2014and enough authority to intervene when the evidence is inadequate.<\/span><\/p>\n Nepal does not need to wait for a comprehensive AI law before improving the way government buys and deploys technology. Contracts for high-impact systems can specify audit rights, limits on supplier access, notification of significant model or configuration changes, incident-reporting obligations, secure data return and a credible exit plan. They can also establish the government\u2019s right to restrict or suspend a service when essential safeguards fail.<\/span><\/p>\n These are not merely legal details. If a ministry cannot examine how a supplier\u2019s system was changed, or cannot disconnect it without losing control of its own records, its sovereignty is fragile no matter where the servers are located.<\/span><\/p>\n Nepal is not starting from zero. The National AI Policy provides for an AI Regulation Council, and the National AI Centre has now been established. The National Cyber Security Policy 2080 and National Cyber Security Centre provide a wider cybersecurity foundation. Yet establishing a centre does not by itself show that it has legal authority to approve, audit or halt a government AI deployment.<\/span><\/p>\n The deploying agency should remain responsible for its system; qualified independent experts can assess it; and national institutions can coordinate guidance within mandates that are publicly defined. The important task is to agree who sets minimum standards, who verifies compliance and who has the power to act. Creating another committee without answering those questions would simply move the uncertainty elsewhere.<\/span><\/p>\n Passing a Security Gate once should not mean being trusted indefinitely. A new dataset, model upgrade, expanded supplier privilege or connection to another public system can change the risk. Significant changes should trigger reassessment, supported by monitoring, incident plans and protected records of important actions. Oversight should continue through eventual retirement, including secure handling of data and access when a system is withdrawn. Audit logs can help show what a system accessed or changed, although they cannot always explain why a model reached a particular conclusion.<\/span><\/p>\n People matter too. Officials, investigators and security teams must be able to recognise abnormal behaviour and exercise the authority to restrict permissions, disable risky connections or suspend a system. Citizens affected by important AI-assisted decisions should know which public body is responsible and how to ask a person to review an outcome. Technical security and administrative fairness are different obligations; public trust requires both.<\/span><\/p>\n Nepal should measure progress in sovereign AI not only by computing units acquired or applications launched, but by whether its institutions can answer basic questions: who approved the system, who can alter it, what independent evidence supports its safety and who must respond when it fails?<\/span><\/p>\n Genuine sovereignty means having the capacity to understand, scrutinise, secure and, when necessary, stop the technology we build. That is what it should mean to hold the keys.<\/span><\/p>\n Suman Tiwari is an Ireland-based cybersecurity professional who writes on cybersecurity, emerging technology and public policy.<\/span><\/p>\n","protected":false},"excerpt":{"rendered":" \u00a0Suman Tiwari, 4 November. As Nepal prepares for its third National Summit on Artificial Intelligence on 7 November, the debate should move beyond whether the country needs more computing power. A harder question deserves attention: before a government AI system affects people\u2019s lives, who checks it, who approves it and who has the authority to […]<\/p>\n","protected":false},"author":3,"featured_media":6809,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_exactmetrics_skip_tracking":false,"_exactmetrics_sitenote_active":false,"_exactmetrics_sitenote_note":"","_exactmetrics_sitenote_category":0,"footnotes":""},"categories":[13,16,3],"tags":[],"class_list":["post-6808","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-nepal","category-opinion","category-top-stories-of-madhesh"],"_links":{"self":[{"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/posts\/6808","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/comments?post=6808"}],"version-history":[{"count":1,"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/posts\/6808\/revisions"}],"predecessor-version":[{"id":6810,"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/posts\/6808\/revisions\/6810"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/media\/6809"}],"wp:attachment":[{"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/media?parent=6808"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/categories?post=6808"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/himalini.com\/en\/wp-json\/wp\/v2\/tags?post=6808"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}Sovereignty is more than server location<\/b><\/h1>\n
When AI mistakes multiply<\/b><\/h1>\n
From policy promises to a security gate<\/b><\/h1>\n
A test that matters to citizens<\/b><\/h1>\n
Make procurement part of sovereignty<\/b><\/h1>\n
Use institutions, but define their powers<\/b><\/h1>\n
Approval is not forever<\/b><\/h1>\n
Who really holds the keys?<\/b><\/h1>\n