AI in Government Organizations: Is It Allowed and Is It Safe?

Yes, government organizations can and should use artificial intelligence — but not "as is", and only under clear rules. Safety depends less on the model itself and more on what data goes into it, where it runs and who is accountable for the output. Public and anonymised material can be processed in public services; personal data and internal information require enterprise modes, local models and a clear internal policy.

Below is a calm breakdown without hype or scare stories: where AI helps public agencies, which risks are real, what Uzbek law says and how to build a safe operating model.

Why the question matters right now

In just a couple of years, AI chatbots have gone from experimental toys to everyday work tools. Staff at ministries, hokimiyats, state banks and state-owned companies already use ChatGPT, Gemini, Claude and other services — often from personal accounts and without any rules. That is the real risk: not the technology itself, but unmanaged use.

At the same time, the state is setting a clear direction: in October 2024 a Presidential resolution approved the Strategy for the Development of Artificial Intelligence Technologies until 2030. We covered it in "Uzbekistan's AI Strategy 2030: what to know". So the question is no longer "whether to use AI" but "how to use it safely".

Where AI genuinely helps government organizations

AI delivers the most value where there is a lot of text, routine and standard documents, while a person still makes the final decision:

• Presentations and reports. Speech structure, slide drafts, talking points, summaries of open data.
• Working with text. Editing, shortening, translating between Uzbek, Russian and English, turning bureaucratic language into plain language for citizens.
• Citizen requests. Classifying requests by topic, drafting standard replies (always reviewed by an officer).
• Open-data analysis. Summarising public statistics, comparing international practice, preparing briefings.
• Communications. Posts for official channels, press releases, FAQ pages for the agency website.
• Internal knowledge bases. Searching regulations and normative documents with an AI assistant deployed inside the organization.

The general rule: AI is great at accelerating draft work and poor as a source of final decisions.

What it looks like in practice: three typical scenarios

To keep this concrete, here are generalised scenarios that occur in most agencies. They are not stories of specific organizations, just typical working situations.

Scenario 1. The press office prepares a speech for the head of the agency

The task: within two days, prepare a talk and a presentation on the year's results for a programme whose data has already been published. The safe approach: the officer gives the model only published figures and texts from the official website, and asks for a ten-minute speech structure, three key messages and slide title options. They then check every figure against the source by hand, and a designer builds the slides in the corporate style. The unsafe approach: pasting an internal memo with non-public indicators and the names of responsible staff into a chatbot.

Scenario 2. The citizen requests department

The task: speed up the handling of a large flow of similar requests. The safe approach: a model deployed inside the organization's perimeter classifies requests by topic and suggests a draft reply based on approved templates, which an officer checks and sends. The unsafe approach: forwarding request texts with names, addresses and phone numbers to a public foreign service to save time.

Scenario 3. A ministry's analytical department

The task: prepare a review of international practice for a sector reform. The safe approach: use AI to find directions and structure the review, then verify every reference, figure and document title against primary sources. The unsafe approach: copying "studies" and links suggested by the model into the report without opening a single one.

In all three cases, the difference between value and an incident lies not in the choice of model but in what data goes into it and who checks the result.

Real risks: what can go wrong

1. Data leakage

Everything an employee pastes into a public chatbot goes to the provider's servers, usually abroad. If that is applicants' passport data, internal correspondence or a draft resolution before publication, it is already an information security incident.

2. Training on your data

Many free and consumer plans allow conversations to be used to improve models by default. Business plans usually exclude this, but you need to verify it in the specific service's terms rather than assume it.

3. Cross-border transfer

Uzbekistan's Law "On Personal Data" (No. ZRU-547 of 2 July 2019) required, from 2021, that citizens' personal data processed over the internet be stored on servers in the country. Law No. ZRU-1125 of 26 March 2026 relaxed this: mandatory localisation now covers biometric and genetic data and data of subscribers of local telecom operators, while other data may be transferred abroad if conditions are met — for example, to a country on the government-approved list with an adequate level of protection, or under standard contractual clauses. For a government body, though, this is no reason to upload citizens' data to a public foreign AI: official information, restricted information and sensitive data fall under internal information security rules, and a free chatbot offers neither a contract nor guarantees. We go into detail in "Uzbekistan's personal data law and AI".

4. Hallucinations

Models confidently invent numbers, references to non-existent regulations and quotes. For an official document that is unacceptable: every fact must be checked against the primary source.

5. Blurred accountability

"The AI wrote it" is not a defence. A specific officer and manager are always responsible for a document, a reply to a citizen or a figure in a report.

The core principle: AI is an assistant, not an executor. It prepares the draft; a person checks and signs.

Myths about AI in the public sector

AI attracts plenty of extreme views. Here are the four most common misconceptions:

• "AI will replace civil servants." In practice it takes over routine work — drafts, summaries, translations — while accountability, communication with citizens and decisions remain with people.
• "If the service is paid, the data is safe." Paying does not answer the question of where the data is physically stored and whether you are allowed to send it there.
• "The model knows Uzbek law." A model may know general principles but easily mixes up versions, numbers and dates of documents. Always check legal provisions on lex.uz.
• "It is easier to ban everything." A ban without an alternative leads to shadow use on personal phones — the most dangerous scenario of all.

What international practice teaches

Governments in many countries have already published generative AI guidance for their staff. The wording differs, but the principles repeat: do not enter confidential or personal data into public services, keep a human accountable for the final document, be transparent about where and how AI is used, and train employees. These principles are easy to adapt to Uzbekistan, taking national personal data requirements into account.

Data classification: what you can send to an AI model

The most practical security tool is a simple three-level classification of information:

1. Public data — published laws, press releases, public statistics, texts from the agency website. Fine for public AI services.
2. Internal data — working drafts, internal procedures, anonymised reports. Only enterprise plans with training disabled or local models, after names and identifiers are removed.
3. Confidential data — citizens' personal data, official secrets, financial information, draft decisions before publication, anything classified. Only isolated local solutions inside the organization's perimeter — or no AI at all.

If an employee is unsure which level a document belongs to, it should be treated as confidential by default.

Examples: which level a document belongs to

• A published law or resolution — public data.
• A press release already on the website — public data.
• A draft press release before approval — internal data.
• An internal procedure or job description — internal data.
• A summary report without names or non-public indicators — internal data.
• A citizen request with full name and contact details — confidential data.
• HR documents, salary information — confidential data.
• A draft decision, budget or tender documentation before publication — confidential data.

Attach a list like this to the internal policy and keep adding examples from the agency's own practice: staff find their way by examples far more easily than by definitions.

Safe patterns for using AI

Anonymisation

Before text goes to a model, remove full names, personal ID numbers (PINFL), passport numbers, phone numbers, addresses, amounts and counterparty names. Tasks like "improve the structure" or "make it shorter" do not need this data.

Enterprise plans

Business and enterprise versions of major AI services generally do not train on customer data and offer access administration. That is a step forward, but on its own it does not settle cross-border transfer questions for personal data (where the data is stored, what contractual guarantees apply) or internal bans on sharing official information.

Local (on-premise) models

Open-weight language models can be deployed on the organization's own servers or in a data centre in Uzbekistan. Data never leaves the perimeter, and IT controls access. More in "On-premise AI for organizations".

Human review

Any AI-generated text is checked by a responsible officer before publication or sending: facts, figures, legal references and tone.

Comparing the options: public service, enterprise plan, local model

Each option has its niche. Here is how they compare on the criteria that matter to a government organization:

• Where data is processed. Public service — on the provider's servers, usually abroad. Enterprise plan — also with the provider, but with contractual guarantees. Local model — inside the organization's perimeter or in a data centre in Uzbekistan.
• Training on your data. Public service — depends on the plan and settings. Enterprise plan — usually excluded by the contract terms. Local model — fully under your control.
• Access management. Public service — personal accounts with almost no oversight. Enterprise plan — a central admin console. Local model — integration with internal access management systems.
• Answer quality. Large cloud models are usually stronger at complex text tasks. Open local models have improved markedly but need to be selected and tuned for your tasks and working languages.
• Cost and timing. A public service is up and running in a day. An enterprise plan means a subscription and contract negotiation. A local model means servers, specialists and support, but no dependence on a foreign provider.
• Suitable data. Public service — public data only. Enterprise plan — public and anonymised internal data. Local model — including confidential data, provided protection is configured properly.

In practice, many organizations end up with a hybrid set-up: cloud services for public tasks and a local model for working with internal documents.

How to anonymise text before sending it to a model

Anonymisation sounds complicated, but for most tasks it takes a few minutes of manual work:

1. Decide what the model actually needs. To edit style or structure, it needs no names, amounts or dates.
2. Replace people with roles. A specific name becomes "Applicant A"; a job title stays without a surname.
3. Remove identifiers. PINFL numbers, passport series, phone numbers, addresses, account and contract numbers.
4. Generalise sensitive figures. An exact amount can become "X soums", a specific district "district N".
5. Check indirect clues. A rare job title, a unique event or a combination of facts can identify a person even without a name.
6. Put the data back afterwards. Insert the real details in your own editor, not in the chat.

If a document cannot be anonymised without losing the point of the task, that is a signal: process it only locally, or without AI at all.

How to check AI output: a short protocol

Human review only works when it is clear what exactly to check. For official texts, five steps are enough:

1. Figures. Every number is checked against its source — a report, the statistics website, the original table. Numbers without a source are removed.
2. Legal references. The number, date and title of each document are checked on lex.uz, including the current version.
3. Names and titles. Models often distort job titles and the names of organizations and programmes, or translate them in their own way.
4. Logic of conclusions. Do the conclusions follow from the facts presented, or has the model "invented" a connection?
5. Tone and language. Does the text match the official style, and does it read naturally in Uzbek and Russian?

For important documents, a "four eyes" rule helps: a draft prepared with AI is checked not only by its author but also by a second officer.

A step-by-step rollout plan for an agency

1. Audit current practice. An anonymous survey: who uses which services and for what tasks.
2. Internal policy. A short document: approved services, data classification, prohibitions, review procedure. See an example in "ChatGPT for civil servants: the rules".
3. Pilot on public tasks. Presentations, website copy, translation of public materials.
4. Staff training. How to write prompts, check answers and anonymise data.
5. A solution for internal data. An enterprise plan or a local model, depending on pilot results.
6. Regular review. Technology and regulation change fast; update the policy at least every six months.

How to measure the value of AI

To keep adoption from becoming a box-ticking exercise, the pilot needs measurable goals. Simple metrics:

• how long it takes to prepare a standard document or presentation, before and after;
• how many edits a manager makes to the draft;
• how many citizen requests are classified correctly the first time;
• the share of staff who have been trained and use approved services;
• the number of incidents where data that should not be there ended up in an AI tool.

After two or three months, these figures will show where AI genuinely saves time and where it creates extra checking work.

Common mistakes when introducing AI in a government organization

• Starting with the most sensitive task. A pilot on citizen requests or HR documents immediately runs into personal data requirements. It makes more sense to start with public materials and build experience.
• Buying technology without a use case. First understand which tasks take the most time, and only then choose a service or deploy a model.
• Writing a policy nobody reads. A document running to dozens of pages does not work. Staff need a single page with clear do's and don'ts and examples.
• Forgetting about languages. Agencies work in Uzbek, Russian and sometimes English. Model quality in Uzbek varies noticeably, so test on real texts in every working language.
• Not appointing an owner. Without someone responsible, a pilot quickly turns into scattered experiments by individual staff.
• Punishing mistakes. If reporting that extra data ended up in a chatbot leads to punishment, people will start hiding incidents.

A checklist for managers

• Is there an approved list of permitted AI services?
• Do employees know which data must never be pasted into chatbots?
• Is training on conversations switched off in settings?
• Is someone in the organization responsible for AI?
• Is there a rule that AI-generated text must be reviewed?
• Do we know where personal data is processed, and is there a lawful basis for any transfer abroad?

The role of IT and information security

AI adoption is not just a matter for the press office or a few enthusiasts. IT and the information security team should be involved from the start:

• assessing services before approval: where data is stored, whether training can be disabled, how administration works;
• creating and disabling staff accounts centrally;
• setting network rules: which AI services are reachable from the internal network;
• keeping an incident log and reviewing incidents without blame, so people are not afraid to report mistakes;
• owning local solutions if the agency decides to deploy its own model.

When security is involved from day one, AI becomes a managed tool rather than a source of anxiety.

Questions to ask an AI vendor

Before buying a service or commissioning a local solution, get written answers to a few questions:

1. Where exactly is data stored and processed, including backups and logs?
2. Are prompts and uploaded files used to train models, and can this be prohibited by contract?
3. How long are conversations retained, and can they be deleted on request?
4. How is access managed: single sign-on, roles, deactivation of staff who leave?
5. Is there a log of user actions, and can the security team access it?
6. How well does the solution work in Uzbek and Russian, and can you test it on your own materials?
7. Who updates the system and fixes vulnerabilities, and how quickly?

Vague answers to the first two questions are reason enough not to use the service for anything except public data.

Transparency towards citizens

Trust in government communication is a valuable resource. If an agency uses AI in services that citizens interact with directly, such as a chatbot on its website, it should say so clearly and always offer a way to reach a human officer. Internal drafts that a person then checks and signs usually need no separate label: the responsibility already lies with the officer. What matters is that these rules are written into the internal policy rather than left to individual taste.

How to start without risk

The safest starting point is tasks that use only public data and where the result is easy to check by eye: presentations for speeches, visualising public statistics, adapting texts for citizens. Here AI saves hours while the risk is minimal.

If you need help with presentations and reports for a public agency, the Dasta Creative team uses AI exactly this way: as an accelerator for draft work, with manual fact-checking and design that matches the organization's status. Learn more on our presentations and reports page.