Governments banned AI tools 18 times across 15 countries between March 2023 and October 2026. We compiled every verified government AI ban into one database, with dates, scope, stated reasons, and current status, because no single source tracks them all.
The pattern is not what most coverage suggests. Only two bans in the database target the public. The other 16 target government devices, and they split into two distinct logics: bans on Chinese models over where data goes, and bans on Western models over who signed the contract.
The database
| Date | Government | Tool banned | Scope | Stated reason | Status |
|---|---|---|---|---|---|
| Mar 31, 2023 | Italy (Garante) | ChatGPT | Public | Privacy: unlawful data collection | Lifted Apr 28, 2023 |
| Jan 29, 2025 | India (Finance Ministry) | ChatGPT, DeepSeek | Ministry devices | Confidentiality of government documents | Active |
| Jan 30, 2025 | Italy (Garante) | DeepSeek | Public | Data practices; EU rules apply | Active: only public ban worldwide |
| Jan 31, 2025 | Taiwan | DeepSeek | Public sector, schools, SOEs, critical infrastructure | Cross-border data transfer, leaks | Active |
| Feb 2025 | South Korea | DeepSeek | Ministries | User-data collection | Partially lifted Apr 28, 2025 |
| Feb 4, 2025 | Australia | DeepSeek | All government devices | National security | Active |
| Feb 2025 | Canada | DeepSeek | Federal phones; British Columbia | Data security | Active |
| Feb 2025 | Netherlands | DeepSeek | Central government staff | Advisory: told not to use | Advisory |
| Feb 2025 | Japan | DeepSeek | Officials | Advisory: asked to avoid | Advisory |
| Jul 9, 2025 | Czechia | DeepSeek | State administration | Chinese law may compel data access | Active |
| Jun 2025 | Germany (Berlin DPA) | DeepSeek | App stores | Requested Apple/Google removal; unlawful data transfer to China | Removal requested |
| Jul 2026 | United Kingdom | DeepSeek | Official work | Not approved for official use | Active |
| Dec 1, 2025 | Belgium | DeepSeek | Federal work devices | Data security | Active |
| 2026 | Sweden (Skatteverket, SiS) | ChatGPT, Claude | Agency systems | Information security | Active |
| Oct 20, 2026 | United States (DHS) | DeepSeek | All federal civilian agencies | Data transfer to China could yield intelligence on US persons | Active |
| Oct 26, 2026 | Sweden (Regeringskansliet) | ChatGPT, Claude | Staff work computers | No contract in place; information security | Active |
Two US-state rows complete the count: multiple states banned DeepSeek on government devices starting January 2025, and France's privacy watchdog opened a formal questioning of DeepSeek in January 2025. That is 18 actions across 15 countries.
Two logics, one destination
Fifteen of the 18 actions target DeepSeek, and their stated reason is consistent: data flowing to China. The DHS order of October 20, 2026 is the bluntest, warning that transferred information "could produce actionable intelligence on US persons and/or government programs." Taiwan, Australia, Czechia, and South Korea use the same logic with different wording. Italy is the outlier that proves the rule: it is the only government to block DeepSeek for the public, not just officials.
The remaining three actions target Western models, and their logic is contractual, not geopolitical. Sweden accounts for two: its agencies and then the Government Offices, which blocked ChatGPT and Claude on staff computers effective October 26, did not claim the models are unsafe. It said staff were using AI services with no agreement in place, which could have "consequences for our information security," and pointed employees to Microsoft Copilot, the one service that cleared a security review. (India's January 2025 order covers both Western and Chinese tools on confidentiality grounds.) Italy's 2023 ChatGPT block, the only public ban of a Western model, was a privacy enforcement action and was lifted within a month once OpenAI adjusted.
The destination is the same in both logics: government work consolidates onto vendors that can sign the security paperwork. Right now that is Microsoft, whose Copilot benefits from exactly this kind of review everywhere it runs.
What the bans do not establish
A government-device ban is not a safety verdict on the model. It is a procurement verdict on the contract. Sweden still lets some staff reach ChatGPT and Claude through an in-house platform it built itself, which means the models are fine and the paperwork was the problem. Similarly, most DeepSeek bans leave public use untouched; the model is legal for ordinary people almost everywhere. Reading a device ban as proof a model is dangerous is the most common misreading of this database. The bans also say nothing about efficacy: South Korea lifted its ministry block after DeepSeek revised its privacy policy, which suggests at least some of these are negotiable rather than permanent.
The approval model, explained
Sweden's approach is the one to watch because it completes the loop most organizations leave half open: ban the unsanctioned tools at the network level, then actually provide a sanctioned alternative that passed review, so employees have somewhere legitimate to go. A policy memo without an approved alternative just drives usage underground; an approved alternative without enforcement leaves the memo as decoration. The Regeringskansliet did both, and paired it with tightened Copilot security settings after its own review. For enterprises, that is the template: the question is never "should staff use AI" but "which AI has a signed contract," and the answer has to be a tool, not a paragraph. The US federal DeepSeek purge shows the same consolidation from the supply-chain side.
By ByteHerald Staff. Italy's Garante blocked ChatGPT on March 31, 2023 and lifted the block April 28, 2023, the only public ban of a Western model. Italy blocked DeepSeek for the public on January 30, 2025, the only public DeepSeek ban worldwide. The US Department of Homeland Security ordered DeepSeek removed from all federal civilian agencies on October 20, 2026. Sweden's Regeringskansliet blocked ChatGPT and Claude on staff computers effective October 26, 2026, directing staff to Microsoft Copilot, as reported by Sveriges Radio's Ekot on October 23, 2026.
Comments
Post a Comment