Historical publication / 25 Apr 2026
Five domains.
Uneven evidence.
AI Services Censorship Report — April 2026. The original study, its measurements and its limits.
A historical record, with gaps.
An anomaly is not, by itself, proof of a government block, a specific mechanism or successful account/API access. This page does not check current AI-service availability.
The page and JSON supply 17 country rows; the Markdown has six additional rows. Published totals and labels are retained as source claims, with the discrepancies below.
Read the source discrepancies before citing
- The report and JSON declare 74 measured countries and 23 with a restriction signal. The page/JSON matrix contains 17; the Markdown contains 23. The extra six rows have labels/sample counts but no numeric anomaly rates in that file.
- Pakistan's Gemini cell is labelled partial at a 5% rate with n=1, despite the stated 50% partial threshold. Several other partial cells equal 50%, while prose says above 50%. These are preserved source labels, not reclassified measurements.
- The narrative says Claude had zero failures across 14 China probes. Both bundled page and JSON leave that cell's rate and n unspecified. Missing values remain missing here.
- The text says Gemini had hard blocks in two countries but names only China as a hard block; the other listed cases are partial/interference. No second hard-block country is invented.
- The blocking-method column is reported empty for all 233 anomalous events. DNS/SNI mechanisms, certificate-based explanations and ISP-level causation are not independently established by this aggregate.
- The current accessibility endpoints have no April time filter. The original SQL has a start bound but no end bound. Neither is, by itself, a reproducible frozen April snapshot.
- The original article says nine additional providers but names ten. Its promised July report, future analyses, probe funding and update schedules remain historical text, not new commitments or verified delivery.
Inspect each cell.
Exact rates from the downloadable JSON, alongside its supplied labels and n. “No restriction” is the source category; an unspecified rate or sample size is not treated as zero.
| Country | ChatGPTchatgpt.com | ChatGPT (legacy domain)chat.openai.com | OpenAI websiteopenai.com | Claudeclaude.ai | Geminigemini.google.com |
|---|---|---|---|---|---|
| ChinaCN | block100%n=14 | block100%n=12 | partial50%n=1 | No restriction¹Rate not suppliedn not supplied | block100%n=14 |
| RussiaRU | interference6.24%n=14 | interference8.07%n=12 | interference3.79%n=11 | interference4.12%n=10 | interference6.94%n=14 |
| PakistanPK | interference28%n=5 | interference5%n=2 | No restriction¹Rate not suppliedn not supplied | interference4%n=2 | partial5%n=1 |
| CanadaCA | interference2.7%n=4 | interference5.6%n=8 | interference2.1%n=4 | interference3.1%n=5 | interference3.4%n=8 |
| SpainES | interference1.9%n=5 | interference2.6%n=5 | interference1.6%n=9 | interference1.3%n=8 | interference1.4%n=5 |
| GermanyDE | interference1.7%n=1 | interference2.6%n=3 | interference1.4%n=3 | interference2.6%n=2 | interference4.1%n=4 |
| United KingdomGB | interference1.9%n=2 | interference7.7%n=1 | interference2%n=1 | interference1.9%n=1 | interference1.3%n=1 |
| United StatesUS | interference1.1%n=1 | interference2%n=2 | interference1.3%n=4 | interference1.8%n=2 | interference1.3%n=2 |
| KenyaKE | interference4.5%n=1 | No restriction¹Rate not suppliedn not supplied | interference4.3%n=1 | interference6.7%n=1 | interference4.3%n=1 |
| KazakhstanKZ | No restriction¹Rate not suppliedn not supplied | interference4.3%n=3 | No restriction¹Rate not suppliedn not supplied | interference5%n=1 | interference4.3%n=1 |
| South AfricaZA | partial50%n=1 | No restriction¹Rate not suppliedn not supplied | interference2.7%n=1 | No restriction¹Rate not suppliedn not supplied | partial50%n=1 |
| BrazilBR | interference6.3%n=1 | No restriction¹Rate not suppliedn not supplied | No restriction¹Rate not suppliedn not supplied | interference3.8%n=1 | interference7.7%n=1 |
| IndiaIN | No restriction¹Rate not suppliedn not supplied | No restriction¹Rate not suppliedn not supplied | partial50%n=1 | No restriction¹Rate not suppliedn not supplied | partial50%n=1 |
| NigeriaNG | No restriction¹Rate not suppliedn not supplied | No restriction¹Rate not suppliedn not supplied | partial50%n=1 | No restriction¹Rate not suppliedn not supplied | No restriction¹Rate not suppliedn not supplied |
| IraqIQ | No restriction¹Rate not suppliedn not supplied | No restriction¹Rate not suppliedn not supplied | No restriction¹Rate not suppliedn not supplied | partial50%n=1 | No restriction¹Rate not suppliedn not supplied |
| AlgeriaDZ | No restriction¹Rate not suppliedn not supplied | No restriction¹Rate not suppliedn not supplied | No restriction¹Rate not suppliedn not supplied | partial50%n=1 | No restriction¹Rate not suppliedn not supplied |
| ArmeniaAM | Not suppliedRate not suppliedn not supplied | Not suppliedRate not suppliedn not supplied | Not suppliedRate not suppliedn not supplied | Not suppliedRate not suppliedn not supplied | partial50%n=1 |
¹ Original source label. It does not establish zero failures, sufficient coverage, successful signup or current service access. Russia's JSON preserves greater decimal precision than the original page; the complete original page matrix remains in the article disclosure.
Six additional country rows in the Markdown only
The public Markdown supplies the rows below, absent from the page/JSON matrix. Labels and sample counts are reproduced literally; missing rates are not inferred. These six rows help explain the 23-versus-17 presentation gap but do not independently validate the published service totals.
| Country | ChatGPT | ChatGPT (legacy domain) | OpenAI website | Claude | Gemini |
|---|---|---|---|---|---|
| Belarus (BY) | n/a | INT (n=1) | n/a | n/a | n/a |
| Australia (AU) | INT (n=1) | n/a | n/a | n/a | n/a |
| Japan (JP) | n/a | n/a | n/a | n/a | INT (n=2) |
| Mexico (MX) | n/a | n/a | n/a | n/a | INT (n=1) |
| Singapore (SG) | n/a | n/a | n/a | n/a | INT (n=1) |
| Taiwan (TW) | n/a | n/a | n/a | n/a | INT (n=1) |
What the publication reports.
Original aggregate figures, not a new measurement or a count derived from the partial matrix. The downloadable source retains these declarations.
- Measurement events
- 3,604
- Countries measured
- 74
- Block-category events
- 50
- Interference events
- 183
Domain totals and exact stored window
- Start
- 2026-04-12T00:00:00Z
- End
- 2026-04-25T00:00:00Z
The stored timestamps span thirteen days and end at midnight on 25 April. The publication's inclusive prose date range is not silently extended to the end of that day.
| Domain / vendor | Reported events | Reported countries with signals |
|---|---|---|
| chatgpt.comOpenAI | 743 | 13 |
| chat.openai.comOpenAI | 670 | 11 |
| openai.comOpenAI | 738 | 12 |
| claude.aiAnthropic | 726 | 13 |
| gemini.google.comGoogle | 727 | 17 |
The complete original article.
Every original prose section, finding, example, source link and citation is retained below. Historical assertions are preserved for scrutiny; the source discrepancies above apply throughout.
Lede
Across 74 countries with active OONI measurements between 12 April and 25 April 2026, ChatGPT (chatgpt.com) is hard-blocked in 1 country and shows interference in 12 more. Claude (claude.ai) is reachable from China in the same measurement window — every probe in our China dataset that tested claude.ai returned a normal response, while every probe testing chatgpt.com failed. Russia interferes with all five major Western AI services we tested (chatgpt.com, chat.openai.com, openai.com, claude.ai, gemini.google.com), with anomaly rates between 3.8% and 8.1%. Gemini is the most-broadly-restricted AI service overall, showing block or interference signals in 17 countries.
These asymmetries matter for the agent economy. An AI agent that depends on OpenAI's API will behave differently in China than one built on Anthropic's; a research tool that calls Gemini will fail in places where Claude works. The "internet is one thing" assumption baked into most LLM applications is not true, and the gaps are widening on a per-vendor basis. This report documents what we measured. Every claim links back to a live API call you can run yourself.
Methodology
Voidly ingests OONI's web_connectivity measurements every six hours via the OONI Aggregation API and records each domain × country × test outcome as an evidence row in our public database. For this report we selected five domains representing the dominant Western consumer-facing AI services (chatgpt.com, chat.openai.com, openai.com, claude.ai, gemini.google.com) and pulled every measurement event between 2026-04-12 and 2026-04-25. The full pull contains 3,604 measurement events across 74 countries, of which 233 events in 23 countries show either blocking (50) or interference (183).
We classify each (country, service) cell using OONI's anomaly rate. A 100% anomaly rate on a sample of ten or more probes is reported as a hard block; anomaly rates above 50% on smaller samples are reported as partial blocks; anomaly rates below 50% are reported as interference (reachability is degraded but not absolute, often indicating ISP-specific blocking, blockpage delivery in some networks but not others, or transient enforcement). All other cells are reported as accessible. We surface the underlying signal value and sample size on every cell so readers can apply their own thresholds.
Limitations. OONI's coverage is uneven: some countries have hundreds of measurements per domain per week, others single digits. Probe placement is not uniformly distributed across ISPs within a country, so a country-level "accessible" verdict can mask ISP-specific blocking, and a country-level "block" can be ISP-localized. We label sample sizes throughout. We do not currently distinguish DNS poisoning from TCP-RESET from blockpage delivery in this report — OONI exposes the underlying packet-level evidence but our aggregation collapses it; future reports will surface this. Provider-side geographic restrictions (where OpenAI or Anthropic terms-of-service exclude a country) are not captured by network-side measurement and so do not appear here as "blocks" — they show up as accessible in our data even though end users cannot sign up. Voidly's own probe network (100+ operators) cross-checks a subset of these results; we report only the OONI-derived numbers in this report to keep the methodology auditable against a single upstream.
Reproduce: every figure in this report can be regenerated with our public API and the SQL query in the "Open data + reproduce" section below. Source data is OONI, CC BY 4.0.
Finding 1 — Country × service matrix
23 countries show some form of restriction on at least one of the five tracked AI services. Sorted by hard blocks first, then by total restricted services.
| Country | ChatGPT | chat.openai.com | openai.com | Claude | Gemini |
|---|---|---|---|---|---|
| CNChina | BLOCK100% anomaly · n=14 | BLOCK100% anomaly · n=12 | PART50% anomaly · n=1 | okRate not supplied · n not supplied | BLOCK100% anomaly · n=14 |
| RURussia | INT6.2% anomaly · n=14 | INT8.1% anomaly · n=12 | INT3.8% anomaly · n=11 | INT4.1% anomaly · n=10 | INT6.9% anomaly · n=14 |
| CACanada | INT2.7% anomaly · n=4 | INT5.6% anomaly · n=8 | INT2.1% anomaly · n=4 | INT3.1% anomaly · n=5 | INT3.4% anomaly · n=8 |
| ESSpain | INT1.9% anomaly · n=5 | INT2.6% anomaly · n=5 | INT1.6% anomaly · n=9 | INT1.3% anomaly · n=8 | INT1.4% anomaly · n=5 |
| DEGermany | INT1.7% anomaly · n=1 | INT2.6% anomaly · n=3 | INT1.4% anomaly · n=3 | INT2.6% anomaly · n=2 | INT4.1% anomaly · n=4 |
| GBUnited Kingdom | INT1.9% anomaly · n=2 | INT7.7% anomaly · n=1 | INT2% anomaly · n=1 | INT1.9% anomaly · n=1 | INT1.3% anomaly · n=1 |
| USUnited States | INT1.1% anomaly · n=1 | INT2% anomaly · n=2 | INT1.3% anomaly · n=4 | INT1.8% anomaly · n=2 | INT1.3% anomaly · n=2 |
| PKPakistan | INT28% anomaly · n=5 | INT5% anomaly · n=2 | okRate not supplied · n not supplied | INT4% anomaly · n=2 | PART5% anomaly · n=1 |
| KEKenya | INT4.5% anomaly · n=1 | okRate not supplied · n not supplied | INT4.3% anomaly · n=1 | INT6.7% anomaly · n=1 | INT4.3% anomaly · n=1 |
| KZKazakhstan | okRate not supplied · n not supplied | INT4.3% anomaly · n=3 | okRate not supplied · n not supplied | INT5% anomaly · n=1 | INT4.3% anomaly · n=1 |
| ZASouth Africa | PART50% anomaly · n=1 | okRate not supplied · n not supplied | INT2.7% anomaly · n=1 | okRate not supplied · n not supplied | PART50% anomaly · n=1 |
| BRBrazil | INT6.3% anomaly · n=1 | okRate not supplied · n not supplied | okRate not supplied · n not supplied | INT3.8% anomaly · n=1 | INT7.7% anomaly · n=1 |
| INIndia | okRate not supplied · n not supplied | okRate not supplied · n not supplied | PART50% anomaly · n=1 | okRate not supplied · n not supplied | PART50% anomaly · n=1 |
| NGNigeria | okRate not supplied · n not supplied | okRate not supplied · n not supplied | PART50% anomaly · n=1 | okRate not supplied · n not supplied | okRate not supplied · n not supplied |
| IQIraq | okRate not supplied · n not supplied | okRate not supplied · n not supplied | okRate not supplied · n not supplied | PART50% anomaly · n=1 | okRate not supplied · n not supplied |
| DZAlgeria | okRate not supplied · n not supplied | okRate not supplied · n not supplied | okRate not supplied · n not supplied | PART50% anomaly · n=1 | okRate not supplied · n not supplied |
| AMArmenia | —Rate not supplied · n not supplied | —Rate not supplied · n not supplied | —Rate not supplied · n not supplied | —Rate not supplied · n not supplied | PART50% anomaly · n=1 |
Finding 2 — Most-restricted services
Number of countries in which each AI service shows blocking or interference signals (anomaly rate > 0). Bars are ranked descending.
Gemini's broader footprint reflects two effects: it shares Google's *.google.com certificate chain (so it is collateral damage in some Google-blocking regimes), and it is the only one of the five tracked services with hard blocks confirmed in two countries (China, plus partial blocks in Armenia, India, Mexico, Pakistan, Singapore, South Africa, Taiwan). Anthropic's claude.ai, by contrast, is on a standalone domain and an Anthropic-controlled certificate, which appears to insulate it from blanket "block all of Google" DNS rules.
Finding 3 — Vendor asymmetries
The most-cited assumption in AI infrastructure planning is that reachability for "the major Western AI services" is roughly equivalent — if one is blocked, all are blocked. The data does not support this.
- China: ChatGPT (both
chatgpt.comandchat.openai.com) and Gemini are hard-blocked (100% anomaly rate on n=12–14 probes). Claude (claude.ai) is reachable in the same measurement window with zero failures across 14 probes. Anthropic's marketing domain (anthropic.com) was not measured in this window. - Russia: All five Western AI services show measurable interference (3.8–8.1% anomaly rate). The interference is roughly uniform across vendors, suggesting an ISP-level filter rather than a domain-level block list. There is no hard block, but reliability for an automated agent is noticeably degraded.
- Pakistan: ChatGPT shows the highest anomaly rate (28% on chatgpt.com, n=5). Claude and Gemini show much lower rates (4–5%). Pakistan's filtering, where present in this window, is OpenAI-asymmetric.
These differences should be assumed to be temporary — censorship policy moves on weeks-to-months timescales — but for any system that depends on AI service availability today, vendor choice matters per country.
Finding 4 — Blocking method breakdown
Of the 233 block/interference events in this dataset, OONI's web_connectivity test reports an anomaly verdict (test failure) but does not always disambiguate the underlying mechanism. Voidly's evidence database carries a blocking_method column that we populate from upstream packet-level evidence when available; for this AI-domain subset, the column is empty for 100% of rows.
What this means: for AI domains specifically, the OONI aggregation we ingest collapses the per-method detail (DNS poisoning vs TCP-RESET vs HTTP blockpage) into a single "anomalous" verdict. Reconstructing the methods requires going back to the raw OONI measurement files, which we do for high-priority incidents but not for the bulk-aggregated feed used here. The China hard block is widely reported elsewhere as DNS poisoning combined with SNI filtering on the GFW; we have not independently verified that for this measurement window. We are treating "method-unattributed" as a known limitation and flagging it explicitly rather than guessing.
Finding 5 — Measurement window
All evidence in this report was observed between 2026-04-12 and 2026-04-25 — a 13-day window. We do not yet have a longitudinal AI-domain dataset of sufficient depth to publish a "newly blocked this quarter" or "trend over the last 90 days" finding with confidence. The evidence database has been collecting AI-service measurements continuously since February 2026, but coverage of small-country probes was uneven before April. We flag this as a limitation and intend to publish the quarterly trend in the July 2026 report once a comparable Q2 window is available.
For real-time monitoring of these services in the meantime, see Voidly's live AI Censorship Index, which updates every five minutes.
Finding 6 — AI access during election windows
Two countries with restricted AI access in this measurement window also have upcoming elections or ongoing political tensions on Voidly's election-risk calendar: Pakistan (highest ChatGPT interference rate in the dataset, 28%) and Russia (uniform interference across all five services). We do not assert a causal link in this report — the measurement window is short, and we have no pre-election baseline to compare against — but the presence of AI services on the same network paths as restricted news, social, and messaging domains means an outage triggered for political reasons is likely to take AI access with it. We will publish a controlled before/after analysis in subsequent reports as more election-coincident data accumulates.
Voidly's election risk briefings API gives 7-day forecasts that combine ML-derived risk with the election calendar.
Finding 7 — Coverage gap on emerging providers
We attempted to measure nine additional AI services beyond the big three vendors: Perplexity (perplexity.ai), GitHub Copilot (copilot.github.com), Mistral (chat.mistral.ai), DeepSeek (chat.deepseek.com), Grok (grok.x.ai), Meta AI (meta.ai), Hugging Face (huggingface.co), Cursor (cursor.com), Replicate (replicate.com), and Cohere (cohere.ai). OONI's measurement footprint for these domains is currently insufficient — fewer than 10 events per domain across all countries in the 13-day window — so we cannot make accessibility claims about them with the same confidence as the big three. This is a real gap. The censorship measurement community has historically focused on social, news, and messaging services; AI-specific monitoring is in its first generation. Voidly is funding additional probes for these domains via our community probe program; if you want to contribute, see /probes.
Implications
For AI labs: reachability is a product feature. China shows that vendor-by-vendor decisions get made independently, and that a domain-naming choice (claude.ai vs chat.openai.com) can determine whether the service is reachable. Whatever the labs do at the policy layer, domain selection, certificate scope, and sub-domain proliferation each have downstream censorship-resistance implications.
For AI-using businesses: single-vendor AI dependencies are a per-country availability risk, not just a price risk. Deployments serving global users should plan for graceful degradation when a given vendor is unreachable in a given country. Multi-vendor routing — call Claude from China, OpenAI from elsewhere — has a measurable basis here.
For agent developers: the agent economy assumes consistent reachability. It is not consistent. Agent frameworks should surface per-country availability of their underlying model providers, and treat AI-service unreachability as a recoverable error class with vendor failover — not as a fatal exception.
For journalists and policy researchers: everything in this report is reproducible against OONI's public corpus. The measurement record will support claims that hold up to legal and academic scrutiny. The data is CC BY 4.0; cite it.
Open data + reproduce
Every figure in this report can be reproduced. Two paths:
Live API: for any (domain, country) pair, query
curl 'https://api.voidly.ai/v1/accessibility/check?domain=claude.ai&country=CN'
curl 'https://api.voidly.ai/v1/accessibility/check?domain=chatgpt.com&country=CN'
curl 'https://api.voidly.ai/v1/accessibility/check?domain=gemini.google.com&country=RU'Full dataset: the underlying SQL query against Voidly's evidence database is:
SELECT domain, country_code, signal_type, signal_value, COUNT(*) AS samples
FROM evidence
WHERE domain IN (
'chatgpt.com', 'chat.openai.com', 'openai.com',
'claude.ai', 'gemini.google.com'
)
AND signal_type IN ('block', 'interference')
AND observed_at >= '2026-04-12T00:00:00Z'
GROUP BY 1, 2, 3, 4
ORDER BY 1, 2;The output of this query is mirrored verbatim in /research/ai-censorship-2026-04/data.json for offline analysis. A markdown copy of this entire report is at /research/ai-censorship-2026-04/report.md.
Subscribe to the Voidly Research RSS at /feed.xml for new findings.
About Voidly
Voidly is a censorship measurement and intelligence platform. We aggregate public network-measurement data (OONI, CensoredPlanet, IODA) with our own global probe network into a queryable evidence database covering 130 countries and 470+ documented censorship incidents. Our data and APIs are free for researchers, journalists, and academics under CC BY 4.0. We also build privacy tools, encrypted agent infrastructure, and a hosted amount-zero task-handoff test that operate inside the same network conditions our research describes. The former experimental credit-ledger interface is retired. The tools and the research share infrastructure so we can document our own products' reachability with the same instrumentation we apply to OpenAI's.
Original citation and press contact
Read / reproduce / cite
Keep the date
with the claim.
The article declares CC BY 4.0. Raw OONI measurements retain upstream terms; see the data licensing notes. Current API responses cannot replace the dated source when citing April findings.