How to Choose the Best AI Platform From Iraq

Whether you can pay for it at all

The World Bank's Global Findex 2025, surveying Iraqi adults in 2024, records 76.2% who used the internet in the past three months and 30.2% with an account at a financial institution or mobile money provider. Debit card ownership is 23.8%, up from 9.8% in 2021. But only 13.8% had bought anything online, and of all adults 7.7% paid cash on delivery for an online purchase against 6.2% who paid digitally. Comparing those two published rates, more online buyers in Iraq settle at the door than settle electronically. A foreign subscription has no door.

What bites at the size of a twenty-dollar subscription is the standing surveillance. A circular of 30 May 2024 requires the anti-money-laundering unit inside every electronic payment service provider to monitor card movements, prepaid, credit and debit alike, under Article 10 of Law No. 39 of 2015: it must check those movements against the limits and ceilings set for the cards and refer suspicious ones to the issuing bank. What the circular does not do is tell anyone to refuse a payment. A card declined without explanation may be that rule working rather than a fault at the till, but that is inference: the circular says nothing about declines, and I found nothing that does.

One assumed route is genuinely closed, and a second thing people blame turns out not to be the cause. A central bank decision of 30 March 2022 bars cards and electronic wallets from speculation and trading in digital currencies of all types; a reaffirmation of 6 February 2023 told banks and payment companies to take administrative, legal and technical measures to prevent and track such transactions under the 2015 anti-money-laundering law. And Iraq's addition on 19 June 2026 to the Financial Action Task Force's list of jurisdictions under increased monitoring is not itself why a payment is refused abroad: the task force's own statement says its standards "do not envisage derisking, or cutting-off entire classes of customers, but call for the application of a risk-based approach". A refusal is a commercial decision taken under the listing, not compliance with it. Whether the listing, or the United States sanctions programme that carries Iraq's name and works through designated persons rather than a country-wide embargo, actually causes an ordinary cardholder to be declined, I found no study measuring.

• World Bank, Global Findex 2025 country-level data (Iraq, survey year 2024)
• Central Bank of Iraq, 30 May 2024: AML units must monitor prepaid, credit and debit card movements
• Central Bank of Iraq, 30 March 2022: cards and wallets barred from digital-currency trading
• Central Bank of Iraq, 6 February 2023: banks and payment companies to prevent and track digital-currency transactions
• FATF, Jurisdictions under Increased Monitoring, 19 June 2026
• US Treasury: Iraq-Related Sanctions programme

What the price is in dinars, and what the gap tells you

Card spending abroad is settled at the official rate, and the official rate is a schedule rather than a single number. The central bank's own converter served 1,310.000 dinars to the dollar on 31 August 2026; its March 2023 card package sets spending outside Iraq on electronic payment cards at 1,320. Those are not two answers to one question. They are two rungs of one managed regime, each with its own scope, and the rung that governs a foreign subscription paid by card is 1,320. Exchange houses were buying dollars near 1,525 in early August 2026 — a rate no public institution publishes, so check the day's number yourself, because it moves.

That spread is the entire local reseller economy in one figure. At the card rate of 1,320 a twenty-dollar subscription is 26,400 dinars; at 1,525 it is 30,500, a difference of 4,100 dinars a month and 41,000 at a two-hundred-dollar tier. I found no published, auditable price list from any reseller of foreign subscriptions here, and could not cite one if I had. What transfers is the arithmetic: divide the dinar price you are quoted by the dollar price on the vendor's own page, and read off the exchange rate the seller is implying. Anything far above the day's market rate is margin.

• Central Bank of Iraq, 14 March 2023: card spending abroad settled at the official rate of 1,320
• Central Bank of Iraq — the currency converter that served 1,310.000 on 31 August 2026
• Iraq currency exchange rates, 5 August 2026 — exchange houses buying at 152,500 dinars per $100

Whether you may lawfully upload the document

Iraq has no comprehensive personal data protection statute, and the regulator says so inside a regulation that is in force. The Communications and Media Commission's Framework Regulations for Digital Platforms and Services, effective 17 February 2025, require providers at section 6.1.3 to comply with Iraqi laws on personal data protection, then add that the commission "will be publishing a personal data protection law in due time". Sections 6.1.5 and 6.1.6 say the same of an artificial intelligence regulation and a cloud regulation. A regulator is writing, in a binding text, that the law it points at does not exist yet.

The same regulation already has a category for the tool you are choosing. Article 4.2 lists artificial-intelligence platforms among the types in scope, and Article 4.1 reaches platforms with no physical presence in Iraq that serve users or earn revenue here. The regulatory-tools table at Article 5 sets foreign-based AI platforms at notification and locally based ones at licence; the licensing requirements sit in Annex 2, and one of them is a commitment to data protection measures "including data localization where necessary". So one question for a foreign provider is whether it has notified the commission. I found no public register of notified platforms, so you can ask it but probably not verify it, and a regulation in force is not evidence of one enforced.

What binds the person doing the uploading is older, and criminal. Article 437 of the Penal Code No. 111 of 1969 punishes anyone who, by reason of his office, profession, trade or work, is privy to confidential information and discloses it outside the circumstances prescribed by law: detention of up to two years, a fine, or one of the two. The text carries its own exception — no penalty where the disclosure was authorised, or was intended to report or prevent an offence. Outside that, the duty is criminal rather than contractual, and it attaches to you personally: no vendor's policy discharges it and no clause moves it onto the vendor. The 200-dinar fine set in 1969 is now nominal; the deterrent is the detention and the professional consequence.

One detail in the professional statutes outweighs the rest. Article 46 of the Lawyers' Profession Law No. 173 of 1965 forbids a lawyer to disclose a secret entrusted to him or learned through his profession, even after his mandate has ended, unless disclosure would prevent a crime. Set that beside the best answer a vendor can give about retention, which is a number of days. The duty has no end date; the retention period does, and the clock you answer for is the one that never stops. With no statutory floor beneath you, the contract is the whole of your protection.

• Framework Regulations for Digital Platforms and Services, in force 17 February 2025 (English PDF) — scope at 4.1 and 4.2, regulatory tools at Article 5, licensing requirements at Annex 2
• Iraqi Penal Code No. 111 of 1969 as amended, English translation — Articles 437 and 438
• Lawyers' Profession Law No. 173 of 1965, Arabic text — Article 46 (the host's file title misdates the law to 1966)
• Constitution of Iraq (2005), Article 17 on personal privacy and Article 40 on correspondence
• What happens to what you type into an AI tool — retention, training and human review

Which Arabic it actually produces

"Supports Arabic" is not an answer to a question anyone asked. It can mean the interface has been translated, or that the model does not error on Arabic input. Iraqi Arabic is not one variety — it names at least two types that split again inside the country, and the evidence for that is in the article on Iraqi Arabic and language models. What belongs here is a test that takes two minutes.

Ask for four sentences in one prompt: "I said to him that I could sell it", "the student is studying in the library", "there is no food in the fridge", "Ali did not study". Each forces a diagnostic environment, on the forms set out in that article; treat the whole procedure as a practical heuristic rather than a validated instrument, and read the reply as a variety identification and not as a pass or a fail. Standard morphology throughout means it drifted, the commonest outcome. Forms consistent across all four mean one real variety; if it is the Baghdadi one, the Muslim Baghdadi variety a published English reference grammar covers — its publisher summarises the book as reflecting "Iraqi Arabic as spoken by Muslims in Baghdad" — that is a defensible answer to "Iraqi" and also a choice made silently on your behalf. Markers from different varieties in one reply are a blend no described dialect produces.

Then three steps the linguistic article does not have. Ask it to name the variety it just produced and compare that to the forms it generated: a tool that produces one variety and calls it "Iraqi" cannot tell you what it did. Test the reverse direction — an Iraqi sentence in, a standard rendering out — because most business use needs comprehension rather than generation, and the two fail differently. And run it twice, because one generation is one sample. Scanning ordinary prose for marker words is the weaker alternative: of 120 curated Iraqi-distinctive lexical cues tested against annotated dialect data, only seven matched any sample at all, and together they recovered .03 of the sentences annotators judged Iraqi. A handful of marker words is not a reliable signal. Forcing the environment is.

• Iraqi Arabic and language models — the varieties, and the forms this test relies on
• Keleg, Goldwater & Magdy (2025), Table 1a — 120 Iraqi cues, seven matching, recall .03
• Georgetown University Press, A Short Reference Grammar of Iraqi Arabic — "the only volume of its kind, reflecting Iraqi Arabic as spoken by Muslims in Baghdad"

Whether you can leave with your work

An Iraqi user has no statutory right to export anything. The communications regulator's own regulation says only, at section 6.1.3, that providers must comply with Iraqi laws on personal data protection and that a personal data protection law will be published in due time; it confers no right on a user to take anything out. So an export button is a product feature that can be withdrawn, not a right you could enforce. For contrast, Article 20 of the European data protection regulation lets a data subject receive the personal data they provided in a machine-readable format and transmit it elsewhere, where the processing rests on consent or a contract and is carried out by automated means — and even that covers what was provided rather than what was derived from it, and runs only against a controller that still exists.

So the question is not whether an export exists but what it contains and what it leaves behind. Your documents are yours; the index built from them is not, because embeddings belong to the model that produced them, so moving means re-embedding everything. A prompt library tuned against one model is a fitted artefact rather than a document, given the up-to-76-point spread across prompt formats that Sclar, Choi, Tsvetkov and Suhr (ICLR 2024) measured on one open model. A fine-tune is worth nothing once its base version is retired, and history and memory are the least portable and the most quietly load-bearing. None of that is peculiar to Iraq. What is peculiar is that here nothing stands behind it.

Ask whether you can export today, in one action, without asking anyone, and whether you can still do it with a lapsed subscription. That is a live local risk rather than a hypothetical one: a card can stop working under the monitoring rule described above, and an account bought through a reseller may not be in your name to begin with. Ask what format, and test the answer: something other than this vendor can open it, and you can read it. If the whole of your answer is a support agent's reassurance you have nothing, and here more than elsewhere, because there is no regulator to complain to afterwards. These questions apply to our own document feature exactly as to anyone else's.

• Article 20 of the EU General Data Protection Regulation, the right to data portability
• Sclar, Choi, Tsvetkov & Suhr (ICLR 2024), up to 76 accuracy points from prompt format alone
• Answers from your own documents — how a retrieval index is built, and why it is model-specific

The claims that prove nothing

A benchmark number a vendor reports about itself, with no method attached, is not a measurement of a model. Sclar, Choi, Tsvetkov and Suhr (ICLR 2024) measured differences of up to 76 accuracy points across plausible prompt formats on one open model, and argue for reporting a range rather than a single number. Ahuja, Gumma and Sitaram (2024) tested seven frequently used multilingual benchmarks against seven models and report that almost all the models show signs of contamination with almost all the benchmarks; the abstract carries no per-language rates, so it is not a finding about Arabic. Singh et al. (2025) documented, on one large public model ranking, undisclosed pre-release testing of many variants with selective retraction of weak scores. None of that proves a vendor cheated; it proves a bare number is not evidence.

An unaudited user count is weaker still, because the word has no standard definition. Registrations, monthly actives, devices, seats and API keys are different populations and all get called users, nobody audits the figure, and an install band on an app listing is a rounding threshold counted against accounts or devices rather than people. Locally it is worth less again: a headline drawn overwhelmingly from another market tells you nothing about whether the tool reads your Arabic or takes your dinar.

The local form of the disease is "used by such-and-such ministry". A pilot, a memorandum and a signed contract are three different things, and the habit of quoting proposals in the past tense is documented in the article on what AI in Iraq consists of. As for "supports Arabic", the one public shared task that translates Iraqi into the standard language asked its reviewers for at least 500 reviewed segments per dialect; every dialect cleared that except Iraqi, which reached 277, and once 200 went to validation the Iraqi test set was 77 sentences. Results were reported as a single score across the five dialects together, so that task publishes no Iraqi number of its own.

• Sclar, Choi, Tsvetkov & Suhr (ICLR 2024), up to 76 accuracy points from prompt format alone
• Ahuja, Gumma & Sitaram (2024), contamination report for multilingual benchmarks
• Singh et al. (2025), The Leaderboard Illusion
• OSACT6 (2024), dialect-to-MSA translation shared task overview — Table 1: Iraqi 277 reviewed segments, 200 validation, 77 test
• How to read AI benchmarks — contamination, harness variation and what a score omits
• What artificial intelligence in Iraq actually consists of

The questions, and what a non-answer sounds like

Every question above has a recognisable good answer and a recognisable non-answer, and the non-answers repeat. Collect them from a dated page or a contract rather than a chat window, before you depend on the tool rather than after.