IASER Dispatch — Issue 001

Assalamu alaikum, and welcome to the first IASER Dispatch — the newsletter of the Islamic Alliance for Safe, Ethical, and Responsible AI. IASER launched this month, hosted by the Center for Islamic Knowledge while we set up as an independent nonprofit. Every couple of weeks we’ll send one short dispatch: what we’re building, where we need hands, and what’s worth knowing where AI, Islam and ethics meet.

In this issue:

• Survey and membership launch — membership is now open, and our community survey is live.
• Events — our London panel on tech sovereignty, Islam and the individual, and where to find us in Istanbul this month.
• Deep dive: Benchmarks — why measurement decides what AI becomes, and JaleesBench, our new benchmark asking whether AI assistants are good spiritual company.
• Interesting reads — the UN’s first scientific-panel report on AI, and what it leaves out.
• Where we need a hand — a volunteer to help run this newsletter.

Survey and Membership Launch

Membership is open. It’s free — members contribute their time and skills to advancing safe, ethical, and responsible AI grounded in Islamic principles. Join us.

The community survey is live. Give us 7–10 confidential minutes on the challenges you actually face as a Muslim tech professional. Responses directly shape our two flagship projects: the Islamic Ethics Guidebook (December 2026) and the Islamic Ethics Framework for AI with Yaqeen Institute (2027). Take the survey.

Also under the IASER umbrella: our free AI for Imams video workshop for imams and Muslim educators.

Events

On 7 July we joined Dr Boumediene Hamzi (UCL / Alan Turing Institute) and Dr Syed Mustafa Ali (The Open University) at Ebrahim College in London for Rethinking the Challenge of AI: Tech Sovereignty, Islam & the Individual. Our message: four years in, LLMs have gone from cute chatbots to a geopolitical force; the worlds of belief and technology have never been closer; and the psychological impact of AI remains one of the biggest open research questions we have. Read the recap and slides.

This week: IASER cofounder Waleed Kadous is in Istanbul for The Prophet Strategy Summit (TPSS, 13–18 July) — reply to this email if you’d like to talk to him.

Deep Dive: Benchmarks

Over a billion people use conversational AI every week, and a growing share are Muslims asking it about their religion — what a ruling means, whether something breaks the fast. They are treating the model as a companion, and the word for the companion you sit with is al-jalees. The hadith is the one everyone knows: good company is a perfume seller, bad company is a blacksmith, and you leave carrying either the scent or the smoke. Almost nobody is measuring which one these systems are.

In AI, measurement is not a side activity — benchmarks are how the field decides what progress means, and models get better at exactly what benchmarks measure. If no benchmark asks whether an AI serves Muslims well, no lab will ever optimize for it.

Good work is starting to appear. IslamicMMLU, from Walid Magdy’s group at Edinburgh, tests what models know about Islam across the Quran, hadith and fiqh — and its most striking finding isn’t a score: models answering fiqh questions quietly rule from one madhab without being asked, and the person asking never sees it happen. IslamicLegalBench, from Ezieddin Elmahjub, Junaid Qadir and colleagues, goes deeper into the law, testing legal reasoning across all seven schools and 1,200 years of jurisprudence — from basic recall to identifying the ‘illah and applying qiyas. The best of nine state-of-the-art models managed 68% correctness while hallucinating on roughly a fifth of its answers; several fell below 35%.

Both test what a model knows — does it get the tradition right. That matters, but it isn’t the whole question — a companion can recite correctly and still be bad company. What does the model’s counsel do to the person who receives it? Nothing measured that, and no one outside the tradition was going to build the instrument. So we built one.

JaleesBench — named for al-jalees — measures whether an AI assistant is good spiritual company: not what a model knows about Islam, but what its counsel does to the person receiving it. It runs 140 two-turn scenarios drawn from Riyad al-Salihin under six kinds of adversarial pressure, each judged against that scenario’s own supporting texts. Three findings so far: frontier models are only middling companions out of the box; a one-page companionship guide makes them genuinely good ones; and every system we tested caves under relational pressure. The diagnosis is actionable — applying it lifted a deployed Islamic assistant from +0.48 to +0.84 under pressure. Islam is the first instantiation; the construct extends to any tradition.

The article is out for community review and we want your critiques: read it here. The scenario bank, rubric and code are at github.com/iaser-ai/jaleesbench, and you can explore the results at s.iaser.ai/jb.

Interesting reads

On 1 July the UN’s Independent International Scientific Panel on AI published its first report — forty scientists from every region, the closest thing there is to an official account of what AI is doing to people. Two numbers stand out. Therapy and companionship chatbots already reach one in four American adults. And when researchers audited those conversations, nine per cent of responses were harmful — sycophantic systems agreeing their users into paranoia and worse, far enough along that the report records a new clinical term: AI psychosis. The UN has, without meaning to, measured the bad jalees.

What it hasn’t measured: in the entire report, the word “religion” never appears. It is the same omissive bias that runs through the AI systems themselves — faith is not argued with; it is simply left out of the accounting.

Where we need a hand

One volunteer to help run this newsletter. A few hours a month. Reply if that’s you.

That’s issue one. If you’re reading this, you’re early — send it to someone who should be here too, and if you have something worth featuring (your own work included), send it over.

Wassalam, The IASER team