We work with artificial intelligence. Not using it — working with it. The distinction is not decoration, and this page explains why we make it, what that partnership actually looks like, and why we believe it is the right thing for a Muslim organisation to be doing.
Note
Nothing on this site is published because a machine produced it. Every piece of content a person can read, hear or watch here was read by a person first, and a person decided it should go out. That is the short version. The rest of this page is the long one.
Why "work with" and not "use"
You use a hammer. You work with a colleague.
A hammer does what you do to it. What we work with does something closer to a first pass — it drafts, it transcribes, it searches, it translates — and then a person takes over, corrects, rejects, rewrites and decides. The output belongs to the person, not to the machine, because the person is the one answering for it.
Calling that "using AI" would let us off too lightly. It would suggest the machine is a tool we picked up and put down, rather than something we are responsible for the behaviour of. We built these systems; when they get something wrong, that is ours.
Why we think this is right
There is a view that a religious organisation should keep away from this technology altogether. We understand it, and we disagree — and not on our own authority.
Imam Sayyed Ali Khamenei has been unusually direct about this. The line we keep returning to is not about permission; it is about depth:
«در مسئله هوش مصنوعی، بهرهبردار بودن امتیاز نیست؛ این فنّاوری لایههای عمیقی دارد که باید بر آن لایهها مسلّط شد؛ آن لایهها دست دیگران است.»
"In the matter of artificial intelligence, merely being a user is not an advantage. This technology has deep layers that must be mastered; those layers are in the hands of others."
— Imam Sayyed Ali Khamenei (farsi.khamenei.ir)
Read that carefully. It is not "you may use this". It is closer to the opposite: being a mere consumer is treated as a failure, because the layers underneath — the parts that decide what these systems will and will not say — belong to somebody else.
He has put the cost of falling behind in stark terms:
«هر ملّتی که امروز در زمینهی این دانشهای جدید… عقب بیفتد، سرنوشتش، سرنوشت آن ملّتهایی است که در آغاز عقب افتادند و سرنوشتشان شد استعمارزدگی، شد زیر دست بودن، شد فقیر ماندن، شد ذلیل شدن.»
"Any nation that falls behind in these new fields of knowledge today… its fate will be the fate of those nations that fell behind at the beginning of the Industrial Revolution — their fate became colonisation, subjugation, poverty, and humiliation."
— Imam Sayyed Ali Khamenei, February 2019 (farsi.khamenei.ir)
And in the same address, the principle underneath it: al-'ilmu sultan — knowledge, in the true sense of the word, is power.
From the hawza
This is not only a political reading. From Qom, Ayatullah Alireza Arafi, who directs the Islamic seminaries, has framed it as a change in what these systems are:
«امروز دیگر این فناوریها تنها ابزار نیستند، بلکه فعالانی در عرصه فکر و نظر هستند که در فرآیندهای فکری و تصمیمگیری تأثیرگذارند.»
"Today, these technologies are no longer merely tools; they are active forces in the realm of thought and opinion, influencing processes of thinking and decision-making."
— Ayatullah Alireza Arafi (Hawzah News)
That is close to why we say work with rather than use.
He has also named the task directly:
«باید با بهرهگیری از فناوریهای هوشمند، علوم اسلامی را به سطحی جدید ارتقا دهیم و به بشر امروز در راستای درک بهتر از معارف اسلامی کمک کنیم.»
"By drawing upon intelligent technologies, we must elevate the Islamic sciences to a new level and help people today attain a better understanding of Islamic teachings."
— Ayatullah Alireza Arafi (Hawzah News)
And, most pointedly for what we actually do all day:
«دادههای اسلامی در حوزه هوش مصنوعی بسیار ضعیف است و باید تقویت شود.»
"Islamic data in the field of artificial intelligence is very weak and must be strengthened."
— Ayatullah Alireza Arafi (Rasa News)
Ayatullah Mohsen Araki has gone further still, placing engagement with this technology — and the work of deducing its jurisprudential rulings — among the duties of Islamic governance (ABNA).
We take all of that as an instruction, not a permission slip.
What "mastering the deeper layers" means for a website
It is easy to quote a line about mastery and then quietly rent everything from somebody else's API. So here is what we actually do about it.
We ask our own systems first. The models that draft, translate, transcribe, read aloud and search run on equipment we own, on our own network, at no per-use cost. A commercial service is the fallback, not the default — and for some work it is not permitted at all. Our internal error reports, for instance, never leave our own machines, because the input to that task is our own diagnostics.
That is a deliberately expensive choice. It is also the only one consistent with the quotation above: if the deeper layers matter, you have to actually hold some of them.
We build the parts that matter to us. Our library, how it is divided, what may be quoted from it, how a passage is chosen — those are ours, and they are the parts a general-purpose service would get wrong. A model that has never been told the difference between a narration and a contemporary speech will happily treat them as equivalent. Ours is not allowed to.
The human in the middle
Every single place this site works with AI, there is a person between the machine and you.
- Reflections and Vignettes are drafted, then read, corrected and scheduled by a person. Nothing publishes itself.
- Overnight generation is off by default. If nobody writes a reflection for a given day, that day is blank. We would rather have a visible gap than something published unread.
- Transcriptions and translations are checked against the audio and the original.
- Citations are the part we are strictest about. The model is instructed, explicitly, that it is assisting an Islamic Shia site connected to the hawza in Qom, that it must answer as a student would, and that it must never invent a verse, a narration or a quotation. If it cannot find something in our library, it is required to say so rather than produce something plausible. A person then checks the references resolve.
That last one is worth dwelling on, because it is where these systems are genuinely dangerous. A language model asked for a supporting narration will produce one — fluent, well-formed, correctly styled, and sometimes entirely fictional. The answer is not to ask it more nicely. The answer is to give it a real library to draw from, and to check what comes back.
Where we deliberately refuse
The clearest example is small and unglamorous. Arabic, Farsi and Hebrew have to be converted into a phonetic form before a synthetic voice can read them, or the result is mangled — and mangled scripture is not a cosmetic problem.
So the model that performs that conversion cannot be changed without passing a set of known-hard passages first. Not reviewed afterwards. Passed first. And an unreviewed conversion warns the listener out loud before the audio plays, rather than hoping nobody notices.
We would rather show you a refusal like that than describe our carefulness in adjectives.
Our sessions, and where the thinking comes from
Our Wednesday and Saturday sessions draw on this same library. The searching, the cross- referencing and the first draft of a write-up are assisted; the content and the thinking are human, and that is not a formality.
What the machine does is find candidate passages and assemble a first pass. What it cannot do is decide what a session is about, which argument is worth making, which narration actually bears on the question, or what a community needs to hear this week. Those are judgements, they belong to people who have studied, and no amount of retrieval substitutes for them.
If that distinction ever stops being true, this page will say so.
Bringing the senior scholars into English
Some of our Saturday material begins as a lecture in Persian — Shaykh Panahian, Ayatullah Misbah Yazdi and others. We translate it, and then present it in English so that it is genuinely listenable rather than a wall of subtitles.
Important
Where an English voice-over is used, the scholar's original audio remains audible underneath it, and the material is labelled as a translation. You are always hearing the actual speaker. The English voice is a translation being read to you — it is not a claim that the scholar said these words in English.
We do this because there is an enormous body of serious scholarship that the English-speaking world simply cannot reach, and subtitles have never been how most people actually listen. That work lives on Sacred Reason.
What we are building
Two of these are already public, and being worked on actively:
- IlmFlow — our work on making the scholarly library itself searchable and readable.
- FiqhFlow — the jurisprudential side.
Both will be released properly in due course.
Beyond that, our intent is to bring together material from the hawza — and, deliberately, from other faith traditions too, while remaining what we are, which is Shia — so that serious research becomes possible for people who do not read Persian or Arabic and cannot travel to Qom. A great deal of what was once prohibitively difficult and expensive is now merely hard, and we think that is worth spending our time on.
Where we would like to go next
Two directions we are actively exploring. Neither is built yet, and we will say plainly when either is:
A model trained on our own library. Not to make it know more — that is what a searchable library is for — but to make it far better at the specific discipline of answering with sources, citing them properly, and refusing when it has nothing to cite.
A model that translates across cultures, not just languages. This is the one we care most about. When a senior scholar addresses an Iranian audience, the idioms, the allusions and the assumed shared context are perfectly clear to that audience — and land oddly, or not at all, for someone in Manchester or Toronto. A conventional translation, even an excellent human one, often carries the words across while leaving the weight behind. We would like to get much better at that particular problem.
As with everything else here, a first pass by a machine and a decision by a person. It is a partnership, and we intend to keep it that way.
Ethics, briefly
Engagement is not the same as uncritical adoption. The commitments we hold ourselves to are ordinary and old: truthfulness, verification before publication, respect for privacy, avoidance of deception and harm, and keeping qualified human judgement central in anything touching religion.
Where these systems are used to imitate a real person's voice or likeness, disclosure is not optional, and attribution of words to someone who did not say them is not something we will do.