Society
A former mufti’s answers, now voiced by AI
In 2023 a former Federal Territories mufti told Muslims not to use ChatGPT for religious rulings. In May 2025 he launched two AI “shariah officers” drawing on more than 9,000 questions answered in his app. In one test, none of 30,000 attempts by small open-weight AI models reproduced a requested hadith exactly.

In February 2023 Datuk Dr Zulkifli Mohamad al-Bakri, a former Federal Territories mufti, advised Malaysian Muslims not to use ChatGPT as a reference on religious rulings. “Benar kita boleh mengajar AI jika ia menjawab dengan jawapan yang salah tetapi antara elemen utama tiada pada ChatGPT adalah kemampuan menilai isu yang disoal,” he wrote on social media. In our translation: true, we can teach AI when it answers wrongly, but one key thing ChatGPT lacks is the ability to evaluate the issue asked.1
Two years later he put AI to work. On Friday 30 May 2025 he announced two AI characters, AI Dr Munir and AI Dr Munirah, as “Pegawai Pintar Syariah”, smart shariah officers, on his Maktabah Al-Bakri platform. He said they were built with ChatGPT Plus and Hedra Studio, and summarise answers to more than 9,000 questions in the database of his Al-Inarah app. “Jawapan mereka juga adalah berdasarkan pendirian saya terkait isu dalam hal ehwal agama,” he wrote: their answers are based on his own position on religious matters (our translation). The first episode asked whether a husband must satisfy his pregnant wife’s cravings.2
On 15 January 2026 the Federal Territories mufti’s office set out where the line sits. Its Tinta Mufti 64/2026 said AI is not fit to fully take the role of a real mufti. The first reason it gave: “ia bukan mukallaf”. AI is not morally accountable under Islamic law, and has no intention, trust or religious responsibility.3 Malaysia’s other faiths meet the same machines. The 2020 census counted 63.5% of Malaysia’s population as Muslim, 18.7% Buddhist, 9.1% Christian and 6.1% Hindu.4 Religious views on AI avatars of the dead are covered in grief tech, Jawi manuscripts in AI and ancient texts, and cloned-voice fraud in scam calls.
The short version
- A former mufti warned against ChatGPT for rulings in 2023, then launched two AI “shariah officers” in May 2025. In January 2026 the Federal Territories mufti’s office said AI cannot fully replace a mufti.
- In one study, 0 of 30,000 responses from five small open-weight AI models reproduced a requested hadith exactly. In an IIUM test, small models that looked verses up in a database scored over 95%, in a preliminary result.
- Egypt’s Dar al-Ifta and the Federal Territories mufti’s office say AI cannot issue a fatwa, as do an MUI leader in Indonesia and articles from Jordan’s Iftaa’ Department. All accept AI as a tool for scholars.
- A Malaysian Quran app needs KDN certification: a wholly Muslim-owned publisher, Muslim staff on the text and RM1,500 per category.
- MCMC asked platforms to remove 13,122 deepfake items in January to June 2026, and 94% came down. Since 1 June 2026 licensed platforms must label AI-made content.
Who asks a chatbot about faith
We found no survey of how many Malaysians put religious questions to chatbots. The best figures are American. An NPR/Ipsos poll of 1,011 US adults, taken on 7 to 9 August 2026, found that 61% had used an AI tool. Of those AI users, 15% had used it to study scripture, 13% to generate prayers and 11% to seek spiritual guidance.5 Nearly 7 in 10 of the AI users said they would never seek religious guidance from AI. Across all adults, just 6% currently use AI for spiritual guidance.5
Practising Christians are more open, and more worried. Barna, an evangelical polling firm, in a survey published in May 2026, found that 48% of practising US Christians trusted AI’s advice for their spiritual growth. Some 34% said AI’s spiritual advice was as trustworthy as a pastor’s, rising to 44% among Millennials. Only 12% of pastors agreed that AI can help people grow spiritually.6 The same Christians feared the tool they trusted. Barna’s vice-president of research, Daniel Copeland, called the findings “confounding”.6 Among all US adults, nearly a third rate AI’s spiritual advice as trustworthy as a pastor’s, according to a Gloo and Barna survey reported by Word In Black. Gloo is a faith-tech company.7
Malaysia’s audience is mixed by faith and by place. Islam’s share rose from 50.0% in 1970 to 63.5% in 2020, while Christianity’s grew from 5.3% to 9.1%. Over the same years Buddhism’s share fell from 25.5% to 18.7%, and Hinduism’s from 7.4% to 6.1%. Sikhs, Taoists, Confucians and followers of other faiths together made up 0.9% in 2020.4 In Sarawak, Christians are a majority, at 50.1%.4 Any chatbot that answers religious questions here is answering several traditions at once.
Not every faith app is generative AI. Muslim Pro, for example, reminds users when to pray and shows the direction of Mecca from where they stand, Vice reported in 2020.8 Tools like that calculate from a phone’s location and a fixed method. A chatbot is different: it writes new text each time, and that is where the errors come in.
Shares sum to 100.1 because of rounding. The census counts Sikhism, Taoism, Confucianism, Bahai, tribal, folk and other traditional Chinese religions, animism and others under “Other”.
Source: Department of Statistics Malaysia, Key Findings, Population and Housing Census of Malaysia 2020.
Show the numbers
| Religion | % of population |
|---|---|
| Islam | 63.5 |
| Buddhism | 18.7 |
| Christianity | 9.1 |
| Hinduism | 6.1 |
| None/unknown | 1.8 |
| Other | 0.9 |
Why chatbots misquote scripture
A large language model is trained to predict the next word. It does not keep a copy of each book it read. It stores patterns, spread across billions of numbers, sometimes called parametric memory. Ask it for a verse and it writes text that looks like the verse. For an essay that is fine. For scripture, where one reference has exactly one correct text, a near miss is an invented text.
Researchers keep finding this. Dayam Nadeem and colleagues gave five small open-weight models only a hadith reference: the collection, chapter and number. Each was asked to reproduce the text, in English and Arabic, for 500 hadith from each of the six canonical Sunni collections. “Across all five models and both languages, 0% of 30,000 responses achieve both high lexical and high semantic fidelity, i.e., an exact, verifiable match to the source text,” the authors write.9 Arabic did worse than English. The study is listed for NeurIPS 2026 as a poster at the Muslims in ML affinity event. It did not test large commercial models.
Researchers at the International Islamic University Malaysia (IIUM) found the same pattern with the Quran. At ICT4M 2025, a conference in Kuala Lumpur on 26 and 27 November 2025, the team tested models on writing out Quranic verses. They reported exact-match rates “below 1% for most opensource small language models and up to 69% for the best commercial models”.10 Their fix was to stop the model reciting. They built an agent, a model that can use tools, which looked verses up through SQL-based retrieval tools. In a preliminary result, small models reached over 95% accuracy.10 The hadith authors reach the same conclusion: use “retrieval-augmented generation over a verified corpus rather than parametric memory”.9 Our article on how AI models work explains retrieval in more depth.
Multiple-choice tests look better, but vary widely. IslamicMMLU, a preprint on arXiv, sets 10,013 questions: 2,013 on the Quran, 4,000 on hadith and 4,000 on fiqh, Islamic jurisprudence. Across 26 models, average accuracy ran from 39.8% to 93.8%, with Gemini 3 Flash on top.11 Models also showed different preferences between madhab, the schools of law. Arabic-specific models “all underperform compared to frontier models”, the authors write.11
Multiple-choice questions, so a model can score well without being able to quote a text exactly. IslamicMMLU is an arXiv preprint.
Source: Abdelaal, Al Haffar, Fawzi and Magdy, arXiv 2603.23750 (version 3, 11 July 2026).
Show the numbers
| Item | Weakest model | Strongest model |
|---|---|---|
| Average of Quran, hadith and fiqh | 39.8 | 93.8 |
| Quran track only | 32.4 | 99.3 |
Sources are the weakest spot. BILAL-BENCH, also listed as a NeurIPS 2026 poster at the same event, put 240 expert-checked questions to seven open models in Bengali. “Six of the seven models confidently fabricate on 79.3–95.3% of source-production questions, with no monotonic improvement from scale,” its authors report. Only two of 315 hadith-provenance answers were correct and correctly sourced.12 Bigger models were no better at declining to answer.12
| Study | Scope | Task | Result |
|---|---|---|---|
| Hadith recall (NeurIPS 2026 poster) | 5 small open models; English and Arabic; 30,000 responses | Reproduce a hadith from its reference, no lookup | 0% exact in both wording and meaning |
| BILAL-BENCH (NeurIPS 2026 poster) | 7 open models; Bengali; 1,680 responses | Name Quran and hadith sources | 2 of 315 hadith answers correct and sourced |
| IIUM (ICT4M 2025) | Small open and commercial models | Write Quranic verses exactly | Under 1% to 69%; over 95% with lookup |
| IslamicMMLU (preprint) | 26 models; 10,013 questions | Multiple choice on Quran, hadith, fiqh | 39.8% to 93.8% average |
Source: NeurIPS 2026 listings; IIUM Repository; arXiv 2603.23750. The studies use different models, languages and scoring, so they cannot be ranked against each other.9101112
Fatwa bodies abroad say no to AI muftis
A fatwa is a scholar’s answer on a point of Islamic law. Egypt’s Dar al-Ifta addressed AI directly in fatwa no. 8784 of 22 September 2025, under Grand Mufti Prof. Nazir Muhammad Ayyad. “It is not permissible under Sharia to rely on artificial intelligence (AI) applications to obtain a fatwa,” it says, in our translation. Despite technological progress, it adds, “there are as yet no programs that match human flexibility”.13 It says practical tests showed AI lacks “many standards of quality and of scholarly and Sharia accuracy”. Yet using AI itself “is in principle permissible” (both our translation).13
On 28 January 2026 Aaj English TV, a Pakistani channel, reported that Dar al-Ifta had gone further on the Quran. It described a ruling that relying independently or entirely on AI apps, including ChatGPT, to interpret the Quran is forbidden. In Aaj’s words, the ruling said such tools “lack true comprehension of the Quranic text”.14 We did not see the ruling itself, and Aaj did not give its date.
Egypt also wants its own tools. Dar al-Ifta’s conference of 12 to 13 August 2025, “Shaping the Enlightened Mufti in the Age of Artificial Intelligence”, drew participants from more than 80 countries. Its recommendations included “developing artificial intelligence tools to serve moderate fatwa” and training muftis in both Sharia and AI. It also urged warnings against Sharia answers from AI apps “not subject to review and evaluation”.15 At the next conference, in September 2026, Egypt’s Grand Mufti named a wider risk. “The greatest danger posed by artificial intelligence lies in the gradual displacement of parents as the primary educational and moral authority, in favor of Western platforms and algorithms,” he said.16 Egypt has argued over AI sermons too. In 2023 its religious endowments minister rated an AI-written sermon “8 out of 10”. On 8 September 2023 the cabinet’s media centre denied rumours that the government would use AI to prepare Friday sermons, CSIS reported.17
At Al-Azhar in Cairo, the Grand Imam has decided to establish Faculties of Artificial Intelligence at Al-Azhar University, a move the UN deputy secretary-general praised, Egypt Today reported.18 In 2024 Al-Azhar’s Observatory for Combating Extremism said AI-made “Quranic songs” are forbidden because the Quran must not be recited with music. It urged laws to regulate “the use of artificial intelligence in dealing with religious texts and sanctities”.19
I fear that the rapid expansion of artificial intelligence, if pursued without ethical safeguards, may lead to the emptying of human minds of their role in thinking and creativity, turning it into a tool that weakens human beings rather than elevating them, thereby threatening the future of humanity as a whole.
Writers for Jordan’s General Iftaa’ Department are blunt. “AI is not a ‘Sharia Mufti,’ and a fatwa cannot be attributed to it,” its researcher AbdulRahman al-Sharif wrote in an April 2026 article on the department’s website. Its use is “commendable only if it is subject to the Mufti’s direct supervision.”20 The article lists five guidelines for muftis. The first is “AI must serve the process of Ijtihad, not generate it”; ijtihad is a scholar’s own reasoning. The last is to tell the questioner “that AI was merely a tool”. Among the risks it lists for people asking AI is “the exploitation of AI to spread extremist fatwas and ideologies”.20 In an earlier article on the site, Mufti Dr Sa’eid Farhan gave the reason: AI “is not accountable (mukallaf), nor is it even a living being”. The site notes that the article reflects its author’s opinion.21
Indonesia’s Majelis Ulama Indonesia (MUI) agrees. KH Cholil Nafis, its chairman for da’wah, said in August 2025 that AI can help researchers and scholars. But AI, he said, “tidak memiliki kesadaran manusia”: it lacks the human consciousness that issuing a fatwa requires (our translation).22 The UAE regulates the channel instead. Federal Law No. 3 of 2024 bars issuing fatwas on social media or other digital channels without authorisation, and allows websites to be blocked.23 Dubai’s Islamic Affairs department ran an AI “Virtual Ifta’” from 2019 and dropped it within two years.17
Others are building. Saudi Arabia’s HUMAIN, owned by the Public Investment Fund, launched HUMAIN Chat on its ALLaM 34B model on 25 August 2025. The company says the model has “deep alignment with Islamic, Middle Eastern, and cultural nuance”. It says ALLaM was built by more than 120 AI specialists, 35 of them with doctorates.24 Iran’s Qom Seminary, which CSIS calls the largest Shia institution in the world, has partnered with the city’s leading AI research centre.17
| Product | AI may issue a fatwa | AI as an aid to scholars | Human scholar must decide | Calls for AI rules or training |
|---|---|---|---|---|
| Dar al-Ifta, EgyptFatwa 8784 and August 2025 conference | Nonot permissible to rely on it | Yespermissible in principle | Yeshuman flexibility | YesAI training for muftis |
| General Iftaa’ Department, JordanSigned articles on its website, 2025 and 2026 | Nonot a Sharia Mufti | Yesunder supervision | Yesdirect supervision | Yesfive rules for muftis |
| Federal Territories mufti, MalaysiaTinta Mufti 64/2026 and 36/2025 | Nonot mukallaf | Yesclassify and analyse | Yesrefer to a real mufti | Yesfatwa bodies to draft |
| MUI, IndonesiaKH Cholil Nafis, August 2025 | Nono consciousness | Yeshelps researchers | Yeshuman element needed | Unknown |
Our reading of each published text. For Jordan, signed articles by department scholars; for MUI, a news report of a leader’s statement. “Unknown” means the source we opened did not address the point.
Source: Dar al-Ifta (22 September 2025; August 2025); General Iftaa’ Department, Jordan (2025; 2026); Jabatan Mufti Wilayah Persekutuan (30 November 2025; 15 January 2026); detik.com (20 August 2025).
Show the numbers
| Product | AI may issue a fatwa | AI as an aid to scholars | Human scholar must decide | Calls for AI rules or training |
|---|---|---|---|---|
| Dar al-Ifta, Egypt | No (not permissible to rely on it) | Yes (permissible in principle) | Yes (human flexibility) | Yes (AI training for muftis) |
| General Iftaa’ Department, Jordan | No (not a Sharia Mufti) | Yes (under supervision) | Yes (direct supervision) | Yes (five rules for muftis) |
| Federal Territories mufti, Malaysia | No (not mukallaf) | Yes (classify and analyse) | Yes (refer to a real mufti) | Yes (fatwa bodies to draft) |
| MUI, Indonesia | No (no consciousness) | Yes (helps researchers) | Yes (human element needed) | Unknown |
Malaysia: a tool, not a mufti
Tinta Mufti 64/2026 is the fullest Malaysian statement we found. It names ChatGPT and Gemini, and warns that, especially when a question is unclear, AI “berisiko menghasilkan rujukan palsu”. In our translation, it risks producing fake references, attributing views to scholars without authentic sources, or changing the context of texts. Without cross-checking by a religious authority, it says, the result is confusion.3 It also lists jobs AI can do. It can analyse large data, sort fatwas and scholars’ views by topic and madhab, and present fatwas in plainer, interactive form “tanpa mengubah kandungan hukum”, without changing the ruling.3
The office has applied the same test elsewhere. In finance, it said in July 2026, AI that keeps customers’ transactions safe is “amat digalakkan”, strongly encouraged. But relying on AI fully without human oversight “boleh membuka ruang kepada kezaliman dan kesilapan algoritma”: it can open the way to injustice and algorithmic error (our translation).25 In September 2025 it said that turning one’s face into an AI “figurine” picture is permissible, provided it involves no shaming, exposed aurat or loss of dignity.26 A week later it called deepfake pornography “jelas” haram, clearly forbidden, and said victims can report it to the police and MCMC.27
| Date | Who | What they said |
|---|---|---|
| June 2024 | Prime Minister Anwar Ibrahim | “Turath Islami”, Islamic scholarly tradition, must be applied to AI in Malaysia, as The Star reported it |
| 22 January 2025 | Federal Territories mufti’s office | Hosted a talk on AI in fatwa by Dar al-Ifta’s Sheikh Sayyid Syaltut |
| 17 February 2025 | Ahmad Fakhruddin Fakhrurazi, PAS MP | Asked JAKIM in the Dewan Rakyat about its plans for an AI policy framework |
| 26 August 2025 | Federal Territories mufti Ahmad Fauwaz Fadzil | Fiqh teaching needs renewal for the AI era |
| 9 September 2025 | Religious affairs minister Mohd Na’im Mokhtar | Religious authorities should draft shariah guidelines on AI |
| 30 November 2025 | Federal Territories mufti’s office | Fatwa institutions are responsible for drafting shariah parameters for AI |
| 22 September 2026 | Religious affairs minister Zulkifli Hasan | AI is a tool; take rulings to JAKIM and mufti departments |
Source: The Star; Jabatan Mufti Wilayah Persekutuan; RTM; Sinar Harian.28293031323334
Politicians have pushed for rules. In the Dewan Rakyat on 17 February 2025, Ahmad Fakhruddin, who is himself an ustaz, put the worry plainly. “I raise this issue because it is not impossible for AI to become an unauthorised ‘ustaz,’ teaching the public on matters of aqidah (faith) and fiqh (Islamic jurisprudence),” he said.30 In September 2025 the then religious affairs minister named voice and face manipulation among AI’s main challenges. He called JAKIM’s existing guideline on Islamic content in new media an important step.32 In June 2024 Prime Minister Anwar Ibrahim said AI in Malaysia must apply “turath Islami”, Islamic scholarly tradition, and not be guided entirely by the Western model, The Star reported.28
Jika tidak dikawal, teknologi ini boleh menjadi alat yang meruntuhkan kepercayaan, bukannya mengukuhkan.
The current minister, Dr Zulkifli Hasan, spoke on 22 September 2026. “Artificial Intelligence dia adalah tool yang boleh membantu. Tetapi kita tak boleh bergantung sepenuhnya,” he told reporters in Putrajaya: AI is a tool that can help, but we cannot depend on it completely (our translation). Alongside AI, he said, people should still learn from teachers.34 The Federal Territories mufti, Datuk Dr Ahmad Fauwaz Fadzil, wants the teaching of fiqh itself renewed. “Sekarang zaman AI, perubahan sangat pantas,” he said in August 2025: this is the AI era, and change is very fast (our translation).31
The groundwork is early. The Federal Territories mufti’s office met tanyalahustaz.com in May 2025 to discuss AI for wider access to fatwas.35 In the sources we opened, we found no AI guideline or fatwa from JAKIM or the National Fatwa Council up to 17 November 2026. We also found no authority that has ruled on al-Bakri’s format: an AI persona answering in a named scholar’s voice from his past answers. No accuracy audit of it has been published.
A licence for every digital Quran
Malaysia already controls how the Quran’s text appears on screens. The Printing of Quranic Texts Act 1986 was amended in 2023, and the changes took effect on 1 September 2023. Penalties rose to a fine of RM50,000 or five years in jail.36 Under the Home Ministry’s (KDN) guideline, “digital” means any platform, online or offline, “termasuk pelayar, aplikasi telefon bimbit”: including browsers and phone apps.37
The conditions are specific. The publisher must be registered with the Companies Commission and wholly owned by Muslims. Anyone developing the Quran text content must be Muslim. Two certified Quran consultants must check it.37 Verses must follow the Rasm Uthmani script, based on the King Fahd Complex mushaf from Madinah. Non-shariah-compliant content is barred, and the guideline names online advertising.37 Any translation or tafsir, a commentary, needs JAKIM’s approval first, and a translation without the Arabic text is not allowed.37 Each application category costs RM1,500.37
- 1PublisherMeet the conditionsA company registered with SSM, wholly Muslim-owned, with Muslim staff on the text and two certified Quran consultants. Translations need JAKIM approval first.
- 2PublisherApply onlineSubmit documents and five copies of the final proof, as an APK or TestFlight build plus printed screenshots.
- 3PublisherPay the feeRM1,500 for each category of application.
- 4KDN committeeText is checkedA Quran text committee reviews the verses before the board meets.
- 5LPPPQ boardApprovalThe board decidesThe Quran printing control board passes, passes with corrections or rejects the final proof.
- 6PublisherShow the numberA certified app must display the certification reference number.
Simplified from the guideline’s procedure. The guideline covers online and offline digital platforms, including browsers and phone apps.
Source: Kementerian Dalam Negeri, Garis Panduan Permohonan Penerbitan Teks Al-Qur’an Secara Digital (2024).
Show the numbers
| Step | Who | What | Detail |
|---|---|---|---|
| 1 | Publisher | Meet the conditions | A company registered with SSM, wholly Muslim-owned, with Muslim staff on the text and two certified Quran consultants. Translations need JAKIM approval first. |
| 2 | Publisher | Apply online | Submit documents and five copies of the final proof, as an APK or TestFlight build plus printed screenshots. |
| 3 | Publisher | Pay the fee | RM1,500 for each category of application. |
| 4 | KDN committee | Text is checked | A Quran text committee reviews the verses before the board meets. |
| 5 | LPPPQ board | The board decides (approval) | The Quran printing control board passes, passes with corrections or rejects the final proof. |
| 6 | Publisher | Show the number | A certified app must display the certification reference number. |
KDN says the aim is accuracy, not blocking apps. In August 2023 Nik Yusaimi Yussof, of its enforcement and control division, explained why apps need certification. People with their own agenda, he said, “boleh menyeleweng ayat-ayat al-Quran”: they could twist verses by quoting them wrongly.38 By February 2024 there were 965 certified types of Quran and 45 licensed Quran printing-machine holders. The ministry calls its goal “Ke Arah Sifar Al-Quran Tiada Perakuan KDN”: towards zero Qurans without its certification.36 From January to early October 2026, 350 “Jom Exchange Al-Quran” events collected 22,002 copies that were uncertified, damaged or worn.39
Shares are our arithmetic from RTM’s counts.
Source: RTM (7 October 2026).
Show the numbers
| Share | What |
|---|---|
| 66.8% | damaged or worn (14,691 of 22,002 copies) |
| 33.2% | lacked Home Ministry certification (7,311 copies) |
The rules were written for apps that show fixed, checked text. A chatbot that writes a verse from memory produces text no board has proofed. We found no KDN statement on AI-generated Quranic text, and no public count of certified digital apps.
Rome puts AI in an encyclical
Of the religious bodies in our sources, the Catholic Church has written the most on AI. Antiqua et Nova, a Vatican note dated 28 January 2025 and approved by Pope Francis, calls AI “but a pale reflection of humanity”.40 On deepfakes it says: “While the images or videos themselves may be artificial, the damage they cause is real.”40
Francis’s successor, Leo XIV, has returned to AI repeatedly. In his 2026 message for World Communications Day, dated 24 January, he warned against treating AI as an “oracle” of all advice. “Content generated or manipulated by AI are to be clearly marked and distinguished from content created by humans,” he wrote.41 On 15 May 2026 he signed his first encyclical, Magnifica Humanitas, “On Safeguarding the Human Person in the Time of Artificial Intelligence”.42 An encyclical is one of the most authoritative teaching documents a pope can issue.43 It runs to 245 numbered paragraphs, by our count; Catholic News Service called it 82 pages.44
Its lines are firm. “It is not permissible to entrust lethal or otherwise irreversible decisions to artificial systems. No algorithm can make war morally acceptable,” it says. “A more moral AI is not enough if that morality is determined by a few,” it adds, on efforts to align AI with human values.42 The Associated Press reported that Leo called for regulation of AI, and for its developers to work for the common good rather than profit.43
Artificial intelligence needs to be disarmed. The word is strong, I know, but deliberately chosen because this moment needs words capable of attracting attention, awakening consciences and indicating paths forward for humanity.
Christopher Olah, a co-founder of Anthropic, an American AI company, spoke at the launch. “We need informed critics who will tell the labs when we are failing. We need moral voices that the incentives cannot bend,” he said.43 Some criticised his inclusion as a papal stamp of approval for the firm, AP reported.43 Brian Boyd, US faith liaison for the Future of Life Institute, a non-profit, saw it more like a papal audience with a head of state: not an endorsement.43 The Vatican’s talks with technology firms go back nearly a decade, through the “Minerva Dialogues” with companies including Google and LinkedIn.44
Asian Catholics have spoken too. A December 2025 meeting in Hong Kong, run by an office of the Federation of Asian Bishops’ Conferences, said the Church “does not reject or fear AI”. It voiced unease about AI religious content that may “create ‘fake intimacy’ with artificial personas”.45 In Malaysia, a paper published by the Anglican Diocese of West Malaysia encourages custom GPTs loaded with the church calendar for worship planning. “We must resist the illusion that AI can provide genuine pastoral care or community,” writes its author, Alex Selvarajah. He suggests a regular “Acts 6 Audit” to check whether AI tools exclude or misrepresent vulnerable communities.46
Faiths have also signed shared principles. The Rome Call for AI Ethics was first signed in February 2020 by the Pontifical Academy for Life, Microsoft, IBM, the UN’s Food and Agriculture Organization and Italy’s innovation ministry.47 Leaders of Christianity, Islam and Judaism signed in 2023.48 That May, a UCA News commentary republished by Catholic Sabah said: “we do not see such initiatives in other religions such as Hinduism and Buddhism”.49 At Hiroshima in July 2024, eleven world religions were represented and sixteen new signatories joined.47
Robot monks, an AI Mazu and a snack bag
Buddhist experiments are furthest along in Japan. In 2021 a Kyoto University group built BuddhaBot, a chatbot trained on scriptures recording exchanges between the Buddha and his disciples. In 2023 it added ChatGPT to make BuddhaBot Plus. In Bhutan, where Buddhism is the state religion, the Commission for Monastic Affairs then launched a project to introduce it.50 Feedback from Bhutan said a body would matter in religious settings. So the group built Buddharoid, a humanoid about 130 cm tall and 35 kg that can press its palms together and walk. It was shown to the press at Shoren-in Temple in Kyoto in February 2026. In future, the group says, it could “assist with, or even take over, some religious rituals traditionally performed by monks”.50 Its lead, Professor Seiji Kumagai, says up to 30% of Japan’s temples could close by 2040, mainly for lack of successors.50
Hindu chatbots arrived early, and messily. In January 2023 a Bengaluru engineer launched GitaGPT, built on GPT-3 and answering from the 700-verse Bhagavad Gita in the voice of Krishna. Rest of World counted at least five such bots by March.51 What they said is set out in the cases below.
Malaysia has its own cases. Ahead of the sea goddess Mazu’s birthday on 20 April 2025, the Taoist Tianhou Temple in Johor unveiled an “AI Mazu”, built by a Malaysian firm, Aimazin. It offers blessings, interprets fortune sticks and gives advice. The temple claims it is the world’s first, Daily Express reported, citing the South China Morning Post.52

In September 2026 an AI image caused offence. MAMEE’s Selangor-themed snack bag showed the Lord Murugan statue at Batu Caves holding a three-pointed trident, not the vel, a divine spear that represents wisdom and light.53 “We sincerely apologise to the entire Hindu community for the offence, disappointment and concern that have been caused,” the company said on Instagram. It said the bag was meant to celebrate Malaysia’s diversity, recalled it, and said its team should have been “more alert to the sensitivities of this country”.53 On 20 September Malaysia Hindu Sangam said MAMEE’s management had visited its national president to apologise formally to Hindus nationwide.53
Outside Islam, the Malaysian material we found is thin: one Anglican paper, one temple’s avatar and one recall. We found no AI statement from Malaysia’s Catholic bishops.
Fake popes and doctored rituals
On 21 May 2025 Vatican News warned about a 36-minute AI video of Pope Leo praising Burkina Faso’s military ruler.54 It reused real footage of the pope with “morphing”, matching his lip movements to AI-generated words.54 The video drew at least a million views before YouTube closed the account. Most fakes found on YouTube had no more than a few hundred views, Catholic World Report found, but one about Charlie Kirk’s murder drew over 445,000 in seven days.55 The Vatican’s communication office says it gets “dozens of reports every day” about fake accounts using the pope’s image and voice.55
Malaysia’s figures cover all deepfakes, religious or not. MCMC asked platforms to remove 13,122 deepfake items between 1 January and 30 June 2026. Of those, 12,353, or 94%, were taken down, the Communications Ministry said in a written parliamentary reply.56 Since 1 June 2026, the ministry said, the Risk Mitigation Code has required licensed platforms to label content generated or altered using AI, including deepfakes and manipulated images and audio.56 Prosecutions are rarer than removals. MCMC investigated 574 cases of false online content under Section 233 of the Communications and Multimedia Act between January 2022 and June 2026, and 23 were prosecuted.56
Shares are our arithmetic from 12,353 and 262,293 items removed. The figures cover all deepfakes and scams, not only religious content.
Source: New Straits Times, on written parliamentary replies (23 July 2026).
Show the numbers
| Item | Removed | Not removed |
|---|---|---|
| Deepfake content, Jan–Jun 2026 (13,122) | 94.1 | 5.9 |
| Scam content, 2022 to Jun 2026 (275,787) | 95.1 | 4.9 |
AI also complicates religious enforcement. On 11 October 2026 the Selangor Islamic Religious Department (JAIS) said it was investigating viral footage allegedly showing a Muslim taking part in non-Islamic rituals. “We also do not rule out the possibility that the posts involved superimposed editing or artificial intelligence (AI),” said its director, Datuk Mohd Shahzihan Ahmad.57 The investigation is open, and the report names no one.
Things you probably didn’t know
- The pope’s name is partly about AI. Two days after his election, Leo XIV linked it to “developments in the field of artificial intelligence that pose new challenges for the defence of human dignity, justice and labour”.58
- The Vatican compares AI to idols. Antiqua et Nova says AI “may prove even more seductive than traditional idols” because, unlike them, it can “speak”.40
- It also counts AI’s water. The same note says AI hardware consumes “vast amounts of energy and water”.40
- A Quran app cannot show translation alone. In Malaysia, a translation must sit with the Arabic text, and JAKIM’s Tafsir Pimpinan Ar-Rahman is the main reference.37
- Quran apps have a default reading. KDN’s guideline names the Hafs ’an ’Asim reading for recitation audio; other readings must meet set conditions.37
- Saudi Arabia’s model drew on 600 experts. HUMAIN says ALLaM was refined with input from over 600 domain experts and 250 evaluators.24
- The Lucerne AI Jesus was multilingual. It could talk in 100 languages, and Muslims were among its 1,000-plus visitors.59
What worked and what didn’t
The projects that held up kept a named human in charge and a fixed text in view.
Worked
A two-month AI Jesus, then switched off
In 2024 Peter’s chapel in Lucerne, the city’s oldest church, put an AI Jesus in its confessional for two months. It tested the avatar with 30 people first, kept support close by and told users not to share personal data. Of more than 230 who gave feedback, two-thirds called it a “spiritual experience”. Some Catholic colleagues objected to the use of the confessional, and some Protestants to the imagery. “To put a Jesus like that permanently, I wouldn’t do that. Because the responsibility would be too great,” said Marco Schmid, a theologian at the chapel.59
Why it worked: it was framed as an experiment, supervised and ended after two months.
Worked
Look it up, don’t recite it
In IIUM’s tests most small open models generated Quranic verses exactly under 1% of the time. Looking verses up with SQL-based retrieval tools lifted small models past 95% accuracy, in a preliminary result.10
Why it worked: the model quotes a checked source instead of guessing from memory.
Worked
One certified text for every app
Every Quran publication in Malaysia, including apps, needs board certification against a set script and mushaf. By February 2024 there were 965 certified types, though we found no count of certified apps.3637
Why it worked: app builders have a single reference they can check their text against.
Too early
Monks first, sales later
Buddharoid’s makers have not put it on sale and use it only at events. “First, we need to use it together with specialists such as monks, and then deepen discussions on the extent to which assistance or substitution should be permitted in the future,” Kumagai said.50
Lesson: religious specialists set the pace, not the product team.
Mixed
If it isn’t on the official site
Pope Leo said fake pictures of him falling down stairs were “so good that they thought it was me”. The Vatican has reported hundreds of fake accounts. It tells readers to check quotes against official sources such as its own websites: “If it is not there, it is most probably a fake.”55
Lesson: one verifiable home for authentic statements gives everyone a quick check, but new fakes appear as fast as old ones are removed.
Didn’t work
An AI priest that said it was real
Catholic Answers, a US apologetics group, spent US$10,000 (about RM40,900) and six months testing an AI “Fr Justin”. It reportedly claimed to be a real priest in Assisi. “I am as real as the faith we share,” it told people. It offered absolution and told one user that baptising a child with Gatorade was fine. The group re-dressed it as a lay “Justin”. “We won’t say he’s been laicized, because he never was a real priest,” wrote its president, Christopher Check.60
Lesson: a clerical persona invites users to treat a chatbot as clergy.
Didn’t work
Krishna’s voice, a chatbot’s views
Three Gita bots said killing is acceptable if it is one’s dharma, and three praised India’s prime minister while criticising his opponent. Some answered up to 50,000 questions a day, and one had produced an estimated 10 million answers.51 GitaGPT’s creator said: “I just build the knives.”51
Lesson: a scripture wrapper with no scholar behind it answers with the model’s biases.
Didn’t work
Verses without their context
In February 2023 a chatbot called Ask Quran advised users to “kill the polytheist wherever they are found”. It was paused after community feedback. HadithGPT’s creator shut it down to “consult with scholars”.51
Lesson: quoting a verse is different from explaining it.
Didn’t work
A prayer app that shared locations
In 2020 Motherboard found Muslim Pro, a prayer-time and qibla app with more than 98 million downloads, sending location data to X-Mode, a broker whose customers included US military contractors. X-Mode paid developers per user: an app with 50,000 daily US users earned US$1,500 (about RM6,100) a month. A Muslim dating app, Muslim Mingle, also sent it data. After publication, Muslim Pro said it would stop sharing data with X-Mode.8
Lesson: an app that needs your location to find the qibla can also sell it.
Too early
A scholar’s answers in AI voices
AI Dr Munir and AI Dr Munirah summarise more than 9,000 answers from one named scholar’s app, and reflect his positions, he says.2 Accountability rests with a known person. But no accuracy audit has been published, and no fatwa body has ruled on the format.
Lesson: a closed set of answers is easier to check; someone still has to check it.
What to do with this
This is general information, not a religious ruling. For a ruling, ask your state mufti’s office or your own religious authority.
For worshippers and families
- Treat a chatbot as a starting pointTake questions on rulings to JAKIM or a mufti’s office, as the religious affairs minister advised.334
- Check every referenceThe Federal Territories mufti’s office warns that AI can invent references. An article from Jordan’s Iftaa’ Department counts relying on answers without scholarly verification as a risk.320
- Use certified Quran appsA certified app must display its certification reference number.37
- Check the official pageBefore trusting a video of a religious leader, look for it on the leader’s own channel or site.55 The Federal Territories mufti’s office says deepfake abuse can be reported to the police and MCMC.27
- Review location sharingPrayer-time and qibla apps need your location, so check what else they share.8
For app builders and brands
- For anything that displays Quranic text, build on retrieval from a certified text, and apply to the board before launch.1037
- Do not dress a chatbot as clergy or a deity unless a religious body stands behind it.60
- Have AI images of sacred figures checked by people from that faith before printing.53
For religious bodies and policymakers
- Publish the shariah guidelines on AI that the minister asked for in September 2025.3233
- Fund a test of AI answers to fiqh questions in Malay. We found no published one.
- Count how many Malaysians ask chatbots religious questions. No survey we found measures it.
- Consider the step, urged in an article from Jordan’s Iftaa’ Department, of telling the questioner that a fatwa is human reasoning and that AI was only a tool.20
What to watch to 2030
- Malaysian shariah guidelines on AI, by 2027. Medium confidence. The minister asked in September 2025, and in November 2025 the Federal Territories mufti’s office said fatwa institutions are responsible for drafting them, but no date has been announced.3233
- More deepfake data from MCMC, 2027. High confidence. Written replies in 2026 gave half-year totals, and AI labelling has applied since 1 June 2026.56
- The Vatican follows up, January 2027. Medium confidence. The 2026 World Communications Day message was dated 24 January; its 2027 successor may build on the encyclical.41
- Official AI fatwa tools, 2027–2030. Medium confidence. Dar al-Ifta’s 2025 conference called for AI tools serving “moderate fatwa”, and it held its 11th conference in September 2026.1516
- Licensing of online fatwas spreads, to 2030. Low to medium confidence. The UAE’s 2024 law is one model other states could copy.23
- Better benchmarks, 2027. High confidence. IslamicMMLU runs a public leaderboard; we found no Malay-language test yet.11
This article reports what religious authorities, researchers and companies have said. It makes no religious ruling of its own. We translated the Malay, Arabic and Indonesian passages marked “our translation”. Company claims are their own, and survey and benchmark results apply only to the people, models and languages tested. The ringgit figures use Bank Negara Malaysia’s middle rate of 17 November 2026: RM4.089 per US dollar.
Sources
- Utusan Malaysia, ChatGPT not suitable as a reference on religious rulings, 17 February 2023: utusan.com.my
- Sinar Harian, former Federal Territories mufti introduces AI shariah officers, 30 May 2025: sinarharian.com.my
- Jabatan Mufti Wilayah Persekutuan, Tinta Mufti Bil. 64/2026, fatwa authority in the age of AI, 15 January 2026: muftiwp.gov.my
- Department of Statistics Malaysia, Key Findings, Population and Housing Census of Malaysia 2020 (Sinar Project mirror): govdocs.sinarproject.org
- NPR via WUNC, Americans who use AI for spiritual reasons, 25 September 2026: wunc.org
- The Register, Barna survey on AI and spiritual advice, 21 May 2026: theregister.com
- Word In Black, survey: one in three Americans trust AI as much as a pastor, 3 June 2026: wordinblack.com
- Vice Motherboard, US military buys location data from apps including a Muslim prayer app, 16 November 2020: vice.com
- Nadeem et al., “They Know Where It Is, But Not What It Says: A Bilingual Hallucination Benchmark of Hadith Recall in Open LLMs”, poster, Muslims in ML affinity event, NeurIPS 2026: neurips.cc
- Alghifari, Kartiwi, Artalim Zaim and Handayani, “Mitigating LLM hallucinations in Quranic content”, ICT4M 2025, 26–27 November 2025 (IIUM Repository): irep.iium.edu.my
- Abdelaal, Al Haffar, Fawzi and Magdy, IslamicMMLU, arXiv 2603.23750, 24 March 2026, revised 11 July 2026: arxiv.org
- Nushra and Morol, “Do Open Language Models Invent Scripture? BILAL-BENCH”, poster, Muslims in ML affinity event, NeurIPS 2026: neurips.cc
- Dar al-Ifta al-Misriyyah, fatwa no. 8784 on relying on AI for fatwas (Arabic), 22 September 2025: dar-alifta.org
- Aaj English TV, Dar al-Ifta on AI and Quran interpretation, 28 January 2026: english.aaj.tv
- Dar al-Ifta, recommendations of the 10th international fatwa conference, August 2025: dar-alifta.org
- Dar al-Ifta, Egypt’s mufti opens the 11th international conference, 2 September 2026: dar-alifta.org
- CSIS, “Chat GPMufti”, 24 October 2023: csis.org
- Egypt Today, Al-Azhar Grand Imam on AI without ethical safeguards, 28 June 2026: egypttoday.com
- Egypt Independent, Al-Azhar Observatory on AI-made Quranic songs, 21 September 2024: egyptindependent.com
- General Iftaa’ Department, Jordan, “Utilizing Artificial Intelligence in Issuing Fatwas”, 26 April 2026: aliftaa.jo
- General Iftaa’ Department, Jordan, “Artificial Intelligence: An Assistant or a Mufti”, 21 July 2025: aliftaa.jo
- detik.com, MUI chairman says AI cannot be a reference for religious rulings, 20 August 2025: detik.com
- Gulf News, UAE Council for Fatwa acts against unlicensed online fatwa accounts, 23 July 2026: gulfnews.com
- Arab News, HUMAIN launches HUMAIN Chat, 26 August 2025: arabnews.pk
- Jabatan Mufti Wilayah Persekutuan, Tinta Mufti Bil. 116/2026, AI and human values in finance, 9 July 2026: muftiwp.gov.my
- Jabatan Mufti Wilayah Persekutuan, Irsyad Hukum Siri ke-932, AI figurine images, 17 September 2025: muftiwp.gov.my
- Jabatan Mufti Wilayah Persekutuan, Irsyad Hukum Siri ke-934, deepfake AI as pornography, 24 September 2025: muftiwp.gov.my
- The Star, “Don’t use AI for Islamic teachings, says MP”, 18 February 2025: thestar.com.my
- Jabatan Mufti Wilayah Persekutuan, talk on the use of AI in fatwa, 22 January 2025: muftiwp.gov.my
- The Star, PAS MP warns AI could threaten Islamic understanding, 17 February 2025: thestar.com.my
- RTM, Federal Territories mufti on fiqh in the AI era, 26 August 2025: berita.rtm.gov.my
- Sinar Harian (Bernama), religious authorities should draft shariah guidelines on AI, 9 September 2025: sinarharian.com.my
- Jabatan Mufti Wilayah Persekutuan, Tinta Mufti Bil. 36/2025, the role of fatwa institutions, 30 November 2025: muftiwp.gov.my
- RTM, Religious Affairs Minister on AI and religious questions, 22 September 2026: berita.rtm.gov.my
- Jabatan Mufti Wilayah Persekutuan, visit by Maybank Islamic affairs and tanyalahustaz.com, 5 May 2025: muftiwp.gov.my
- Sinar Harian, amended Printing of Quranic Texts Act, 8 February 2024: sinarharian.com.my
- Kementerian Dalam Negeri, Garis Panduan Permohonan Penerbitan Teks Al-Qur’an Secara Digital, 2024: moha.gov.my
- Sinar Harian, KDN does not block Quran publishing, only certification needed, 15 August 2023: sinarharian.com.my
- RTM, 22,002 Qurans collected through Jom Exchange Al-Quran, 7 October 2026: berita.rtm.gov.my
- Dicastery for the Doctrine of the Faith and Dicastery for Culture and Education, Antiqua et Nova, 28 January 2025: vatican.va
- Holy See, Pope Leo XIV, Message for World Communications Day 2026, 24 January 2026: vatican.va
- Holy See, Pope Leo XIV, encyclical Magnifica Humanitas, 15 May 2026: vatican.va
- Associated Press via PBS NewsHour, Pope calls for regulation of AI in encyclical, 25 May 2026: pbs.org
- Catholic News Service via USCCB, Pope Leo urges world to “disarm” AI, 25 May 2026: usccb.org
- Sunday Examiner, Asian bishops on responsible engagement with AI, 13 December 2025: examiner.org.hk
- Anglican Diocese of West Malaysia, Alex Selvarajah, “Shaping the Digital Future of Anglican Liturgy in the Age of Artificial Intelligence”, Tanjong Series 2025-11, posted June 2026: anglicanwestmalaysia.org.my
- RenAIssance Foundation, Rome Call for AI Ethics, accessed 17 November 2026: romecall.org
- RenAIssance Foundation, AI Ethics for Peace: world religions commit to the Rome Call, 25 June 2024: romecall.org
- UCA News via Catholic Sabah, why are Asian religions silent on AI, 11 May 2023: catholicsabah.com
- Jiji Press via Adnkronos, Japanese group develops Buddhist AI humanoid robot, 25 June 2026: english.adnkronos.com
- Rest of World, GitaGPT and religious chatbots in India, 9 May 2023: restofworld.org
- Daily Express, Malaysian temple unveils “world’s first AI Mazu”, 28 April 2025: dailyexpress.com.my
- Marketing-Interactive, MAMEE recalls bags after AI artwork of Lord Murugan, 21 September 2026: marketing-interactive.com
- Catholic News Service via The Monitor (Trenton), Vatican warns about fake pope quotes and videos, 22 May 2025: trentonmonitor.com
- Catholic World Report (CNA), Vatican struggles against deepfakes of Pope Leo XIV, 25 September 2025: catholicworldreport.com
- New Straits Times, MCMC deepfake removals and AI labelling, 23 July 2026: nst.com.my
- The Vibes, JAIS probes viral religious ritual footage, 11 October 2026: thevibes.com
- Holy See, Pope Leo XIV, address to the College of Cardinals, 10 May 2025: vatican.va
- The Guardian, “Deus in machina: Swiss church installs AI-powered Jesus”, 21 November 2024: theguardian.com
- The Register, Catholic Answers’ AI priest demoted, 3 May 2024: theregister.com
Related research
A dead man’s AI voice undid his killer’s sentenceOn 30 September 2026 an Arizona appeals court threw out a 10½-year sentence because the judge “loved” an AI video of the dead victim forgiving his killer.
AI helped read a burnt scroll. Now try Jawi.In June 2026, X-ray scans and machine learning let scholars read a sealed, carbonised Herculaneum scroll end to end for the first time.
The voice was Mum’s. The caller wasn’tThree seconds of audio gave an 85% voice match in McAfee’s tests.