The experiment

Building a research archive of Nastaliq typography means confronting an uncomfortable fact: most of the "knowledge" floating about these fonts is unsourced, recycled, or wrong — and increasingly, it is machine-generated.

So we ran the experiment on ourselves. We collected six independent AI research sets — ChatGPT, Claude, DeepSeek, Gemini, Grok and Meta AI — each asked to research Nastaliq font history without seeing our database. Then we checked their claims, one by one, against our graded sources, 56 fresh verification searches, and — decisively — against font binaries analysed with fontTools (name tables, glyph counts, GSUB inventories, checksums).

Finding 1: every AI mixed gold and fool's gold

None of the six was simply right or simply wrong. Each mixed correct material, wrong material and material nobody else had:

• Meta AI contributed the most unique verified families — Tamannai 1965, Wathiq, Diwan, Shekanj, Naqsh Display, Nastaleeq ASC, Indopak — because it surfaced social-channel announcements (Behance portfolios, foundry posts) that text-web searches miss. • Gemini alone attempted the pre-digital metal-type arc , and most of it survived verification. • DeepSeek had the best name-table detail — version strings, exact glyph counts. • Claude produced the largest bibliography and the most honest dispute list. • ChatGPT showed the best epistemic hygiene — every one of its explicit cautions proved warranted. • Grok had the most accurate high-level narrative.

The lesson is not "use AI" or "avoid AI". It is that an AI answer is a lead , never a source .

Finding 2: two systematic failure modes

Across almost all six sets, the same two bugs repeated:

1. Date conflation An AI reads "designed 2014", "announced 2015", "released 2016", "added to a catalogue 2020" and emits "released in 2014". We watched a single font's history compress four milestones into one wrong year, repeatedly. (Our archive now records each family with separate development, release and catalogue-addition dates for exactly this reason.)

2. Family conflation Similar names fused into single fonts: Manzar Lahori merged with Faiz Lahori (different designers, different decades). Three unrelated Naqsh fonts — PakType's Naskh face, Rawaz Hammas's 2024 Nastaliq, and Hoor Zubair's Naqsh Display — collapsed into one "Naqsh". Jameel Noori blended into the Noori lineage until the dating problem became unsolvable.

Both failure modes are plausible-sounding . That is what makes them dangerous in a citation context.

Finding 3: the font binary is the arbiter

When accounts conflicted, the file settled it:

• Alvi Nastaleeq's binary says version 1.2, "Octuber 2008" — typo included — ending a multi-source date argument in one string. • Structural comparison (cmap + GSUB overlap) proved AA Sameer Sagar is a bold restyle of Nafees Nastaleeq 1.02 at 100% structural identity — a claim no web source states outright. • Noor Dotted Nastaleeq fell out as a dotted restyle of Fajer Noori; the Shekasteh cluster resolved into IranNastaliq descendants. • Two AI-originated claims died completely: four Pakistani "community families" (Gohar, Aftab, Barq…) with zero web evidence , and "Gulzar supports 417 languages", contradicted by Google's own blog.

Binary evidence resolved 13 of 17 disputes in one research cycle. It is the highest-leverage verification tool in type history — cheaper than archival travel, more honest than memory.

The protocol we would recommend to anyone

1. Collect answers independently — never let one AI see another's output. 2. Split every date claim into design / announcement / release / catalogue milestones before checking. 3. Disambiguate names first — build a table of same-named families before research begins. 4. Grade every surviving claim (A primary → D weak) and cite the grade, not just the fact. 5. Prefer artefacts over accounts — a name table beats a blog post, every time. 6. Publish the failures — our archive keeps unverified candidates, disputes and gaps in the open, because a research resource you cannot audit is just marketing.

The full verdict record — every AI, every claim, every resolution — is in the Nastaliq Research Archive, and the families named above all carry their evidence on their own pages. Start with the archive's open questions section if you want to help close one.

Methodology and verdicts: rtlfonts.com Nastaliq Research Archive, database v1.4.0, cross-verification session 4 (2026-11-04).