--- title: "Nothing on the Page" published_at: "2026-10-04" canonical_url: "https://www.truthpromoters.com/dispatches/nothing-on-the-page" type: "dispatch" --- # Nothing on the Page _The machine that reads our library's scanned pages was given a blank page from al-Ṭūsī's Tahdhīb al-Aḥkām, and wrote an essay about George Orwell's 1984. So we measured all 24,368 pages it has read, found every page like it, and changed the reader so that a page with nothing on it is never put in front of it again._ > يَـٰٓأَيُّهَا ٱلَّذِينَ ءَامَنُوٓاْ إِن جَآءَكُمۡ فَاسِقُۢ بِنَبَإٖ فَتَبَيَّنُوٓاْ أَن تُصِيبُواْ قَوۡمَۢا بِجَهَٰلَةٖ فَتُصۡبِحُواْ عَلَىٰ مَا فَعَلۡتُمۡ نَٰدِمِينَ > > **Literal:** O you who have believed, if there comes to you a disobedient one with information, investigate, lest you harm a people out of ignorance and become, over what you have done, regretful > > **Contextual:** You who believe, when someone known for breaking the rules brings you a report, check the facts first so you don't hurt an innocent group through carelessness and end up regretting what you did. > > — Quran, Surah Al-Hujurat (The Rooms), Chapter 49, Verse 6 ## A Blank Page Our research library holds thousands of books in Arabic and Persian, and many of them are scans: photographs of printed pages, with no text inside the file at all. To make those pages searchable, a machine reads each one: a vision model that looks at the image of a page and writes down the words it sees. Its readings are never offered to anyone as a quotation. Every one is marked not quotable. It is a finding aid: it tells a person which page to open, and a person reads the page. But a finding aid that finds things which are not there is still broken. A few nights ago we were testing a second reading engine on the pages the first one struggles with. One of them was a page of our copy of Shaykh al-Ṭūsī's *Tahdhīb al-Aḥkām*, one of the Four Books of Shīʿī ḥadīth. We opened it to see what made it difficult. It was blank. Not faded, not damaged: white, edge to edge, the kind of page a printer leaves between sections. The machine had read it anyway. It had written a short essay, in English, about George Orwell's *1984*. **Nothing was on the page. The machine filled the silence.** ## Why a Machine Does This A reading model learns from millions of pages that have something on them. Its job, as far as it knows, is to produce text, and it has almost never been shown a page where the right answer is nothing. Give it one, and it does what a nervous student does with an empty exam paper: it writes something anyway. What it writes comes from its own memory, not from the page and not from our library. That is how a book of ḥadīth came to hold a dystopian novel. Of all the books it could have conjured, it reached for one about a world where the record is quietly rewritten. The checks we already had did not catch it. When these models fail, they usually fail loudly: they get stuck and repeat one phrase hundreds of times, and we look for that. This failure was fluent. No repetition, no stutter, nothing to trip on. **A loud failure announces itself. A fluent one has to be looked for.** ## We Checked Every Page So we looked. Every page the machine has read so far, 24,368 pages in 54 books, was measured for one thing: is there anything on it at all? 141 pages had nothing. On 111 of them the machine had correctly written nothing. On the other 30, it had invented something: - on 24 blank pages of the *Tahdhīb*, versions of the same essay about *1984*; - on 3 blank pages of Shaykh al-Ṣadūq's *Kamāl al-Dīn*, several thousand characters of computer drawing code; - on 2 blank pages of another work, a description of the paper itself: "a white background with black dots… several small black dots scattered". The dust, in other words; - and on one, a single full stop. Those 30 readings have been taken out of the index, and each page is now recorded as what it is: blank. ## How Do You Tell a Machine a Page Is Blank? It sounds easy. A blank page is white, and a page with writing has black on it. Count the black. A scanned page is not that simple, and the reasons are worth naming, because each one needed its own answer. **Paper is not white.** It is cream, or yellowed with age, or grey from the scanner. So we never ask whether a mark is black. We ask whether it stands out from that page's own paper, by a margin wide enough that the grain of the paper, and the faint ghost of text showing through from the other side, never reach it. **Scanners leave dust.** A speck of dust is as dark as ink. What it is not is large. So we look at each page at low resolution, where every point we measure is the average of a small patch of the scan. A speck dissolves into the paper around it; a stroke of a pen fills its patch and stays dark. The machine that reads the page still gets the full image. Only the question "is anything here?" is asked of the small one. **Not every page is light.** A book's cover can be dark cloth with gold lettering: lighter than its background, not darker. So we measure contrast in both directions. ## Where to Draw the Line The last question is how much makes a page not blank. Our first instinct was half a percent of the page. It sounded reasonable. Measured against our own pages, it would have thrown away 84 pages of real writing: the last lines of a chapter, a line on its own, a short dedication. So we let the pages decide. The faintest real writing anywhere in those 24,368 pages, a line of thirteen characters, covers about a fifty-thousandth of its page. The empty pages, dust and all, sit below that. We drew the line four times lower still. Between that line and the faintest real writing there is a narrow band of pages that are probably empty but not certainly. Those still go to the machine, and that is deliberate. **When in doubt, the page is read. Losing a real page in silence is worse than reading an empty one.** ## What We Got Wrong on the Way Our first measure looked only for marks darker than the paper. It called two book covers blank, dark green cloth with gold lettering, and would have stopped the title of the *Tahdhīb* itself from ever being read. We found that by looking at the pages it flagged, not by trusting the number. It now measures both ways. And the check for a stuck machine let a fluent invention straight through. So there is now a second, separate check: a reading from a book in Arabic or Persian that contains no Arabic script at all is set aside for a person to look at, whatever it says. We mention these because we said we would tell you this kind of thing. ## What Changes Before a page is put in front of the machine now, it is measured. A page with nothing on it is recorded as blank, with the measurement that says so, and the machine never sees it. Some printers put a line in the middle of an empty page, *this page intentionally left blank*, so that anything reading it has something true to say. Most books do not. Now it does not matter. > لَنْ يُصْدَقَ الخَبَـرُ حتّى يَتَحَقَّقَ العَيانُ. > > **Literal:** Reports will never be accepted as true until [they are] verified by the eyes. > > **Contextual:** You won't believe a report to be true until you have seen it with your own eyes. > > — Ghurar al-Hikam, Narration and Transmission of Information **The page is measured. The reading is checked. A person reads the page.** --- Read this dispatch on the web: https://www.truthpromoters.com/dispatches/nothing-on-the-page