OCR and AI quality problems
Summary
Two kinds of apps guess: recognition apps (OCR) guess letters from pixels, and AI apps guess meaning from text. Both can be wrong in ways that look right. Good results come from good input (straight, upright, correct language, real text) and from checking what matters. This page gives the fixes in the order that pays off most.
Details
Part 1: OCR that reads badly
OCR lives in OCR Image (text out), Searchable (invisible text layer in your PDF) and as a switch inside other apps (Converter, Resize tiling). Same cure for all of them.
The checklist, in order of impact:
- Orientation first. Sideways or upside-down pages hurt recognition more than anything else. Searchable has a "Source photo is sideways (rotate 90°)" switch but no automatic orientation detection. For a PDF, rotate with Rotation first; for a photo, rotate it on your device.
- Choose the document language. Wrong language is the number one cause of gibberish. The language list has about 35 entries, including Cyrillic languages, Arabic, Chinese (simplified and traditional), Hindi, Japanese, Korean and Thai. The default is English, so a non-English document needs the setting.
- Straighten. Keep Deskew on in Searchable for phone photos. For a file that is already a PDF, run Deskew first, then Searchable (Deskew outputs page images, so OCR after it, not before).
- Clean the background. For yellowed, shaded or stamped pages, tick "Remove background/watermark".
- Resolution. Tiny or blurry pictures give poor text. Take or rescan at a higher resolution; use a straight, well-lit photo.
- AI correction last. "Correct OCR errors with AI" can fix recognition mistakes, but it can also change words. Spot-check names and numbers.
- Split long files. Searchable refuses very long documents with a message asking you to split; use Splitter, then merge the parts.
Symptoms
| Symptom | Likely fix |
|---|---|
| Output looks like gibberish | Language, orientation, resolution |
| "Words are wrong or missing" | Same, then try background cleanup and AI correction |
| PDF "looks the same" after Searchable | Correct: the text layer is invisible. Press Ctrl+F in your viewer to check |
| Text selectable but order is odd | Columns and tables are approximate; use Parser or Excel output with "tables only" |
| Tables came out as lines of words | Plain text cannot hold tables; Searchable to Excel with "Extract tables only" |
| File got bigger | The text layer was added; run Compressor |
| A search for a word finds nothing but you see it | In Search, tick the scanned box and set the language. On a scan, a misread word cannot be found, so "no results" does not prove absence |
Part 2: AI apps and scans
Most AI apps read the text layer of your PDF. A scan has none.
| App | What happens with a scan |
|---|---|
| Ask PDF | Stops with a message that the PDFs "look scanned" or "No extractable text was found" and asks you to make them searchable |
| PDF Knowledge Base | Shows the status "No text (scanned?)" for that document |
| Checklist | A placeholder checklist, or "No checklist items could be extracted" |
| Cheat Sheet | Error asking for a text PDF or a topic; with a topic, the scan may be silently ignored |
| Chat Create | Nothing useful happens after a scan upload |
| Unwatermark | Empty result: there was no text |
| Abstracts | Exception. It recognises scanned pages itself, but slowly, and unread pages are left out within the time limit. If nothing is readable you get "No text could be read..." |
| Resume | Works, but quality depends on recognition |
The rule: run Searchable first, check that you can select text in the result, then upload that PDF to the AI app.
Part 3: AI answers that look wrong
- Trust but verify. All AI output is a well-organised draft (Trust, verification and limits). Check numbers, dates, names and legal wording against the source.
- Open the citations in Ask PDF and PDF KB. A kept citation proves the passage exists, not that the model read it correctly.
- "The documents don't cover it." Believe it first, then rephrase using the document's own words; search is word-based as well as meaning-based.
- Long documents are sampled. Abstracts, Cheat Sheet and PDF KB read parts of long documents (the free plan reads fewer parts; the progress text says "Extracting key points from N of M parts..."). Ask about specific chapters to get detail.
- Charts and photos are invisible. Only text counts.
- Topic-only Cheat Sheets come from the model's general knowledge and nothing checks them; give a source PDF.
- Vague request, vague answer. Checklist needs a specific topic; Resume Improve needs the job text; give audience and outcome.
- Compare two runs. Items that appear in only one deserve a look. Output varies between runs (TOC AI, SVG Maker, Illustrator, Collage).
- Verification built in: non-fiction Abstracts are checked against the document after writing when time allows; fiction is not.
See also Good inputs for AI apps.