If you've used AI for medical record review, you've probably seen this happen: the AI confidently claims a patient has "no history of back pain" when the physical therapy note on page 37 clearly says the exact opposite.
Most people assume the AI hallucinated. In reality, the problem often starts much earlier. With the PDF file itself.
The Problem: Spatial Blindness
PDFs are visual documents, not data documents. The format was created in the 1990s so files would look identical on every printer and screen. When you upload a PDF document with tables, grid selections, and handwritten notes into a standard AI, the result is often text without structure.
AI isn't reading the document you see. It is reading something closer to word soup.
A Simple Test: One Medical Progress Grid
We ran a simple experiment using a typical physical therapy progress report grid. We tested what different systems extracted from the same progress report grid.
A typical physical therapy progress report grid used in our extraction test.
| Software | What the AI Actually Reads | Evaluation | Legal Risk |
|---|---|---|---|
| Google Document AI | Never Seldom Occasional Activity Frequent Sitting ☑ Standing Lifting Driving ХХО | Fail Scrambled headers and misaligned selections. | Confident hallucination. Information is assigned to the wrong row. |
| Amazon Textract | Activity Never Seldom Occasional Frequent Sitting X Standing X Lifting X Driving | Fail Extracts text but ignores the grid structure. | Again: Confident hallucination. |
| ChatGPT | Aetraty Never Seldom Dewing boo wl A oO kK KX OOOO | Fail Complete gibberish. The meaning is lost instantly. | Someone has to manually review and re-enter the information. |
| AttorneyAide | Activity Never Seldom Occasional Frequent Sitting ☐ ☐ ✗ ☐ Standing ☐ x ☐ ☐ Lifting ☐ ✗ ☐ ☐ Driving ☐ ☐ ☐ ☐ | Pass Maintains grid logic and checkbox relationships. | Preserves the structure needed for reliable automation. |
Why This Matters for Your Law Practice
When spatial relationships are lost, AI fills the gaps. Not maliciously. Not randomly. But statistically.
When the structure disappears, the AI guesses the most statistically likely interpretation. That's how hallucinations appear in legal workflows. The output looks professional, but the underlying data is wrong.
AI doesn't hallucinate randomly. It hallucinates to repair broken data.
To ensure your legal AI saves you time, money, and effort use software specifically designed for your domain.
How to Hallucination-Proof Your Practice
For professional automated workflow, you need a strategy that prioritizes data integrity. Here are four practical safeguards when deploying legal AI automation:
The "Notepad" Audit
Copy a complex table from your OCR'd document and paste it into a plain text editor (like Notepad). If the text appears scrambled, your AI is already set up to fail.
Use the "LEGAL" Prompting Framework
Use this professional framework to improve reliability:
Legal Role: Assign a specific persona (e.g., "Act as a Senior PI Paralegal specializing in medical records review"). This primes the AI to respond within a professional context instead of responding like an average internet user.
Explicit Instructions: Define the exact task (e.g., "Identify all objective findings and tabulate medical expenses").
Guidance & Constraints: Tell it what to ignore (e.g., "Exclude subjective patient complaints and do not include duplicate entries").
Audience: Who is this for? (e.g., "Format this for a lead attorney preparing a demand letter").
Layout & Format: Be specific (e.g., "Output as a chronological table with page citations for every claim").
Audit and Validation
Never rely on a single AI output for litigation workflows. Use independent systems to audit and validate the output generated. This could include replicating the same prompt and analyzing the differences. Additionally, complementary prompts can identify inconsistencies.
Human-In-The-Loop Verification
Even the best AI requires verification. Incorporating citations and audits in the prior steps will speed up this phase. The gold standard is a domain specific system that does all this for you and allows you to click any sentence and review it against the exact page in the original PDF.
The Bottom Line
AI will not replace attorneys or paralegals. But firms who use domain-specific AI will absolutely outperform those relying on generic tools or no tools.
In Personal Injury law, the difference between a strong settlement and a missed detail is often hidden inside a checkbox grid, a progress table, or handwritten notes.
Don't let a 30-year-old file format introduce risk into your case preparation.
Curious how your own medical records would extract?
Try uploading a sample and see what the AI actually reads. Tools like AttorneyAide are designed specifically for medical record review — generating chronologies, tabulating medical expenses, performing audits, and linking everything back to the exact page in the original document.