Quick Answer
This guide introduces "grounding prompts" that drastically reduce AI hallucinations for lawyers. By using a strict "Phase 1: Input Audit" and "Phase 2: Strict Drafting" framework, attorneys can force AI models to verify facts before writing, saving hours on petitions (like I-589s or O-1As) while ensuring citations are real and evidence is accurately referenced.
Lawyers lose hours drafting petitions. Discover "grounding prompts" that stop AI hallucinations, force fact-checking, and draft complex legal documents (I-589, O-1A) with verifiable accuracy.
Awais Haq
CEO Time Technologies LLC
The Daily Struggle of Petition Drafting
Drafting tends to be the most time consuming part of an immigration lawyer's routine. You may find yourself spending 3 to 6 hours a day, simply hunched over papers and by your computer.
Drafts take up time that could be better spent in court, negotiating or counselling clients. You lose competitive edge the longer you spend writing statements and declarations. The number of cases you take up is also limited by how much time you can free up without burning out.
Grounding and Prompts: How AI Can Help
You must have heard of fellow attorneys using AI to generate templates. They may even draft entire petitions. But equally, you must have noticed a high propensity for hallucinations. How do you avoid those?
Freely available AIs like ChatGPT or Grok provide generic and approximate responses, rife with hallucinations. Grounding is a method that can remedy this. It is the process of confining a model's abstract knowledge to sets of verifiable, real-world data. If prompts for generic AIs are grounded, they result in statements with minimal errors.
Through extensive experimentation with many publicly accessible AIs, I have identified the perfect workflow:
- The most reliable GPT to use
- The optimal grounding prompt
- 10 practical AI prompt samples
The Most Reliable GPT To Use
I don't claim that certain AI is more reliable than the other for general use. However, after testing my prompts on many AIs, certain patterns emerged. The AI that had the lowest number of hallucinations turned out to be Gemini 3, so my prompts have been fine-tuned to it.
The AI with the highest hallucination rate was Grok. It was explicitly told to demand further input if it had insufficient information. Instead, Grok filled in the blanks and created reports with excessive false details.
As such, Gemini 3 should be used to get the most out of these prompts. Note however that even Gemini 3 will result in some errors, so lawyer oversight is necessary.