If your firm is rolling out AI tools, or you’ve been quietly wondering whether you’re allowed to use one, the useful question isn’t whether AI can help a paralegal. It can. The question is how to use it so your supervising attorney can stand behind every piece of work it touched.
That comes down to three disciplines: using only approved tools, verifying everything, and keeping a record the attorney can review. This guide covers the ethics framework that applies, which tasks suit AI and which don’t, a verification protocol, how to protect client confidentiality, and wording for raising concerns. It uses the ABA Model Rules and ABA Formal Opinion 512 as the reference point. Many state bars have issued their own AI guidance, and state rules differ, so your firm’s policy and your jurisdiction’s guidance control.
What does ABA Formal Opinion 512 mean for paralegals?
ABA Formal Opinion 512, issued in July 2024, doesn’t create new rules for generative AI. It explains how existing Model Rules apply, and several of them land directly on paralegal work through the lawyer’s duty to supervise.
- Competence (Rule 1.1). Lawyers need a reasonable understanding of a tool’s capabilities and limits, and output must be checked to a degree that fits the task.
- Confidentiality (Rule 1.6). Before client information goes into a tool, the risk of disclosure has to be evaluated. For self-learning tools that could expose information to others, the opinion calls for informed client consent.
- Communication (Rule 1.4). In some situations the client should be told how AI is being used on the matter.
- Candor and meritorious claims (Rules 3.1 and 3.3). Anything submitted to a court must be accurate, whatever produced the first draft.
- Supervision (Rules 5.1 and 5.3). Lawyers with managerial authority should set clear policies on permitted AI use and train staff on them.
- Fees (Rule 1.5). Hourly billing covers time actually spent, not time the tool saved.
For a paralegal, Rule 5.3 is the practical center: the attorney is responsible for making sure your work is compatible with their professional obligations, which means your AI use has to be visible and checkable. The guide to Rule 5.3 and supervising AI tools covers the firm side, and what Formal Opinion 512 means for your firm covers the opinion in more depth.
Which paralegal tasks are suited to AI?
AI fits best where the output is a draft or a working aid that a person checks against a source. It fits worst where the output is a final judgment, a legal conclusion, or something that goes straight to a court or client.
| Task | Fit | Required check |
|---|---|---|
| Extracting text from screenshots and scans | Good | Compare names, numbers, dates and times against the image |
| First-draft deposition or document summaries | Good | Confirm each point and page-line cite against the transcript |
| Drafting a chronology from documents | Good with care | Tie every entry to a source document; recheck dates |
| First drafts of routine letters and requests | Good with care | Attorney review before anything is sent |
| Summarizing legal research | Caution | Pull and read every authority; confirm it exists and says what is claimed |
| Privilege or relevance decisions | Not appropriate | These are attorney judgments; AI can only help sort for review |
| Advice or conclusions sent to a client | Not appropriate | Only the attorney gives legal advice |
| Creating or altering exhibits | Not appropriate | Exhibits must reproduce the source faithfully |
The last two rows matter most. AI doesn’t change the rule that nonlawyers don’t give legal advice; ABA Model Rule 5.5 bars lawyers from assisting the unauthorized practice of law, and a client email built from AI output is still the firm speaking.
How do you verify AI output?
Verify against the source, not against your sense that it sounds right. Fluent, confident output that is wrong is the characteristic failure of generative AI, and it is hardest to catch when you’re tired or rushed.
- Know the source. Identify exactly what the output was based on: which transcript, which documents, which export. If you can’t say, you can’t verify it.
- Check every citation. Pull each case, rule and statute. Confirm it exists, the citation is correct, it says what the output claims, and it is still good law.
- Check every quotation word for word against the original.
- Tie every fact to a record cite. Each factual statement should point to a Bates number, exhibit, page and line, or message ID.
- Recompute numbers and dates. Totals, durations, deadlines and time-zone conversions are common failure points.
- Look for omissions. Ask what the summary left out, especially anything unhelpful to the client.
- Record the check. Note what you verified and anything you couldn’t.
This matters most for anything that will be filed. Under Federal Rule of Civil Procedure 11(b), the signer certifies after reasonable inquiry that legal contentions are warranted and factual contentions have evidentiary support, and state rules impose similar duties. Some judges also have standing orders on AI use in filings. More techniques are in preventing AI hallucinations in legal work.
How do you protect client confidentiality when using AI?
Use only tools the firm has vetted and approved, through firm accounts, and share only what the task needs. ABA Model Rule 1.6(c) asks lawyers to make reasonable efforts to prevent unauthorized disclosure of client information, and that extends to the tools staff use.
- Never paste client information into a personal or free consumer AI account
- Know whether the approved tool retains inputs or uses them for training, and who can access them
- De-identify where the task allows: “Parent A” and “Child 1” work for many drafting tasks
- Don’t upload sealed, protected or third-party confidential material unless the attorney confirms it is permitted
- Follow any protective order in the case, which may restrict where produced documents can go
Vetting cloud tools is its own topic; the guide to client confidentiality and cloud software lists what to look for.
Why keep an AI-use log?
A simple log makes your AI use visible to the supervising attorney, which is what Rule 5.3 supervision requires in practice. It also answers billing, disclosure and court-order questions later without anyone having to reconstruct what happened.
Template: AI-use log entry
Date / matter: [date], [matter name or number]
Tool: [firm-approved tool name]
Task: First-draft summary of [witness] deposition, volume 1
Input: Transcript only; no other client documents
Verification: All 38 page-line cites checked; 3 corrected; 1 omission added (pp. 112–114)
Reviewed by: [attorney initials and date]
How should you raise AI concerns with a supervising attorney?
Plainly and early, in writing, with specifics. Attorneys need to know when AI was involved and what you could and couldn’t confirm.
Sample wording
“I used [tool] to draft the attached chronology from the client’s texts and emails. I checked every entry against the source and corrected four dates. Two entries rely on screenshots where the year isn’t visible; I’ve highlighted them for you.”
“The research summary cited five authorities. I pulled all five: three are accurate, one is real but doesn’t support the proposition, and I couldn’t find the fifth at all. I’ve removed it and flagged the paragraph.”
What does responsible AI tooling look like in litigation support?
Responsible tools keep a person in the loop, keep originals untouched, and keep AI out of the steps that must be exact. Those design choices make supervision and verification much easier.
Look for tools where AI-generated content is clearly marked, human corrections are preserved, and final outputs are reproducible. In Evidence Helper, for example, AI-assisted text extraction from screenshots is reviewed and corrected by a person, corrections are stored separately from the original extraction, originals are hashed and preserved, and the exhibit PDF export is generated without AI. That separation lets you use AI where it saves time while keeping the exhibit itself faithful to the source.
None of this replaces judgment. The supervising attorney decides how AI is used on each matter, and when a question involves a specific jurisdiction’s rules, the answer comes from a licensed attorney there, not from the tool.
