No, ChatGPT is not enough to run a personal injury firm. ChatGPT can draft, summarize, and brainstorm well enough to be genuinely useful, and firms that dismiss it are leaving real time on the table. But it cannot carry the parts of a personal injury practice where the money is actually made or lost: reading a 900-page medical record without inventing a diagnosis, valuing a case against comparable outcomes, and producing a demand an adjuster will pay on. Those are the constraints that decide whether a general-purpose model is a tool in your firm or the operating system of it.
This guide covers what ChatGPT does well in plaintiff work, where it breaks, what your ethics obligations are when you use it, and how to tell whether your firm is small enough to get by on it alone.
If used properly, ChatGPT can be good enough for work where you are the verifier and the stakes of an error are low:
The common thread: you already know the right answer, and the model is saving you keystrokes rather than supplying judgment. In that lane it is fast, cheap, and legitimately good. A solo who uses ChatGPT this way and nothing else is not making a mistake.
The failure begins the moment the model is asked to supply the judgment instead of the keystrokes.
ChatGPT starts to fall apart on the three tasks that really determine case value: reading the medical record, valuing the claim, and building the demand package.
Medical record review. A serious PI case arrives as hundreds or thousands of pages of scanned, out-of-order, partially illegible provider records. General-purpose models are not built to ingest that volume with the accuracy the record requires. They lose the thread across a long document, silently skip pages, and produce a chronology that reads fluent and is wrong in ways that are expensive to catch. The error mode is not “it fails obviously.” It is “it produces a confident treatment timeline missing the two ER visits that establish causation.”
Valuing a claim. Asking ChatGPT what a case is worth returns a plausible-sounding range assembled from whatever it absorbed on the open internet. It is not anchored to actual verdicts and settlements in your venue, for your injury type, against that carrier. A number with no comparables underneath it is not a valuation. It is a guess with good grammar.
Winning demand package. ChatGPT will write you a demand letter. It will not know which ICD codes matter, which treatment gaps the adjuster will exploit, or how to construct damages that survive scrutiny. The letter will look right. Adjusters read hundreds of these, and they can tell.
Underneath all three sits the same limitation: general-purpose models are trained broadly and know personal injury the way they know everything else, which is to say, generally.
Partially, poorly, and not at all, in that order.
| Task | Can ChatGPT do it? | The catch |
| Legal research | Partially | It does not have reliable access to current case law and will fabricate citations that look real. Every cite must be independently verified in a real database. |
| Document drafting | Yes, as a first draft | Output quality depends entirely on what you feed it and how carefully you review it. It has no knowledge of your firm’s standards or prior work. |
| Medical chronology | No | Volume and accuracy demands exceed what a general model reliably handles. |
| Case valuation | No | No access to structured verdict and settlement data. |
| Case management | No | It has no memory of your caseload, no deadlines, no integrations, no audit trail. |
| Client intake | No | It cannot run a workflow, route a lead, or follow up. |
The pattern: ChatGPT is a drafting assistant. It is not a system of record, and it is not a source of truth. Firms get into trouble when they let a drafting assistant behave like either one.
Using ChatGPT does not transfer any professional responsibility away from you. Three duties are directly implicated. These are not new obligations invented for AI.
In July 2024, the American Bar Association issued Formal Opinion 512, its first formal ethics guidance on generative AI, which confirmed that the existing Model Rules of Professional Conduct apply directly to a lawyer’s use of these tools. Three of those duties bear most heavily on plaintiff work.
| Competence (Model Rule 1.1) | You are responsible for the accuracy of what you file and send, regardless of what produced the first draft. Formal Opinion 512 is explicit that lawyers must understand the capabilities and limitations of the AI tools they use, and that a general-purpose tool cannot substitute for a lawyer’s own competent legal work. “The AI wrote it” has never once worked as a defense. |
| Confidentiality (Model Rule 1.6) | Client information entered into a consumer AI tool may be retained or used for training depending on the product tier and settings. Opinion 512 advises lawyers to understand how a tool uses input data and to obtain the client’s informed consent before entering client confidences, noting that boilerplate language buried in an engagement letter is not sufficient consent. A free-tier consumer account is not a safe default for protected health information. |
| Supervision (Model Rules 5.1 and 5.3) | Opinion 512 places an affirmative duty on managing and supervising lawyers to establish clear policies on permissible AI use and to ensure both lawyer and nonlawyer staff are trained to follow them. If your team is using AI tools, supervising that use is your obligation, not theirs. |
The practical translation: ChatGPT is not prohibited, and the ABA did not prohibit it. It is simply not a place to put responsibility.
The hallucination problem is not that ChatGPT is wrong sometimes. It is that it is wrong confidently, and the confidence scales with how little you know about the subject.
The most cited example is Mata v. Avianca. In June 2023, Judge P. Kevin Castel of the Southern District of New York sanctioned two attorneys and their firm $5,000 under Rule 11 after they filed a brief citing six judicial opinions that ChatGPT had fabricated entirely, and then stood behind the fake cases when the court questioned them. Fittingly, the underlying matter was itself a personal injury claim. The court was careful to note there is nothing improper about using a reliable AI tool for assistance; the sanction was for failing to verify its output.
That case became the template, and it has not been the last. Courts have continued to sanction lawyers for the same failure in the years since.
For PI firms, the exposure is quieter and more common than a fabricated cite. It is a chronology that omits a visit. A damages figure with nothing under it. A demand that misstates a diagnosis. None of these get you sanctioned. All of them cost you money on the settlement.
Here is the trap that makes this economically self-defeating: the verification burden scales with the volume you are trying to automate. If you have to read every page of the record yourself to confirm the AI read it correctly, the AI saved you nothing. Automation you cannot trust is not automation. It is a second copy of the work.
That is the actual dividing line between a general-purpose model and purpose-built software, and it is not about intelligence. It is about whether the output arrives verified.
The distinction is not model quality. Frontier models are extremely capable, and PI software is often built on top of them. The distinction is everything wrapped around the model:
| General-purpose AI | Purpose-built PI platform | |
| Training data | The open internet | Structured personal injury case data |
| Medical records | Degrades at volume | Built for high-page-count ingestion |
| Valuation basis | No comparables | Verdict and settlement data |
| Verification | You do it, every time | Expert review before delivery |
| Case memory | None | Persistent across the entire case lifecycle |
| Audit trail | None | Required and retained |
The honest version of the comparison: firms need the comprehensive capabilities of purpose-built PI tools. For example, EvenUp is a proactive, AI-native operating system that supports the entire case lifecycle, from intake through resolution.
ChatGPT is a capable drafting assistant and a poor system of record. It can save a plaintiff firm real hours on correspondence, summarization, and first drafts. It cannot read a medical record accurately at volume, cannot value a case against comparables, and cannot produce a demand that holds up under adjuster scrutiny, and those three things are where plaintiff firms make their money.
The right question is not whether ChatGPT is good. It is which parts of your practice you are willing to leave unverified.
