Will AI replace lawyers? What the court record shows (2026)

If you practise law, you have read the prediction. The model reads the contract, drafts the brief, finds the cases, and the billable hour goes with it.

There is a public record to test that against, and it is not a vendor demo. It is one famous bank figure and a growing pile of signed court orders.

The figure first. In 2017 Bloomberg reported that JPMorgan’s COIN software does work that “consumed 360,000 hours of work each year by lawyers and loan officers” [source]. It is still the number quoted whenever someone says AI will replace lawyers.

What “will AI replace lawyers” actually asks

Will AI replace lawyers asks whether software can take over the work a licensed lawyer is paid and held responsible for, not just speed parts of it up. The answer depends on two things: what the tool can do, and who the legal system holds accountable when it is wrong.

This post looks at both, using five TIN case files: one automation claim and four court rulings.

Will AI replace lawyers?

No, not on the record so far. The automation evidence is thinner than its reputation, and the court evidence shows judges making the lawyer more accountable for AI output, not less.

Start with COIN, because the obvious reading is that it already happened. The bank’s software “reviews documents in seconds, is less error-prone and never asks for vacation,” Bloomberg wrote, on contracts that took 360,000 hours a year [source].

Read what the figure covers. It is “lawyers and loan officers” together, reading commercial-loan agreements [source]. It is a single-origin company claim from 2017, and JPMorgan “has been tight-lipped about the details of the proprietary technology” ([source]).

No one outside the bank ever measured those hours. No public source puts a lawyer headcount next to them. COIN is evidence that clause reading can be automated. It is not evidence that a lawyer was replaced.

What happens when AI does the lawyer’s drafting?

When AI drafts the filing and nobody checks it, courts punish the lawyers who filed it. That is the consistent result across the four rulings below, and across them the consequence moves from fines to removal from the case.

The canonical case is Mata v. Avianca in 2023. The lawyers “submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT” [source]. The brief “included six fictitious case citations” ([source]). The sanction was $5,000, jointly and severally [source].

Two years later the firms were bigger and so was the bill. In Lacey v. State Farm, a special master found “approximately nine of the 27 legal citations in the ten-page brief were incorrect in some way” [source]. Ellis George and K&L Gates were ordered, as firms, to pay $31,100 [source].

Then a court said fines were not enough. In Johnson v. Dunn, Judge Anna Manasco said fabricated authority “demands substantially greater accountability than the reprimands and modest fines that have become common” [source]. She reprimanded three Butler Snow lawyers, disqualified them from the case and referred them to the state bar [source].

Does a legal AI tool change who is responsible?

No. A purpose-built legal research product moves the responsibility nowhere: the 2026 Sixth Circuit ruling in United States v. Farris sanctioned the lawyer, not the software.

The court’s “first tell was the file name of the principal brief: ‘CoCounsel Skill Results.’” CoCounsel is Westlaw’s AI platform [source]. The cited cases were real; the quotations from them were not.

The court said that difference “does not absolve him” [source]. The lawyer was denied his Criminal Justice Act pay for the appeal and referred for possible discipline ([source]). In a separate order, the court removed him from the case [source].

This is the part the replacement forecast misses. A lawyer is not paid only to produce text. A lawyer is paid to be the person a court can sanction for that text, and no court in this record has moved that duty onto a tool.

Will AI replace paralegals?

AI could take over parts of the review work around a filing, but the record shows that review is exactly the step that failed. TIN holds no case file that measures paralegal headcount, so this section makes no claim about job numbers.

What the record does show is the hand-off. In Lacey, the order found that “No attorney or staff member at either firm apparently cite-checked or otherwise reviewed that research before filing the brief” [source]. The AI wrote an outline; the missing step was a human check.

In Johnson v. Dunn the firm itself was released without sanction, and an independent account notes the court credited Butler Snow’s existing AI policies and its response [source]. The firm with a verification process was spared; the individuals who skipped it were not.

The proof

TIN’s case file on COIN reports the 360,000-hours figure exactly and in context, and says plainly that it was a single-origin corporate claim never independently measured: JPMorgan’s COIN.

TIN’s case file on Mata v. Avianca reads the signed sanctions opinion and two independent newsrooms: the six fabricated cases and the $5,000 sanction.

TIN’s case file on Lacey v. State Farm checks the 9-of-27 count and the $31,100 order against the order text and two newsrooms: Ellis George and K&L Gates.

TIN’s case files on Johnson v. Dunn and United States v. Farris read both opinions in full: three lawyers disqualified and referred, and a lawyer unpaid and removed over CoCounsel quotations.

Read the limits honestly. These are one automation claim and four sanctions rulings, not a survey of legal employment. None of them measures how many lawyers AI has displaced, and this post does not claim a number. For the wider sanctions pattern, see what the AI fabricated citations record shows.

What each case settled

The bottom line

The forecast that AI replaces lawyers treats the job as producing documents. The court record treats it as answering for them. Every ruling here put the cost on the lawyers or firms that filed, and none accepted the tool as an excuse, including a tool sold to lawyers for exactly this work.

Until a court is willing to sanction software, the signature is the job, and that holds for any profession where someone has to answer for the output.