AI Answer Summary

Most lawyers using AI are not using a legal platform. The ABA's Law Technology Today, reporting the 2026 Legal Industry Report, puts personal use of general-purpose tools at 69% of surveyed legal professionals, against 34% of firms using legal-specific products.

On accuracy, peer-reviewed testing measured purpose-built legal research tools hallucinating on 17% to 33% of queries, and GPT-4 at 43%. Anything produced this way needs checking against a real source before it leaves your desk.

On privilege, a federal court held in February 2026 that a defendant's exchanges with a public AI tool were protected by neither attorney-client privilege nor work product, because the tool is not an attorney. The ruling is narrower than the headlines suggest, and the part most coverage omits matters more to practising lawyers than the holding itself.

What Lawyers Use: General-Purpose Tools, Not Legal Platforms

Most writing about AI for lawyers answers the question of which legal platform a firm should buy. Most lawyers are not using one.

The ABA's Law Technology Today, reporting the 2026 Legal Industry Report from 8am, found that 69% of 1,300 surveyed legal professionals personally use general-purpose AI platforms for work, more than double the 31% recorded a year earlier. At firm level, the same report has 46% using general AI platforms against 34% using tools built for legal work.

Our reading of that gap: the tool people adopt is the one they can start using this afternoon without asking anyone for a licence.

The last row explains the rest of the table. If you are reading this on a personal account your firm does not know about, the sections below are written for you.

For the firm-level version of this decision, our comparison of AI legal assistants by product type covers what each category does and who it suits.

What General-Purpose AI Does Well in Legal Work

These tools are good at reshaping material you give them and unreliable at supplying material you do not have.

That distinction does more work than any feature list, and it matches the pattern in independent benchmarking. Vals AI, with Legaltech Hub, tested legal AI products against a control group of practising lawyers and found the tools ahead on document questions, summarisation and transcript analysis, and behind on redlining. Reading beat amending across the board.

What that supports in daily practice:

• Summarising a document you already hold, then checking the summary against the document
• First drafts of routine correspondence, which you rewrite rather than send
• Restructuring an argument that is already yours
• Orienting yourself in an unfamiliar area before you verify anything in a real source
• Reformatting, converting and tidying, where mistakes are visible on sight

What it does not support is anything where the tool supplies the authority.

On the question this audience keeps asking, whether Claude or ChatGPT is better for legal work: no independent benchmark answers it. Both sit in the general-purpose class rather than having been tested against each other on legal tasks. Choose on the terms attached to your tier, which the section below covers, rather than on a comparison nobody has published.

Where AI for Lawyers Fails: What the Accuracy Research Shows

Tools built specifically for legal research hallucinate at rates that would end a career if the output went unchecked. General-purpose tools measured worse.

The source is Magesh and colleagues, published in the Journal of Empirical Legal Studies, the first preregistered evaluation of these products. Two limits worth holding: the testing ran in May 2024, and one of the products tested has since been retired and replaced.

A newer benchmark changes the comparison, not the conclusion. In October 2025, Vals AI tested three legal AI products and ChatGPT against a control group of practising lawyers on 200 general legal research questions, and found the AI tools, including ChatGPT, averaging about 80% accuracy against 71% for the lawyers. That study measured whether an answer was right; it excluded generating formatted citations, which is the failure behind the sanctions below. Read together, the two studies test different things. On answer accuracy, legal-specific tools and general-purpose models are now roughly tied. On citation generation, the cause of every sanction in this article, neither study measures it, so every citation still needs opening in a real source.

The consequence is personal rather than institutional. A public database of court decisions involving AI-fabricated material recorded 805 decisions attributable to lawyers out of 2,022 in total in early September 2026, and it counts only the ones a court noticed and wrote up. The name on those filings belongs to a practitioner.

So the rule is the old one arriving faster. Every citation gets opened in a real source before it is filed. Every proposition gets read back against the authority it claims. A tool that is wrong 17% of the time is useful. A tool that is wrong 17% of the time and trusted is a disciplinary matter waiting to happen.

Privilege After United States v. Heppner: What the Ruling Held

A federal court held that a defendant's exchanges with a public AI tool were not privileged. It did not hold that lawyers cannot use AI.

The facts are narrow and they matter. In United States v. Heppner, No. 25-cr-00503-JSR, Judge Rakoff of the Southern District of New York ruled from the bench on 10 February 2026 and issued a written opinion on 17 February. A criminal defendant, having received a grand jury subpoena and engaged counsel, used a consumer AI tool on his own initiative to generate 31 documents about his defence, then shared them with his lawyers. Federal agents seized them. The court held they were protected by neither attorney-client privilege nor the work product doctrine, describing the question as one of first impression nationwide.

Three strands of reasoning, each with a practical edge:

The tool is not an attorney. Because Claude is not an attorney, that alone disposed of the privilege claim. Rakoff noted that recognised privileges require a trusting human relationship with a licensed professional who owes fiduciary duties and is subject to discipline.

The privacy policy defeated confidentiality. The court pointed to terms under which the provider collects inputs and outputs, uses them to train, and reserves the right to disclose to third parties.

Purpose was the closer call. Rakoff found the defendant had not communicated with the tool for the purpose of obtaining legal advice, since he acted of his own volition without the suggestion or direction of counsel.

That third strand carries the line most coverage drops. Judge Rakoff wrote that had counsel directed Heppner to use Claude, the tool might arguably be said to have functioned in a manner akin to a highly trained professional who may act as a lawyer's agent within the protection of the attorney-client privilege. Analysts at Debevoise read that as pointing toward the Kovel doctrine, which extends privilege to non-lawyer professionals engaged by an attorney to assist in providing legal advice.

So the operative distinction is not AI against no AI. It is direction. A tool a client picks up alone sits outside the relationship. A tool counsel directs, on terms that support confidentiality, has an argument available to it. As McDermott's analysis puts it, the ruling applies settled principles about third-party disclosure to a new setting rather than rewriting doctrine, and the court's work product analysis turned on the same absence of counsel direction. The Harvard Law Review has criticised the opinion for drifting toward categorical exclusion where a fact-dependent analysis would serve better.

One caution against over-reading in either direction: Rakoff declined to follow an earlier magistrate judge's decision in this district, so the question is live rather than settled.

How to Use AI for Legal Work Without Creating a Problem

The tier you are on decides what you can put into the tool, and most lawyers have never read the terms of theirs.

Heppner turned partly on the terms of a consumer tier, which is the practical reason the distinction matters rather than an abstraction.

The obligations sitting on top are not new. ABA Formal Opinion 512 applies existing Model Rules to these tools. Competence includes understanding what the tool does and where it fails. The opinion recommends informed consent before client confidences go into a tool and says boilerplate language in an engagement letter is not adequate. And on hourly work you bill the time you spent, so working faster does not create billable hours.

Three things to do this week: find out which tier your firm has provisioned, read what its terms say about training and retention, and stop using a personal account for anything a client would recognise as their own information.

What Your Firm Needs to Decide

Individual discipline does not substitute for a firm policy, and plenty of firms still do not have one.

The decisions belong above your desk: which tools are approved, what data may go into each, who reviews output before it leaves the firm, what gets recorded, and who owns the decision when it needs revisiting.

One pattern worth raising if you are the person who ends up raising these things. Banning AI without provisioning an approved alternative pushes people onto personal accounts, which removes the firm's visibility completely. Read against Heppner, that is the outcome a firm should worry about most, since the terms attached to a personal tier are the ones nobody negotiated.

Our guide to AI for mid-market law firms sets out the firm-level version: what to automate first, what it costs, how long it takes, and what the policy needs to cover. Codebridge builds workflow automation for law firms, and that guide is where our own view sits.