In 2023 a federal judge in the Southern District of New York sanctioned two lawyers whose brief cited judicial opinions that did not exist. The citations came from a general-purpose chatbot, the lawyers did not verify them, and when opposing counsel could not find the cases, they doubled down. The court in Mata v. Avianca imposed monetary sanctions and wrote an opinion that is now assigned reading in professional responsibility courses. The following year the Second Circuit referred another attorney to its grievance panel in Park v. Kim over a fabricated citation.
Those cases are the honest starting point for any discussion of legal technology, because they show exactly where the value and the risk sit. The tools are genuinely capable at drafting, summarizing, and searching. They are unreliable at the one thing lawyers are paid for, which is being accountable for whether a statement is true. Everything useful in this field follows from taking that division seriously. Consumer document platforms such as Lawdistrict illustrate the same split at the retail level: automation can assemble a competent lease or will from structured inputs, but it cannot tell you whether your situation is the one where the template is wrong.
Where the Technology Is Actually Delivering
Strip away the conference-stage language and a short list of applications is producing measurable results in practice today.
Discovery and document review
This is the oldest and most proven use, and it predates the current wave by a decade. Technology assisted review, sometimes called predictive coding, trains a model on attorney-coded samples and ranks the remaining corpus by likely responsiveness. Courts have accepted it since Magistrate Judge Andrew Peck approved its use in Da Silva Moore v. Publicis Groupe in 2012, with later decisions such as Rio Tinto v. Vale reinforcing that the question is no longer whether TAR is acceptable but how it is validated. The 2015 amendments to Federal Rule of Civil Procedure 26(b)(1), which put proportionality at the center of discovery scope, made cost-effective review a rules issue rather than merely a budget issue.
Research and first-draft production
Retrieval-based research tools that ground answers in an actual case database, cite to the source, and let the lawyer open the opinion are meaningfully different from a general chatbot answering from memory. The productivity gain is real for first passes, issue spotting, and summarizing a long record. The verification obligation does not shrink; it just moves later in the workflow.
Document automation and intake
Assembly systems that generate documents from structured client data eliminate the transcription errors that cause a large share of malpractice claims, and they enforce version control on clause libraries. Intake automation, conflict checking, and deadline calculation are unglamorous and reliably profitable, and they carry far less ethical risk than generative drafting.
Evidence that is itself digital
The bigger shift may be on the evidence side rather than the practice side. Telematics, event data recorders, wearables, smart home logs, and driver assistance system records now generate contemporaneous data about what happened, which changes how liability is proven. Analysis of what that data shows on the roads, covered in Speeding and Safe Driving Tech, is a useful illustration of how quickly reconstruction arguments give way to logged facts. Federal Rules of Evidence 902(13) and 902(14), added in 2017, allow certain electronic records and hash-verified copies to be self-authenticated by certification, which removed a practical barrier to getting that data in front of a jury.
The Ethics Rules That Already Cover This
No new rulebook was needed. The existing professional conduct rules reach generative tools directly, and the ABA Standing Committee on Ethics and Professional Responsibility said so in Formal Opinion 512, issued in 2024, which addresses generative AI across competence, confidentiality, communication, fees, and supervision.
- Competence, Model Rule 1.1. Comment 8 has required lawyers to keep abreast of the benefits and risks of relevant technology since 2012, and the large majority of states have adopted some version of it. Competence now includes understanding that a language model can produce fluent, confident, false output.
- Confidentiality, Model Rule 1.6. Entering client information into a consumer tool whose terms permit training on submitted data is a disclosure question, not a preference. Enterprise agreements with no-training terms and data residency commitments exist for this reason.
- Communication, Model Rule 1.4. Whether clients must be told that AI was used depends on circumstances, including whether the client’s information is being submitted to a third-party tool and whether the use is material to the representation.
- Fees, Model Rule 1.5. A lawyer cannot bill an hour for work that took six minutes because a tool did it. Efficiency gains belong to the client under hourly billing, which is one of the quieter arguments for flat and fixed fee structures.
- Supervision, Model Rules 5.1 and 5.3. A firm deploying these tools owes the same supervisory duty it owes a junior associate or a vendor, which means written policies, training, and an actual verification step rather than an assumption that someone checked.
Courts have added their own layer. A number of federal judges issued standing orders beginning in 2023 requiring counsel to disclose the use of generative AI and to certify that any AI-assisted text was verified by a human. Check the judge’s standing order and the local rules before filing; the requirement is not uniform and the consequences of missing it are not trivial.
Matching the Tool to the Task
The practical question is not whether to use these tools but which tasks tolerate their failure modes. A model that is wrong ten percent of the time is a gift on a first draft and a catastrophe on a citation table.
| Task | Fit today | Required human step |
|---|---|---|
| Summarizing a deposition or long record | Strong | Spot-check against the transcript for omissions and reversed meaning |
| First-draft correspondence and internal memos | Strong | Standard editorial review before anything leaves the firm |
| Contract clause extraction and comparison | Strong | Verify against the executed document; models miss defined-term interactions |
| Privilege and responsiveness review at scale | Proven, with validation protocols | Statistical sampling and a documented validation methodology |
| Legal research in a grounded, citator-linked platform | Good | Open every cited authority and confirm it says what the summary claims |
| Legal research in a general chatbot | Poor | Do not file it; this is the Mata v. Avianca failure mode |
| Predicting case outcomes or judge behavior | Limited | Treat as one input among many, never as advice to a client |
| Advising a client on a novel question | Not suitable | The judgment is the service; the tool cannot hold the responsibility |
Blockchain, Smart Contracts, and E-Signature: Separating the Real From the Pitch
Electronic signature is settled law and unremarkable in daily practice. The federal ESIGN Act and the Uniform Electronic Transactions Act, adopted in nearly every state, give electronic signatures and records the same legal effect as paper for most transactions, with defined exceptions including wills and certain family law and notice documents.
Smart contracts are a narrower story than the marketing suggests. Several states, including Arizona, Tennessee, and Wyoming, enacted statutes recognizing blockchain signatures and smart contract enforceability. What those statutes do is confirm that code-based performance is not void for being code. What they do not do is resolve the hard questions: what happens when the code executes correctly but produces a result neither party intended, how mistake and impracticability doctrines apply, who has jurisdiction over a decentralized system, and how a court unwinds an irreversible transfer. Self-executing performance works well for narrow, objectively verifiable conditions such as escrow release on a confirmed event. It works poorly for anything requiring interpretation, which is most of contract law.
Regulation Is Arriving Unevenly
The European Union’s AI Act entered into force in 2024 and takes effect in phases. It classifies AI systems intended to assist judicial authorities in researching and interpreting facts and law as high-risk, which triggers obligations around risk management, data governance, human oversight, and documentation. Firms with EU clients or EU operations should treat that classification as a procurement question now rather than a compliance question later.
In the United States, the action is at the state level and in the courts rather than in a single federal statute. Utah’s regulatory sandbox for legal services innovation and Arizona’s decision to permit alternative business structures both loosened the traditional bar on nonlawyer ownership, creating room for technology-led service models that Rule 5.4 forecloses elsewhere. Meanwhile state supreme courts, bar associations, and individual judges are issuing guidance at different speeds and reaching different conclusions, so the applicable rule genuinely depends on where you practice.
The Access to Justice Question
The most important argument for legal technology is not efficiency for firms. Research from the Legal Services Corporation has consistently found that the large majority of civil legal problems faced by low-income Americans receive no or insufficient legal help, and the gap is not closing through pro bono hours alone. Guided interviews, plain-language form assembly, court self-help portals, and online dispute resolution platforms now used for small claims and some family matters in several states reach people who were never going to hire counsel.
The caution is equally real. A tool that helps someone file a competent answer is valuable; a tool that gives a confidently wrong answer to a person with no way to evaluate it is worse than nothing, because it produces a waived defense or a missed deadline that no one can fix later. For a grounded look at where self-help genuinely works, see how technology helps you respond to a lawsuit alone.
Frequently Asked Questions
Can a lawyer use AI without telling the client?
Sometimes, and it depends on the use. ABA Formal Opinion 512 frames disclosure through Model Rule 1.4: if the use is material to the representation, or if client confidential information will be submitted to a third-party tool, the client generally should be informed and in some cases must consent. Routine internal use of a self-contained tool is treated differently from feeding a client’s documents to an outside service.
Will AI replace lawyers?
Not on current evidence, but it is redistributing work. Tasks that are volume-heavy and verification-light are moving first, which compresses the traditional junior associate training path and raises a real question about how the next generation develops judgment. What does not transfer is accountability. A model cannot be sanctioned, cannot owe a fiduciary duty, and cannot appear before a tribunal.
Is it safe to put client documents into a chatbot?
Not into a consumer tool whose terms allow the provider to train on your inputs. Model Rule 1.6 governs, and the practical answer is an enterprise agreement with contractual no-training terms, access controls, encryption, and defined data retention. Firms should also read ABA Formal Opinions 477R and 483 on securing communications and responding to a data breach.
Are smart contracts legally enforceable?
In several states, statutes confirm that a contract is not unenforceable merely because it is executed through a blockchain or contains a smart contract term. That is narrower than it sounds. Traditional doctrines of formation, mistake, unconscionability, and remedy still apply, and code that executes irreversibly can create disputes courts must resolve after the fact.
What should a small firm adopt first?
Not generative AI. Start with the operational layer: a document management system with real version control, conflict checking, calendaring with automated deadline calculation, secure client communication, and multi-factor authentication everywhere. Those reduce malpractice exposure immediately and carry almost no ethical risk. Add drafting tools afterward, with a written verification policy in place before the first use.
What to Do Next
Write the policy before you buy the tool. A one-page firm standard that names which tools are approved, what categories of client information may never be entered into them, who verifies AI-assisted work before it leaves the firm, and how the use is documented will prevent more problems than any product evaluation. Then check the standing orders of the judges you appear before, because the disclosure requirements are already in effect and are not uniform. For more practical coverage of legal practice and technology, browse the Legal advice section.
This article is general information about legal technology and professional responsibility and is not legal advice; consult your jurisdiction’s rules and a qualified attorney for guidance.
