The Advice I Would Give a Law Student Today

If I were beginning a law degree today and someone asked which additional competence could do most to improve my professional position five years from now, my answer would be quite simple:

Learn artificial intelligence properly.

I do not mean merely learning to write good prompts or using ChatGPT to summarise study notes.

I mean understanding modern AI well enough to build and use work systems around it: providing context, connecting tools, retrieving documents, preserving memory, defining permissions, checking results and determining when the machine must stop or request authorisation.

“Building” does not necessarily mean training a model or becoming an engineer. It may mean assembling and configuring existing components responsibly to solve a specific legal problem.

I believe one combination will be especially valuable over the next few years:

Law × applied technological competence.

It does not guarantee a job. Employability also depends on the market, languages, experience, grades and many other circumstances. But it can become a meaningful advantage.

Not because law firms need programmers disguised as lawyers, but because they will need legal professionals who can introduce technology into legal work without losing the law, confidentiality or professional responsibility along the way.

The gap between the partner and the engineer

Consider an experienced partner at a good law firm.

The partner may have negotiated hundreds of transactions, led complex litigation and developed a professional instinct that cannot be acquired by reading a textbook. But that person may not have time to study the difference between a model and an agent, test local systems, design document retrieval or discover why an automated workflow loses information after ten steps.

Not because the partner is incapable. Because there are clients, meetings, responsibilities and a business to run.

That partner does not necessarily need to become an AI engineer. The partner needs someone who can sit beside them and ask:

• How do you actually do this work?
• What do you look at first?
• Which documents do you need?
• Which exceptions concern you?
• Which parts can be delegated?
• What always requires review?
• What result would you consider sufficiently sound to deliver to a client?

The next step is to turn that professional knowledge into a process that other people and tools can execute, verify and improve.

An extraordinary opportunity may lie here for the next generation of legal professionals: acting as a bridge between those who understand the law deeply and those who build the technological infrastructure.

In July 2026, Reuters described how several associates were moving to legal-AI companies to work as legal engineers: hybrid professionals who bring legal knowledge to the design, testing and implementation of tools for legal work. This is a trend observed through reporting, not a general labour-market statistic, but it gives a name to a need that is already beginning to materialise.

The traditional entry point is beginning to change

For decades, law firms hired young lawyers partly because there was an enormous quantity of relatively straightforward work that someone had to perform:

• researching case law;
• reviewing documents;
• preparing chronologies;
• comparing contracts;
• extracting clauses;
• summarising case files;
• and preparing a first draft.

No one hired a recent graduate because that person possessed the judgement of a partner with twenty years’ experience. The graduate was hired in part for productive capacity. And while producing, the graduate learned.

Some of those tasks are being automated or transformed in certain settings. That does not prove that AI is generally eliminating junior positions. It does raise a difficult question: how will we train the lawyers of the future if part of the work through which they gained experience disappears?

In a 2026 article by Steven B. Katz published by the American Bar Association, the author poses precisely this problem. Document review, research, initial drafting and chronologies were not merely billable tasks: they also taught lawyers to read, organise facts, distinguish authority and learn from corrections. This is professional analysis, not an empirical study or a formal institutional position of the ABA.

In Norway, another signal deserves attention. The Norwegian Bar Association’s 2025 annual report states that the number of lawyers and trainee lawyers employed by firms fell in 2024, and that the decline continued in 2025, driven by fewer new advokatfullmektiger. The Association itself cautions that it is still too early to determine the causes. In the Secretary General’s message, she considers it plausible that increased use of AI may be contributing to the trend, but does not claim that the relationship has been established.

We should maintain the same caution. Task automation, reduced recruitment and job destruction are different phenomena. But it would be equally unwise to assume that the route into the profession will remain exactly as it is today.

The largest models are not always necessary

A few years ago, it might have seemed that this transformation would depend entirely on whether a law firm had access to the world’s most powerful models.

The reality is more interesting.

For many structured tasks, we do not need an intelligence superior to that of a partner. We need a system capable of extracting specific fields, comparing clauses against a checklist, classifying documents, proposing a chronology, detecting differences or working with a defined document collection. Another layer can then verify the output, and a lawyer can decide.

Downloadable models also expand the options for running certain operations within controlled infrastructure. In June 2026, Google presented Gemma 4 12B as a model designed to participate in local agentic experiences and to run on devices with around 16 GB of unified memory or VRAM, according to the provider’s specifications.

The model on its own is not an agent and does not use tools autonomously. The surrounding application—the harness—provides tools, context, permissions, memory and limits. And an execution will be genuinely local or offline only if the sources, connectors and tools it uses are local or offline as well.

That does not mean any laptop can solve any legal matter, that the model is suitable for legal Norwegian or Spanish, or that running it locally automatically makes it secure. Accuracy depends on the task, language, configuration, hardware, sources, licences and controls.

The lesson is different:

For repeatable work, architecture and verifiability can matter as much as the model’s raw power.

What it means to learn AI properly

Before long, saying “I know how to use ChatGPT” will probably do little to distinguish a candidate.

The advantage will lie one level above: in understanding the environment within which the machine works.

Basic use is already becoming normal. In Fremtidens Jurist 2025, 53.9% of Norwegian participants reported using AI tools daily or weekly. The figure combines 19.6% reporting daily use and 34.3% weekly use among 1,380 participants recruited through Lovdata customers; it should not be interpreted as a necessarily representative sample of all Norwegian legal professionals, nor as a measure of the depth of use.

If I were studying law today, I would try to understand at least five broad areas.

1. Models and agents

What a model does well, where it fails and when a single query is enough. And what changes when an agent receives an objective, chooses steps, uses tools and continues working with a degree of autonomy.

2. Sources, retrieval and memory

How to provide access to specific documents, retrieve only the relevant information and preserve facts or decisions across sessions. A retrieval system does not by itself guarantee a correct answer: sources must be cited and checked.

3. Tools and automation

How to connect AI to bounded tasks: converting documents, extracting data, consulting a register, completing a template or comparing versions. The question is not only what the system can do, but what it is authorised to do.

4. Controls and evaluation

Permissions, limits, logs, tests, deterministic rules and mandatory human-approval points. Also, how to measure whether the system works and identify the circumstances in which it fails.

5. Privacy and data architecture

What information may leave the organisation for an external provider, what should remain local, how to minimise or pseudonymise data before transmission and why confidentiality must be resolved before the system is connected.

It is not necessary to master these subjects to the level of an engineer. It is necessary to understand them well enough to apply them to legal problems and to hold a rigorous conversation with the person who does build the infrastructure.

Technological competence also includes knowing when not to automate.

Your advantage will not be technological if you do not know the law

I would add an important warning here.

I would not sacrifice law to learn technology. I would try to learn law better, because technology increases the value of the judgement capable of directing it.

A mediocre legal professional who knows how to build agents may simply end up automating mistakes.

To explain to a machine what it must do, you must first understand the problem. To evaluate the output, you must know the subject. To identify which exception matters, you must have experience or work with someone who does. And to assume professional responsibility one day, you must first become a professional.

The formula is therefore not AI instead of law.

It is law × technology.

The multiplication matters. If either factor approaches zero, so does the result.

If I were twenty-two

I would complete my law degree determined to become a better lawyer than my peers. But at the same time, I would build.

Not twenty course certificates.

Three verifiable, well-documented prototypes.

For example:

1. A system for comparing contracts against defined criteria.
2. Another for researching a problem using verifiable sources.
3. A third for organising and retrieving legal knowledge with traceability.

I would use only public, synthetic or expressly authorised data. I would explain the problem each prototype solves, what lies outside its scope, where it fails, what information it uses, how it verifies its conclusions and at what point it requires human review.

I would not try to conceal the errors. Documenting them would demonstrate a better understanding of the problem than a flawless demonstration prepared for an interview.

The conversation with a law firm would then be different.

It would no longer be only:

“I am a law graduate and I want you to train me.”

It could be:

“I am legally trained, and I have learned to build with the technology that will transform part of your work. You have the knowledge, the clients and the experience. I can help you turn some of that into safe, verifiable systems.”

That is a different value proposition. It does not promise to replace anyone already working at the firm. It offers to collaborate in redesigning how the firm works.

Training is not solely the student’s responsibility

It would be unfair to place the whole problem on those entering the profession.

Universities must teach the responsible, practical use of these tools. And firms that automate junior tasks will have to create new learning pathways: bring junior lawyers into meetings, strategy and review earlier; teach them to audit outputs; require them to open the sources; and evaluate not only what they produce, but the reasoning by which they justify a decision.

The Thomson Reuters 2026 Law Student Pulse Survey, based on 1,874 responses collected online in April 2026, captures the tension well. 72% regarded AI literacy as an essential professional skill, while 74% recognised that excessive reliance on it could weaken their legal skills. 32% disagreed that their law school was providing the AI skills they would need for their future career.

It is a survey, not a universal picture of every law school. But its participants appear to understand something important: learning AI and protecting legal education are not opposing goals.

The final paradox

Artificial intelligence may reduce precisely some of the work that, for decades, justified hiring junior lawyers.

At the same time, it creates a new need within law firms.

Someone has to understand the technology, implement it, evaluate it and speak both the partner’s language and the engineer’s. Someone has to turn years of professional experience into procedures, tools, memory and controls without destroying what made that experience valuable.

A law student may enter a profession full of people who know far more law than they do. But that student can arrive understanding a technology that many organisations are still learning to integrate.

That combination may be worth a great deal.

Because in a few years, knowing how to use artificial intelligence will probably be normal. The opportunity lies in learning first how to apply it with judgement.

Learn law.

Learn technology.

And become the person capable of bringing the two together.

Scope note

This article offers professional advice and draws several inferences about the evolution of the market. The cited data do not prove that AI is causing a general reduction in junior positions, nor do they guarantee that any particular technological competence will lead to employment. Confidential or client information must not be used in training projects without authorisation and a suitable architecture.