What AIs should I get for law school?
As students begin to head back to law school, a question confronting them should be, what AI should I be purchasing. It's an important decision. AI should be at the forefront of their legal education; getting the right AIs is important. Fortunately, I think there is a clear answer. Students should spend $20 per month on the paid "Plus" version of ChatGPT. That will get you extensive access to the new GPT-5, which is outstanding. And students should, if at all possible, get the free-for-students version of Google Gemini Advanced. That will get you good access to the Gemini Pro model, NotebookLM and other features (like 2 TB of storage on Google Drive). If they can't get it for free, they should spend another $20 per month and just pay for their addiction now, just as they will a year from now when Google's generous offer comes to an end. You will be assimilated!
And that's it. Unless you have some special reason to do so or an unusually large budget, do not pay for Claude*, do not pay for Grok, do not pay for Perplexity, and certainly do not purchase specialized proprietary AIs for "students" likely populating your Instagram feed. Most are inferior, limited, and readily replicated in the general purpose AIs discussed here. You can use the AI capabilities of Westlaw and Lexis – sometimes they are invaluable – but recognize that they are not general purpose models and are often vastly inferior to the current frontier models.
- If you are going to be doing more than occasional but less than intensive coding, a paid $20/month subscription to Claude gets you access to Claude Code, which is fabulous. If you become obsessed with coding as a full time occupation, you'll need the $100/month plan to do the most intensive work.
Don't Cheap Out
Many of the complaints I hear from faculty or students about AI come from those using only free models. And because they're unhappy with the free models, they don't pony up for the paid ones. Falling into that cycle is a mistake. Free AI is definitely capable of summarization and basic retrieval. Indeed, it often does quite a good job. You are already likely paying tens of thousands a year for law school and, at many schools, the difference between being toward the top of the class and being toward the bottom can means hundreds of thousands of dollars in lifetime earnings. In a competitive world, you don't want to be relegated to technology that is so early 2025 (ChatGPT 4o or even o3). Moreover, because of the high cost of failure in the law world, "often" is not good enough. Your professor and your employer will increasingly expect something close to perfection. The latest models provide more incisive analysis, more coherent prose, and fewer hallucinations. (OpenAI claims they devoted substantial resources to reducing hallucinations in ChatGPT5 but I have not verified whether the improvements extend to law). Paying $20 per month for the vast capabilities of modern AI is an incredible bargain. Plus, for some it will be more entertaining than Netflix, which you likely don't have time for anyway.
Get Two
Can you get away with only one of these top AIs, particularly since Gemini is likely to be free? Well, yes. But let me tell you why you should get both. Getting two AI opinions is better than one. Have one AI do an analysis and feed the output into the other AI. Ask the second AI ruthlessly to critique the first and then improve the product. Iterate by telling the first AI that a "friend" set forth the copy-and-pasted criticisms and a revised version. Ask the first AI ruthlessly to critique the work of the second and to improve the product. Repeat the dialectic until the result shines. True, you can emulate this iterative development with a single AI by prompting them to assume different personas, but having two AIs with genuinely different "personalities" and training data is likely to avoid the tunnel vision to which AIs can be prone.
In addition to the benefits of synergy and dialog, having two models harnesses the principle of comparative advantage. You may find that one model is consistently better at certain kinds of tasks than the other. For example, right now Gemini is better at producing long-form content than is GPT-5, although the latter is starting to come close. GPT-5, however, has "agentic workflow" built in, which lets you create multi-stage prompts that involve gathering and analyzing information and then producing derivative products such as PowerPoint presentations. No need to go through the steps I outlined in a prior blog post. Although Google has an experimental agentic product (Opal) that may actually be better than GPT-5's implementation, it currently requires a little more technical skill to use it properly and is only experimental anyway.
Chat GPT-5's Strengths
GPT-5 from OpenAI is the standard. It's like buying an iPhone. You are never going to go badly wrong by picking the world's leading AI. And it's a leader for good reason. It has the now-standard features of modern consumer AIs such as web search, multi-modality (input images instead of text), conversation history search, and a "Canvas" mode that acts as an intelligent editor of text.
But, as they say, there's more. GPT-5 comes with a panoply of features likely to be useful in legal education. GPT-5 has Study and Learn mode that acts like a (very) soft Socratic professor guiding you through material that you select. Instead of just telling you the answers and hoping that you will passively absorb the result, Study and Learn mode asks you to participate by answering questions that test your understanding.
It lets you easily create CustomGPTs that let you combine a standard prompt with a library of materials and obtain high-quality grounded results. By "easily," I mean that anyone admitted to law school should be able to handle the intuitive interface with ease. Should you feel so inclined, you can share your GPT with the world and put it on your resume. Moreover, the world of CustomGPTs is large and searchable so there's often no need to reinvent the wheel. There are numerous writing aids such as a specialized Grammar checker, a Bluebooker, an editor that favors active voice, or an editor that attempts to "humanize" AI-drafted prose excessively littered with "tapestry" and "meticulous." (This latter capability is one of many reasons that software that attempts to detect unauthorized use of AI in education is likely to fail). There are research aids such as "Scholar GPT," for example that searches through hundreds of millions of scholarly documents looking for matches to your query.
GPT-5 likewise supports extended research and thinking. It has both a Deep Research mode that spends extra time scouring the web for materials relevant to your query and a Think Longer feature integrated into GPT-5 that devotes extra resources to "reasoning" about a particular query.
There are three other features of note. First, there's Voice Mode that lets you use conversation instead of typing and reading to communicate information. It is still not perfect but it is now pretty darned good. You can even simulate job interviews, oral argument, or witness interviews using this feature, Second, for occasional artwork, GPT-5 is outstanding. I use it to generate the feature images for my blog. And third, there is a new Agent Mode that lies on the frontier of AI capabilities. Agent Mode in GPT-5 is designed to let the model behave more like a persistent, semi-autonomous research or drafting assistant rather than a one-shot Q&A engine. Instead of waiting passively for your next prompt, it can be given a broad or multi-step mission and then actively plan, gather information, and execute subtasks on its own—calling tools, running searches, retrieving and reading documents, and iterating without you having to break the request into small pieces manually. Law professors (or students) can imagine creating a virtual seminar around their article draft: the AI searches out the leading experts, simulates their interests and biases and then does a detailed critical review of your document.
Gemini's Strengths
Gemini likewise handles the basics brilliantly and has great features. Its Gemini Pro model – paying customers can stay away from the speedier but decidedly inferior Gemini Flash – is one of the two best models currently out there for law. Having been trained on the vast case law contents of Google Scholar and the gargantuan training grounds available to Google, Gemini Pro usually provides solid, sensible, and hallucination-minimal responses to the sort of questions law students are likely to encounter. Moreover, its user interface enhancements make it extra useful. Its new study and learn mode is a great way to be interactive with legal material. Its Deep Research mode is the best current consumer AI at producing long form responses to legal questions. And the responses can be readily exported to Google Docs or slurped into NotebookLM. DeepThink lets the LLM mull over the subtleties of the arguments you or your opponent is advancing in some brief or lets you think collaboratively with "someone really smart" on legal issues you find perplexing or troubling. Although this last feature is currently available only to Ultra subscribers paying a monthly fortune, I am confident at least limited use will shortly trickle down to those at lower subscription tiers. And even without DeepThink, Gemini is smart enough to engage as a collaborator or pretend adversary in honing your arguments.
Second, getting essentially unlimited access to Google's NotebookLM is worth the price of admission alone (particularly if that price for students is zero!). As discussed in a prior blog entry, NotebookLM can serve as a great organizational device for law school. Set up a notebook for various units of your courses: a notebook for Intentional Torts, a notebook for Erie Doctrine, a notebook for the Commerce Clause. Put your own materials into a notebook, use the "discovery feature" to have NotebookLM search for more, and then start harnessing the Gemini engine to produce briefing documents, quizzes, synthetic podcasts, timelines, and more. Plus, you can get grounded answers to custom questions that can then be fed back into NotebookLM to make its future responses even better. NotebookLM documents are easily shareable so students can form consortia and professors can provide them to students.
Third, Gemini Gems include prebuilt Gems such as the Storybook one featured in a prior post, a Career Builder gem that interactively guides your efforts to land a better job, and a Learning Coach that serves an interactive tutor. You are not limited to pre-built Gems. Gemini lets you build the equivalent of CustomGPTs, i.e. repeatable complex prompts grounded by material you provide. You can generate multiple choice question generators, ask Gemini to help you prepare for class like a law professor, or create an expandable compendium of resources that then provide grounded answers to whatever question you have in mind.
Fourth, coming down the road I suspect is preferential access to an experimental product called Opal. It's an implementation of "Agentic AI" in which the AI engages in a multistep process that may involve gathering information, taking action in the real world, and producing reports, presentations and other materials. I've been playing around with it for the past couple of weeks. I'm impressed. Expect a blog post about it and Agent Mode in GPT-5 in the next month or so.
Why not this alternative
Why not Claude?
Claude is good, don't get me wrong. And, having been built by the pioneers of a connectivity technology known as MCP, it is particularly strong these days as a chat-based front end into a number of services such as Google Calendar, GMail, Notion and various platforms favored by code developers. This strength in connectivity leaves it well positioned to be a great chat-based front end into legal databases at some point in the future. Moreover, Claude has many of the convenience features of GPT-5 and Gemini: a Canvas (which it calls an Artifact) for editing, searchable conversation histories, "Projects," and more. For now, however, Claude is, in my experience, behind ChatGPT and Gemini for the most common AI-appropriate in law school. It lacks certain features such as study and learn mode or use of "Gems" or Custom GPTs that the two flagships now contain. More fundamentally, however, most of the time, Claude's answers just aren't quite as good in the legal arena. So, if you have some extra funds or want to experiment, Claude is hardly a mistake. But it should not be your first choice.
Why not Meta.ai?
Meta.ai is Facebook/Meta's model. Let me succinct. For anything serious, it stinks. Shallow answers to legal questions often with hallucinations thrown in. Very 2023. There's a reason Mark Zuckerberg is scouring the planet and throwing eight-figure salaries at leaders in AI to get them to come work for him. The product his existing developers built is lousy. I will concede two strengths. It's free at least if you have an account within the meta family. And you get your bad answers fast.
Why not Grok?
Because it just isn't as good. Grok3 was crummy and Grok4, though vastly improved, is still a step behind the leading two AIs in the quality of answers it tends to provide in response to legal questions. Plus it lacks some of the nice features of Gemini and GPT-5: no ability to generate or search for customized helpers, for example. If it were super fast, I might forgive these lacunae, but it is not. Yes, it is a bit less "censored" than other AIs , as one might expect from a product by X, and it is very good at math and coding. And, again, if it were 2024, Grok4 would be seen as miraculous. But it's 2025 and the market has moved. Grok4 is a runner up.
Why not the Chinese Models?
This is a very good question. There are a host of models coming out of China that are looking very good. DeepSeek, Kimi K2, Qwen, and the latest entry, GLM-4.5 from Z would again all have been state of the art six months ago. And they are all free for basic use. But there are several reasons not to make them your "go to" for work in legal education. First, perhaps it is naivete or my susceptibility to ambient bias, but I do worry a bit about privacy with Chinese models in a way that I do not for models maintained by American companies such as OpenAI an Gemini. And while it's true that most of what I do is exceedingly dull, occasionally I do ask AIs for assistance on matters that I would prefer large government not know about. You may have similar attributes. You should also check to make sure your state or university does not prohibit you from using these models, particularly when doing so using state-owned equipment.
Second, these models currently lack some key features. For example, they don't have CustomGPTs or the equivalent, they don't have voice mode, they don't have Study and Learn. And finally, outside of math and coding, where some of them really do excel (GLM-4.5 in particular), I do not find their responses, on average, to be quite as good as the flagship models I am recommending.
A year from now I may be changing my tune. The models from China are evolving at a rapid pace. Chinese researchers seem to be making advances in self-evolving AI, in which AI is used to evolve itself. (Scary!) Given sufficient privacy assurances and some feature additions that may not be too difficult such as Study and Learn, several of the models out of China may fairly soon overtake American models. But for now, I would use these as third opinions or as backups. Don't stake your law school career right now to these models.
Why not Local Models?
Some in the legal community might want to run large language models locally on their own machines rather than on the cloud-based services I have been discussing. Local models guarantee privacy in a way that no cloud-based service can. Your communications never leave the safety of your own storage devices. And with programs such as Ollama, installing and using a local LLM has become relatively painless in a way that I could not claim even a few months ago. Moreover, for many purposes a local LLM will be powerful enough. I would trust them with a recipe for Kung Pao chicken or even a basic recitation about the dormant commerce clause. But to get a model with enough sophistication to give you high quality responses to more challenging legal questions, you need a machine that is more powerful than most of those owned by those in the legal community. Even the best local models that fit on most consumer hardware such as the new GPT-oss 20B hallucinate on legal questions in a way that seems like 2023 all over again. Responses also tend to be slower than those provided by commercial providers. And while more expensive computers with gobs of memory can host more sophisticated local models such as the better versions of Llama, Qwen, or GPT-oss 120B, even these models do not, in my experience, quite compete with those you can get for $20 per month.
So, if you have a particular need for privacy or you want a backup in case the Internet is unavailable, these local models are worth considering. For most in the legal community, however, commercial providers should be the first choice.
The Bottom Line
If you are serious about succeeding in law school, invest in the best tools from day one. Net month I might have a somewhat different answer, but right now, that means (1) paying $20/month for ChatGPT Plus to get GPT-5, and (2) grabbing the free-for-students version of Google Gemini Advanced while it lasts—or paying for it if you must. These two models complement each other: GPT-5’s agentic workflows, CustomGPTs, and study modes pair beautifully with Gemini’s deep research, NotebookLM integration, and long-form drafting strengths. Everything else—Claude, Grok, Meta.ai, Chinese models, and local models—has its place, but for most law students they are optional extras, not core tools. The modest monthly cost is trivial compared to the potential upside in grades, skills, and career opportunities. The cost of a bad answer in law can be huge; the cost of a good AI is small.
Note: Want to gamble $10?
There's one more dark horse I want to talk about. Midpage. You can think of it as kind of like Lexis or Westlaw but with ChatGPT as the orchestrating AI. The image below shows some of its capabilities. Right now, law students and law faculty can get it for $10 per month. I looked at an earlier version of Midpage and found it promising but limited. Development on the product has accelerated, however, and at some point soon I want to give it a second look. For the same price as a modern fast food combo meal, and a product with considerably less fat, you too can try it for a month. Send me a note at aiforlegaled at gmail dot com if you give it a try.
