Mike: open-source legal AI
mikeoss.com
mikeoss.com
What legal professionals actually pay for, and that is virtually impossible to replicate unfortunately, is to give the AI access to a legal database of case law. Without case law, you can't do accurate legal research, and you are inviting disaster if you're doing things like drafting statements of case or skeleton arguments.
There's a reason why companies like Thomson Reuters have an oligopoly on these types of products, and can get away with charging thousands a year. They are the only ones with access to a comprehensive set of case law, and they've entrenched their position by having exclusive contracts with the law reporting companies. Without that, your model is just relying on publicly available cases that it can find on Google etc., and that's just a fraction of the full set.
With that said, these types of competitor products can be useful if you're just doing simple tasks like drafting letters or reviewing contracts and you accept that you need to do the legal research separately. But again, you can get that with just ChatGPT + a good prompt.
I'm not in the legal field, but can someone explain that further? I would have expected that all case law is public access. Not necessarily easy access, but when a judge writes an opinion, why on Earth would that opinion be gated behind a corporation? What am I missing?
To access the DB through the modern archive (well modern as new rules) you'd have to be an accredited professional passing through a few legal hardles and digital chancellor's office for each copy. It's like going to a bureaucracy^bureaucracy office.
Some early companies given their initial foothold were not required these checks so they were able to get hold of bigger archives (it's also important to remember often legislation or conformity is done through consulting or lobbying done by these entrenched players).
They can also build on the Data professionals themselves submit.
It doesn't work that way in the US. Legal judgements are documents of public record, and they are normally published in full - it's not uncommon to search for someone's name and see legal cases pop up that they have been involved in.
There are specific instances where a judge can seal a judicial record (and records for minors are sealed automatically), someone may petition for their own records to be expunged, and parties may ask for some information to be redacted, but these are normally (except in the case of minors) not done automatically. As I understand it, the US has much more lax rules around the publicity of legal proceedings than other jurisdictions. For example, even though someone is deemed "innocent until proven guilty", arrest records and mugshots are reported all the time in the media even though this would be illegal in many other areas of the world.
In theory, any member of the public can obtain a judgment by applying for one at the court and paying a fee. That's fine if you just need a one-off judgment, don't mind paying the fee, and you're not in a hurry. It also assumes that you know which case you need.
For realistic legal research, you might need to wade through dozens of cases just to even know if any of them are relevant, you might have a deadline of tomorrow to get it done, and you might not want to pay that fee for a bunch of cases that you aren't going to end up needing. Only a company which already has a comprehensive copy of virtually every important case can help you here.
A typical workflow for a complex piece of legal research might look like this:
1. You need to research a legal topic.
2. Do some Googling, or chat to your LLM, to get a rough overview and some pointers for further research (but don't completely rely on what you find).
3. Read some professional content (e.g. Practical Law articles relevant to the topic, or a legal textbook).
4. Read the relevant legislation.
5. Use a legal database to download all the cases you found from steps 2 and 3 which seem like they might be relevant.
6. Use a legal database to download all the cases which cite the relevant legislative provisions you found in step 4 and seem like they might be relevant.
7. Use the legal database to confirm that those cases are still good law (not overridden or criticised by a later case).
8. Skim read them, discard those that turned out to obviously not be relevant.
9. Read the remaining ones more closely.
10. Note any useful-looking cases which are cited in the ones from step 9, and recursively work your way through those cases as well.
Relying on court-provided copies of judgments won't realistically help you with most of these steps.
If tech companies invested 10% of what they have in AI assisted coding tools into AI assisted legal tools, they would be able to do those steps easily.
It is definitely coming.
It's obviously even more important for judges (compared to lawyers) to be able to easily search all of the relevant case law to see which cases are controlling and would have precedence. Seems bizarre to me that this critical function would be gated behind a corporation.
I'd not use generative AI for anything but a cursory check anyway⁰. Even if it is trained on clean up-to-date data rather than all the wrong information that is out there, it could still give a wrong answer and I have no leg to stand on if I rely upon it. At least if I pay a human and they trust the LLM too much, I'll hopefully have some call to pursue them for giving bad advice when it bites me.
--------
[0] Or at all… But even if I wasn't someone actively avoiding LLMs, the point would still stand
Theoretically speaking if someone scraped all of it and added it to something like this open source Mike project would that then be a much better tool for lawyers?
Better than before, yes. Good for general legal work that doesn't require robust legal research, yes. Sufficient for full legal research, no.
The problem is that "a lot of case law" isn't enough case law. You need close to everything. Otherwise this can happen: Canlii case X -> Legal principle Y. Westlaw case Z not on Canlii -> X overriden, Y no longer good law. Or you might simply not find a case which cogently supports your argument, when one does in fact exist. Or, conversely, you are unaware of a detrimental case which your opponent knows about because they have Westlaw.
And this is not likely to change in the future, as the legal market is so small and niche that the leading makers of LLMs have put legal analysis near the bottom of their list of priorities in terms of improving model performance. There is very little if any effort by the major LLM companies to curate sources of additional high-quality legal training data or fine-tune their models to improve performance on legal tasks. Law is also a field with very low tolerance for error, where tiny mistakes can have big consequences, and getting the models to perform well under these constraints would require a lot of investment without a sufficient payoff.
The true reason big firms are buying Harvey and Legora subscriptions is simply to use an LLM for LLM-type tasks, like document review, spotting issues in user-provided documents, and other things that LLMs do well. True, services like Harvey and Legora have lots of cool templates and features for legal work, but you will find that most of the people who use these services in these firms use them much the same way they'd use ChatGPT, Claude, or any other AI chatbot.
The reason law firms can't just use ChatGPT or Claude is that they can't allow confidential or privileged client data (such as documents provided through the prompts) to be stored and hosted on a third party service like ChatGPT or Claude, as these companies may have to turn over client data in response to subpoenas, and depending on the type of LLM account you have, these companies could use your user prompts to train future models thus risking leakage of client data to third parties and potential privilege waiver.
Services like Harvey and Legora solve this problem by accessing the LLMs through APIs, and all client data, prompts/responses, etc., are stored encrypted on Harvey or Legora servers and protected by keys held by the customer. For many law firms, this is 95% of what they're paying for.
The big challenge "Mike" presents to services like Harvey and Legora is that it exposes how little additional value they offer over ChatGPT or Claude, for the vast majority of law firms. A system like "Mike" can provide the same security benefit at basically $0 cost, and can be hosted on the law firm's own internal servers. This is going to put a lot of pricing pressure on services like Harvey and Legora; law firms are notoriously cheap when it comes to IT and software spend and will switch quickly if cheaper alternatives arise. This confirms that Harvey and Legora are going to have to sell their services based on the value they add to lawyer productivity, and not just on being a protected wrapper around GPT or Claude.
One question I have about legal AI startups/products, is how do they maintain or improve upon billing practices of law firms?
Having worked with a bunch of lawyers, I know that I'm often paying $500/hr to that firm. That work is actually done by a paralegal who is being paid $40/hr, and then I'm being billed through the partner for an extra $460/hr. This is a gross oversimplification, but you get the point.
If the partner needs to bring in $5M a year, how does any addition of tech solve that?
If I'm the customer of the law firm, I would love to have a more cost efficient way to get legal advice. But, I don't understand how those incentives are matched by the partner? I don't really think they want a more efficient result for their customers, they want a better way to get more billable hours. Adding "tech efficiency solutions" does not solve that issue at all.
Inevitably, customers will use LLMs on their own, and as people have noted, lose attorney client privilege (and often get hallucinated bad advice). There will probably be some very comical court room dramas when people try to represent themselves with an LLM on their shoulder.
Am I misunderstanding something fundamental about the legal world that will make a major law firm adopt this tech? I feel like there are some strong reasons they will universally avoid moving in this direction. Long term it will win and there will be blood on the floor, but why would any large firm adopt this stuff right now?
I just don't understand how decision makers at a big firm are going to say yes to tech solutions when those solutions will kill the goose roaming their hunting grounds.
Also, an interesting example: in English litigation (where, broadly, loser pays unlike America where each side pays), maximising billable hours is not always a viable strategy for anybody if those costs aren't recoverable on success. Someone involved in large-scale commercial litigation involving disclosure of millions of documents who doesn't use algorithmic document classification (now pretty broadly accepted as normal) potentially runs the risk of a judge determining that the costs of going through all the documents by hand isn't recoverable. Insurers/litigation funders aren't going to want to risk padding the costs so much that the judge prevents them from recovering their stake in the litigation.
Customers using their own LLMs: yep, they might do that. I think the pitch from the legal LLM providers is "we've got legally trained people doing RLHF to make it more accurate" mixed in with "also we've got a partnership with Lexis/Westlaw/etc. so we can do legal research that's better than what's on the open web", with a little bit of "if you get sued for professional negligence, 'I used the legal AI thing that's built into Westlaw' is gonna be more convincing to a judge and jury (and your insurance company) than 'I used ChatGPT, yes, like the app you've got on your phone'...".
We don't have paralegals/attorneys handle cases from beginning to end. We have different positions handle different tasks. One person may only do scheduling, another does discovery, another handles reviewing releases.
For us, adopting tech to make us more efficient is a priority. Our setup is a bit unique, but I can see PI and collection firms adopting tech similar to this.
Or maybe it will be the more established open source model where the code is free but the maintainers offer hosting/some default product
It's not a big leap to apply that model to a company and its customers, where the company builds a well-abstracted, easily extensible base that 1) Customers can easily extend/customize for their workflows 2) Customers can self-host or run fully isolated, much easier (probably not quite there yet, but is a possible world)
Sounds like your developers are relegating themselves to being review monkeys instead of developers
Engineering has moved up another layer of abstraction (just like we moved past managing buffers & writing machine code)
Potentially if used with a local LLM and not a service provider, this might protect attorney-client privilege?
But that only applies for clients using the chatbot. If a lawyer is using the LLM it is definitely protected. No different if a lawyer searches something on Google or Lexis Nexis. The search itself is protected. I guess you could debate metadata but the content surely is protected.
In Google you're generally entering fairly generic and short search queries. The example you provided ("how to give yourself an alibi after murdering someone"), is generic and could apply to anyone or could have been entered for other purposes such as writing a crime novel.
With ChatGPT and Claude, the risk is much higher, because you're basically uploading entire documents with potentially privileged material to a third party, as part of your prompt. To use your analogy, instead of entering a generic "how to give yourself an alibi" query, you'd be providing privileged interview notes and other attorney work product as part of your prompt to the LLM. In the Heppner case (which actually involved a client and not a lawyer), detailed reports and discussions of potential strategies were uploaded.
Cool project regardless!
We're going to have to re-train ourselves on what hard work looks like (and thus what should be upvoted here).
I don't know whether the project's creator (@willchen96?) is a lawyer, or if they work at a law firm that helped them shape this, or how much time and effort they put into this, or whether law firms even want or need a vibe-coded open source project for their legal AI stack, but we should be considering the totality of those things when looking at new projects these days.
There's a lot of red flags here.
I don’t actually care that much about the work having been hard - I care about the result being good.
If that's true, how does it actually achieve anything with respect to client confidentiality or anything else? (For example, there's the claim "the assistant keeps full context across every conversation and every document." --- but isn't that a function of the model one uses, which is on Anthropic or Google? Ditto the claim "Documents never leave your perimeter. Compliance, residency, and privilege stay under your control." But this is only true if you're not piping them to Anthropic or Google...) Is this just a user interface?
It would be nice if these product webpages included an easy way to find documentation so that one could figure out what the product actually does. I can't find any obvious way to discern if it can be easily used with a local model running via ollama or something, for e.g.
Since this is HN, I guess it's fair to assume it's for the US, but since English is used in more countries than the US, wouldn't it be a good idea to say outright what countries legal systems this actually understands and supports? Or is it maybe meant to be country-agnostic somehow? If so, that isn't very clear either.
That may be confusing on the naming.
Except that the font that it is using is EB Garamond and Apple was heavily using the Garamond font in the mid-1980s to 2000s.
Given that almost everyone is copying both, it is now garbage.
https://github.com/anthropics/claude-code/tree/main/plugins/...
That, plus an Anthropic-like logo.
go look at the auth - it's a call to supabase.
go look at the migrations - it's like 5 tables.
There is a real need in the space and a real opportunity for a solution like this but this is a complete nothing burger of what exists in the underlying code.
The requirements for this kind of product are extensive and complex. The shape of the data layer is complex and nuanced. Absolutely none of this is considered or implemented in the project but it sure is blowing up.
Everything the incumbents ship, in an open codebase your firm owns.
vs
Everything the incumbents ship in an open codebase, your firm owns.
laywers live in docx not pdf
I use AI extensively in my legal work. But I check every citation myself, manually. That means that I read the entirety of every case that I plan to cite in my output, and I check on Westlaw that it hasn't been overridden by a later decision. If you're just producing the AI's output verbatim, then you have only yourself to blame when things go wrong in the courtroom.