Legal firms struggle to prevent data leakage from AI tools
15 posts across 2 communities in the last 30 days; 93% from r/legaltech.
“I don't disagree. My original point was only that telling people "just get an enterprise plan lol" is an annoyingly common answer and it isn't a helpful or productive response to people who are asking about best practices for those who can't afford an enterprise plan. A…”
“Why does it not comply with confidentiality? Google has had my firm’s emails for over a decade now. Tons of confidential client information. I don’t think the length of time is what matters here.”
“Are the files with client names in pdfs/spreadsheets/docs/images? Is it the volume of data that's an issue? creating safe zones for sensitive data? Making sure any uploads to copilot / chatgpt / claude are pre-redacted? I'm more than happy to work with you to solve any name…”
Several legal professionals express frustration with the difficulty of controlling data security within AI-powered legal research platforms like Harvey and Claude. They describe building their own solutions to avoid unintended data scope, highlighting concerns about client confidentiality and compliance with professional ethics rules regarding Zero Data Retention (ZDR). The posts reveal a need for more robust data isolation features.
SOURCES (60)
“I don't disagree. My original point was only that telling people "just get an enterprise plan lol" is an annoyingly common answer and it isn't a helpful or productive response to people who are asking about best practices for those who can't afford an enterprise…”
“It's actually quite hard to do with AI. And the tooling to redact would be manual and time consuming.”
“everyone in this thread is arguing about training and ZDR but the thing that I've seen working with firms is they run into abuse monitoring. for example in a family law firm if the client consult gets recorded, client walks through what their ex did to her and the kids in detail and they paste the transcript into chatgpt to draft a summary. that transcript reads exactly like the stuff the safety filters are built to catch, so it gets flagged, and flagged content can get pulled for a human at”
“I think there are many. There are definitely many, many out there basting away on their $20/mo consumer accounts, which usually offer no more protection/privacy.”
I use it but never run privileged or confidential information through it
“What do people think about using a two step process like this: (1) redaction tool to remove personal information, plus (2) only using incognito?”
“By official instances like the Data Protection Commission/Authority/etc. In the EU you can get fined because of the fucking GDPR if you don't adhere to some ridiculous rules. But I guess if you are in the US it's much more free.”
“Can you explain that a bit more? Do you mean objectively determine whether data is used for training, OR do you mean interpret the agreements to guarantee no data to be used for training? I thought it was pretty standard to contractually prohibit training/monitoring etc.”
“I fucking hate when lawyers and people in general do that. Say it as short as possible. AI has made it so much worse.”
“Maybe. These companies are losing billions and I doubt they’ll be able to profit from subscriptions alone. Monetizing data is what we’ve seen historically from big tech companies.”
“Even if its zdr or whatever it is, there's no way one can decode if the data is used in training...”
“The question is the right one, and a few commenters already landed on the important split: Team or Business "do not train" is not the same thing as Enterprise/ZDR terms. Claude Team (and similar paid mid-tiers) usually includes a no-training setting or default, which is a real upgrade from consumer. That still does not automatically answer retention duration, whether provider staff can access flagged content for abuse review, admin-enforced controls versus per-user checkboxes, or what”
“That's the best way and the hardest: fully air-gapped, no questions asked. We developed our own private version as well for clients that require it. What do you think about data redaction? Is that a good middle ground?”
That's not how this works or the result of that case though....
“We don't use ChatGPT at all. We have an Enterprise level plan with Claude, have it installed locally rather than using the web app, set it to not use any data for training, and never input confidential client data.”
“If you think non-commercial plans are any less private than what Microsoft does to everyone’s laptop with Onedrive then you frankly just don’t understand the tech enough to be in charge of these decisions. The average cloud file storage is not truly private, and Microsoft forces every office user with a laptop into automatic syncing, which probably includes maybe…98% of the attorneys in your bar association. Agreed that free ai plans are a problem, as are any plans where the privacy feature has”
“That's true, if your decision is that only Zero Data Retention by default is the only acceptable use, then you have to use enterprise plans and configure them appropriately. I personally see that as overkill given how much data attorneys regularly entrust to digital providers without that same standard. How many times have you transfered files via dropbox, or kept client information in sharepoint, or saved a PDF to adobe cloud, or sent an email through google or outlook? Obviously none of th”
“That’s the prosumer level though, not enterprise. From what I understand, enterprise has much stronger data protection, audit trails, private data center use, etc”
“Same guarantee microsoft isn't doing it when you use sharepoint or onedrive, or dropbox when you use their service, or any other digital provider.”
“You don’t really explain how turning off training is not enough beyond “only these plans have a need to turn training off.””
“That’s a pretty reasonable position but not realistic for solos and small firms, no? Also I’m in the AI section of a bar association in another state. I might DM you if that’s ok to piggyback off your work haha”
lol this is fair... my pops puts absolutely everything into copilot
No, a paid plan is probably insufficient; see my explanation above.
“Obviously this is starting to vary by jurisdiction. I chair the AI Law section of a bar association in California. We talk about this all the time. Highlights: The general rule (ie best practice) is do not use consumer (non-commercial) AI plans because they are bound to non-commercial TOS/Privacy polices, so: ❌ $0-$200 plans (basic/pro/max) ✅ Teams/Enterprise seat licenses ✅ BYOK (Bring Your Own API Key) ✅ A third party AI legal tool which accomplishes the above While you can turn off data train”
“I would expect at least a paid plan and verification that the training retention setting is turned off. Outside of due diligence on the T&C of any model being used, the safeguards are really the same expectations we'd apply to a jr. atty who is accessing sensitive information: 1- a clear, intentional map of the data/workflow (explicit inputs/outputs specific to the project); and 2- (most importantly) an experienced/trusted attorney in the loop .”
“Exactly. OP is pretty fast and loose with "without realizing that the data may be used for training". I am not sure I buy that there are many practitioners out there just blasting away on free chatGPT plans lol.”
“Get an Enterprise plan with the contractual data protection compliance that comes with it?”
“Yes for most of it. The risk is in the files that leave the firm. Easy to move: file storage (Nextcloud), scanned paper (Paperless-ngx), passwords (Vaultwarden), backups (Restic), PDF work (Stirling-PDF). Where it goes wrong: Documents going out and coming back. Tracked changes and clause numbering can shift between Word and ONLYOFFICE or LibreOffice. Test with a real, heavily redlined agreement first. A safe middle path is to keep Word only for people who exchange drafts outside the firm. Redac”
“If I were an associate and you made me use some janky PoS to save $100/year on a license, I would be looking for another job”
“Assuming you are referring to the lawyer wearing the glasses, not, just, like, the increased availability of video evidence. The only thing I can think of is recording client interviews for the purpose of automatically generating notes/affidavits via AI.”
“Realistically I’d say you can’t and what’s going to actually prevent you is email. While it is technically possible to host your own email, you are going to get randomly and silently blocked. There are many stories of this occurring online. So practically speaking you need to pay someone to host it and if you do that then you should probably just use Microsoft or Google”
“Lmao. You can. Let us know how it works out for you. But seriously there are some European governments that are seriously looking at this after giving consideration to the potential of the American government and/or American based companies perhaps not willing to play ethically if circumstances were to favor them in that decision.”
“Why does it not comply with confidentiality? Google has had my firm’s emails for over a decade now. Tons of confidential client information. I don’t think the length of time is what matters here.”
“Claude Team doesn’t protect confidentiality. They keep all data for up to 5 years or so.”
“The instinct to pause before connecting Claude to a data warehouse of time entries and client financials is the right first step. A few things I would separate before you pick Harvey vs anonymize vs Copilot. A Team or Pro "do not train" UI setting is not the same as Enterprise or Zero Data Retention (ZDR) contract terms. Opting out of training answers one question. It does not automatically answer retention duration, logging, whether humans at the provider can access flagged content, o”
“Are the files with client names in pdfs/spreadsheets/docs/images? Is it the volume of data that's an issue? creating safe zones for sensitive data? Making sure any uploads to copilot / chatgpt / claude are pre-redacted? I'm more than happy to work with you to solve any name redacting scenarios. You can see some of my recent work by getting some fake PHI data here and testing name detection here (fully private in-browser).”
“Re: Copilot - I agree. I'm going to ping the IT group and see how much I can leverage it, as right now we're limited to outlook. Beyond that, it requires uploading Excel / Word documents. Again, if we're trusting it to look at emails, I truly can't see this as being any worse. I'm fairly new here, but do have a query out to our IT group, as we may have a team plan with ChatGPT. With respect to customer data, each client has different outside counsel guidelines, but these are”
“Well considering I am only an expert in what I do business in. I hire experts to handle things I don't understand, then to teach me what I need to understand, but I rely on them and their licensing and their profession to do what they're supposed to do.”
“I think there are really two separate questions here: confidentiality and IP. If the document is genuinely stripped of client-identifying information, and you would be comfortable with the underlying drafting style or clauses becoming more widely known, then the confidentiality issue is obviously much smaller. The part I would still look at carefully is whether “anonymized” really means that nothing identifying remains, rather than just names and addresses being replaced. With specialist contrac”
“I think the better question for firms is less “AI or no AI” and more what actually happens to the client data when a particular system is used. I’d want clear answers on retention, whether data is used for training, who can access it, what contractual protections exist, and what the firm allows lawyers and staff to submit in the first place. That’s also how we think about Referent. AI can help across the legal workflow, but the lawyer should remain in control of what information is used and what”
Ain’t there confidentiality obligations as legal counsel?
“Besides the frontier power (which I guess some might consider it as an argument to use this) I wonder how this addresses the confidentiality and trust issues that legal professionals need”
“Wow. Although, there’s the difference between a contractual agreement as to zdr and whether the provider is actually secure and complaint. Is local inference on hardware you control viable? Set up a cluster ($$$) and run an open weight model? My single RTX 5090 pretty good for personal use.”
“OP here, thanks for all of the comments, I have a follow-up question. What is the harm in having the AI "train" on my forms, if there is not confi info? I work in a highly specialized field of commercial contracts, and IMHO the contract forms and mechanisms I have developed along the way are really solid. Definitions clear. Default mechanisms solid and fair. Change of law and FM clauses exactly how I want them. We've been doing that in my industry by hand for years. I would be grat”