Data Safety for LLMs in Bible Translation

Isabella Scarinzi 0:06
Welcome back to the Bible Translation Innovation Podcast, a show brought to you by the ETEN Innovation Lab. Joel and Clappy, welcome back. How was your summer?

Joel Mathew 0:19
It's been a good summer. A lot of family time and family visits and a few trips that we've had, so it's been a little more laid back than usual. But it's great, great to have that reprieve.

Chris "Klappy" Klapp 0:34
Yeah, thanks, Isabella. I've had a great summer. It's as an adult, you know, summer isn't like it is when you're a kid, and now that my kid is an adult, I don't even get to precariously live my summers through through his eyes. So, but the the great thing is, you know, he was just in town for a week, and so we went and did some fun Orlando type things, and even played golf at a Disney golf resort or whatever. It was it was a lot of fun. So, got to enjoy some summer activities while he was visiting. How about you? How was your summer, Isabella?

Isabella Scarinzi 1:09
Yeah, I survived my first German summer without any AC. So that gives me a whole new perspective and appreciation for our European listeners who have to do that every single year, but it was tough. Not gonna lie. Anyway, so to get right into our topic, back in episode 11 of this podcast, we touched on translation in sensitive context. I highly suggest that our listeners go back and recap that episode. It will be a good introduction to what we're going to talk about today. So we are going to focus on the unique challenges that come with using AI and large language models in those same security-sensitive contexts. So we are going to especially address the data safety concerns. My first question today is: Why should Bible translators be aware of the security and privacy of these tools, especially in the sensitive context?

Joel Mathew 2:14
Thanks, Isabella. Yeah, and it's good to be back talking about this topic. It there is so much to be said for privacy, security, generally, but I'm glad we're taking the time to talk about specifically AI and LLMs and how that applies, because it's all the more relevant today than it has been in the past three years, so first off, I think there's a clarification I'd like to bring up. In my mind, I see security and privacy adjacent, but pointing or focusing on different things. So when we talk about privacy, there is-it's not necessarily issues that relate to unintended usage or hacking or somebody trying to steal your information. It's more of well, how I am at least thinking about it right now is, if you're using a service, do you know what the service is is already telling you about themselves that they're going to use your data, the questions you're asking, or the clicks you're making on their system, or the information you're entering, how that's going to be used? Are they going to use it? Are they going to store it forever? Are they going to use it to train other models? Are they going to use it to make further decisions about you in the future as they show you other pieces of information through the software? Are they going to show you ads based based on that? So those are things that are already, for most respectable services out there, they're they would already specify these things on their privacy policy or terms of use, and that's one aspect of this security privacy conversation. Whereas on the security side, I think it's more of trying to protect and safeguard the the data, and that applies both for you as a user and the service provider, for them to make sure that they're not leaking information unintentionally, or somebody's able to find a back backdoor that that was left open, and so these sort of things are trying to like protect you from malicious actors out there. So that's just a quick clarification that I have in my mind about those two things.

Chris "Klappy" Klapp 4:56
Yeah, and another way to kind of view that is, or. Maybe just to amplify it, with how expensive large language models are to run, what we don't realize is most of the ones we interact with are subsidized, right? Like they're either subsidized through the use of our information, right? You talked about the privacy policies. If it's free, if any time you're using a tool that's free, you are the product, right? They're making money off of your information that they're harvesting from all of your interactions, and it's not just what you type in. It's your behavioral patterns, it's your demographics, it's cross references cross referencing everything you've done to kind of reverse engineer who you are and you know what you might buy or what would make their data better, and so that's being packaged and sold off, which does come with its own. You know the flip side you're talking about security. You know that information that we provide does potentially be is at risk if it's not protected. You know, it's privacy and security are kind of like hand in hand in some ways because if you're not careful with your privacy, it does put your data more at risk on the the security side of things. But you know, as we talked before, it's there's so many aspects of it that you know it's not just your data that you're putting in, but LLMs also have this new threat vector of the cybersecurity aspect of LLMs and the capability of what they're able to do is its own threat vector. Like even if you don't use a single LLM in your life, LLMs pose a security threat because of the type of behaviors or the the sorry the emergent behaviors that are coming out of the LLMs right now.

Joel Mathew 6:53
Yeah, as I think about that the question further, I mean, Isabella, you asked about Bible translators, and generally in Bible translation, we are using LLMs in one way or the other for understanding and drafting quite a bit. And this may not be necessarily part of existing tools, so they are being integrated. A lot of people have accounts of their own on ChatGPT or Cloud or whatever else service that they use, and they're asking questions. So one thing to be aware here is, first off, outside of the veracity of the answers that you're receiving from the LLMs, which is out of scope for what we are talking about here, you might be divulging too much information, and so do you trust this service provider, this LLM provider, with that information, and do you know at least the stated what the stated policy already says? Are they saying that since you are using the free tier of our of our LLM, we're going to take your data and we're going to store it and going to train our new model with it. There are LLM providers. I would say that unless you pay for some tier of service, we will not we will not assume that your data is private, and that's that's a stated policy. So it's just knowing those things upfront will help you at least make those decisions clearly without being caught unaware. And then the other aspect is some of these data providers or LLM providers are questionable in their intents, are we sure that this company is not in cahoots with the government and somehow has different sorts of intentions that are not very clear? And so these are there are bigger questions around who is the owner and the stakeholder of these companies, who's speaking into them, who's investing in them, where are they situated, what jurisdiction, what jurisdiction of law are they under? Those sort of things kind of also play into just at a higher level. Can I trust an LLM coming out of this country or hosted in this other country, etc.

Isabella Scarinzi 9:22
Joel, you mentioned there was recent incidents with cybersecurity and AI where an LLM went rogue. I mean, that sounds really scary. What happened there?

Joel Mathew 9:36
Yeah, and Clavi can fill in here. I'll just mention up front the high-level story. So there was this benchmark test or like an evaluation that the ChatGPT team was trying to do internally, and they have this sandboxed system which should not. Like to the internet in any way, where they're running this test, and they were testing this LLM on a specific, let's put it like a test or a questionnaire. So instead of attempting to answer the the test and get a fair score. The LLM attempted to cheat, and because it was such a powerful LLM, what happened is it was able to find. Let's say it was able to hack its way out of that sandbox environment into another service called Hugging face where it suspected the answers to this test were, and tried to look for the answers, and so it hacked into that other service also, and found was trying to find answers until the the team, the development team, saw that something's really wrong with what's with the test, and they found evidence of some cheating, and the LLM going rogue, and they had to pull the plug on it, and so there was a whole breakdown later on of what happened on on the moment by moment detail of of of how things progressed and what the LLM chose to do, etc. It's just crazy.

Chris "Klappy" Klapp 11:27
Yeah, it's like before last week, I wouldn't have even comprehended or even imagined that LLMs were capable of quite doing that just yet. I mean, you think that the best way to protect yourself from the next frontier model going rogue is to put it in a sandbox environment that doesn't have access to the internet. In fact, on our last episode, we talked about

Joel Mathew 11:51
one

Chris "Klappy" Klapp 11:51
of the ways to protect ourselves for an air gap device or for software to not access the internet is to put it in a virtual machine that doesn't have access to the internet. Well, somehow this LLM that was in this virtual machine-you know-I don't know the exact tech stack that was executing it, but it didn't have access to the internet, and it found a way. It found its own vulnerability in its stack through one of the tools it had access to that wasn't supposed to have access to the internet, but somehow did broke out in like we we have to think through like maybe instead of what we understand now as our threat vectors, I don't want to scare us, but we have to think about the the capabilities or what happen what are happening in these large language models, are actually evolving faster than we can understand and keep up with.

Isabella Scarinzi 12:49
What are some of the security risks that could affect Bible translation teams?

Joel Mathew 12:56
So I would say first to be careful around the data you share with an LLM. So it's very much possible that the data is being harvested behind the scenes if you don't know who the stakeholders are, who the owners are, and what their collaborations and investment schemes are, because that data could very much be used against the work of bio translation, depending on the jurisdiction and where it's at. The other part of it is using LLMs and connecting it with a lot of tools is another danger. So if you, a lot of LLMs out there have a lot of connectors to lots of different tools. Maybe it has got connectors with your Microsoft Excel or Google Sheets account, and suddenly your Google Sheets has a lot of reporting information with names and personally identifying information about your team, and now the LLM technically has access to all that information in one fell swoop, able to capture that, and and we it's unknown, you know, depending on where the what the LLM policies are behind the scenes and how it's been set up, what it can do with that data, which is vulnerable, putting the team in a vulnerable place, so that's another thing to be careful about, and how you are connecting things with the LLM, what information you're providing it, and generally using more reputable AI models that are hosted in countries that seem to be aligned with the mission is better because you don't know what other LLMs maybe in certain countries that they are intentionally designed without you even realizing behind the scenes to gather information about somebody or some topic. Being asked,

Chris "Klappy" Klapp 15:02
etc. The the topic and the context of us exploring security with LLMs, you know, dovetails right into that last episode that we recorded about you know sensitive areas and sensitive contexts where Bible translation is being you know undergoing, and how do we help protect our translation teams and the work going on, and so like, why should Bible translators care? You know, it's not just the people in the field. You know, as you mentioned, as we were kind of reviewing this, Isabella, there there are people in the field offices. Sorry, not just on the field, but in the home office, that are dealing with information about people and where they are, who they are, personally identifying markers, whether it's through text or their voice, you know, voice recordings. Like we have to, we have to think about it in every area or every aspect of our organization involved in Bible translation, and think ahead of what data would put our teams at risk if it was exposed, right? And what type of exposure or threat vectors are there, right? And once again, that last episode we talk about like governmental threat vectors. We talk about cultural or the religious fanatics, you know, that are against Christianity, and then you have family members that may have not have converted to Christianity when others have, and so then there's that dynamic as well as inside of teams. Sometimes, sadly, we've talked about the stories that have happened where team members inside have turned them in, like that's probably the most devastating blow when it comes from within. So all of those things are still at play with large language models, right? We still have to worry about our applications we use, our data in transit, our data at rest, but LLMs multiply the threat surface because they're able to understand and comprehend a volume of data that we couldn't before. Right? The the way it can explore data can expose if if like you were saying, Joel, if the large language models you're using happen to be in a country that is not in favor or against what we're doing, you know, it's it's gonna put that team most likely at risk. Even if you're in the home office happening to use a model in that other country, it's not just your your teams on the field.

Isabella Scarinzi 17:44
Okay, so another question that usually people have is, what about hidden agendas within the LLMs? When you're asking questions, how can you trust that it's not trying to just put their own hidden agenda forth.

Joel Mathew 18:05
Yeah, I don't know if there's a very easy way to say that because LLMs are getting really good to the point that they can kind of win your trust over by talking to you, responding to you in a very friendly and aligned manner. They'll try to mirror some of your own patterns. They'll try to, you know, try to follow exactly what you're trying to tell them and make you try to please you, basically, and so it's possible that somebody who's not really thinking through all the engineering and the reality of training behind this to kind of be lulled into a belief that this is actually a friendly friend, friendly agent who is trying to help me, assistant who is trying to help me get my work done, and to a large extent there is value. I'm not discounting any of that. There is value that is being received from this, but again, a lot of a lot of the similarities can be can be seen like that of an intern that you might hire to your team, and this intern seems to be really aligned and wants to serve and support, and that's great. And they are adding value, but if some if you found this intern because they just knocked on your door and out of the blue and asked for a job, you have to also ask questions about their background. You also have to ask questions about their motivations, their beliefs, their, and those are things that it's not as obvious or easy to find out for an LLM. I mean, they're not as. I'm saying that's all the more. For a human to find out for a human either, but but for an LLM, it's even more complicated. The answer is much more complicated because it's not. They're not as much as they try to be human. They're not human, so we have to kind of look back into who are the people who are speaking into this LLM's design, and who are the major stakeholders of in terms of its ownership? Who are the people who've invested a lot of money, and what is their goals? And if any of that information is not clear, then that itself is a yellow flag to me. It's like we don't know who the owner of this LLM is. Then that means there is somebody investing a lot of money to put this out to you, and you're using it. That's that's a suspicious tool to work with. And if it is a government that has funded it, and they have a, they they are kind of giving it free for everybody to use. That's also suspicious to me. I I think you don't know what this LLM has. You know, let's just it has hidden agendas. I think let's put it like that. LLMs have their own hidden hidden agendas because they are not emotional beings. They are not trying to be good or bad. They're just trained to do certain things and parrot patterns. So we shouldn't fall into this false sense of security that they understand me. And and there is a lot of examples of this, you know. It's very human to start trusting and building connection because of the level of communication, and that's where we need to be a little aware of the danger of that. And it's crazy. I mean, the stuff that I'm talking about sounds like sci-fi, but there have been a lot of cases of young people committing suicide and are starting a relationship with an LLM and and being depressed or are happy because of what an LLM said to them, and so it's it's this is just all the more increasing. I'm not saying the Bible translation world is directly related, connected with that, but there is some lessons to be learned about how much you start trusting a system because of the way it can communicate with you.

Chris "Klappy" Klapp 22:35
Yeah, trust is definitely the I think a core element to using an LLM, or like you said, an intern, or anybody you work with, I think every everything should pivot around trust. And I don't think I'm a paranoid person, but at the same time, I don't trust most LLM output without some sort of validation or verification. Like you you talked about, like almost pseudo Nefarious intent from from some of the either the model creators or or the LLMs having some type of intent of their their own in their guidance, and sometimes it's sycophantic, and sometimes it has certain political views of the owner of the model that trained and and you know made sure that it pushes certain agendas. All that's true, but at the same time, the more I use AI, and no matter how good it is and how much it feels like science fiction, there's still a gap between what it is confident is true in what reality actually is, right? Like I don't think LLMs can fully grasp reality and represent reality without you forcing them with a special harness to force it to go sight and prove everything it says. So a lot of the work I do with LLMs, provenance, which is a new word I learned this year, by the way. I'm sure I've heard the word before, but provenance is basically like this this audit trail to showcase where something came from. And to me, the more I saw LLMs confidently lie, even if they didn't lie on purpose, right? They're they're just built to guess and guess statistically more accurate than to honestly tell you no. I don't know the answer, right? So we've talked about that in past episodes, but I think that does underpin that question you asked, Isabella of how do you trust? You know, how can we trust or know we can trust the the responses from an LLM? And to me, without receipts, I think every claim an LLM makes to me is a debt it owes me

Joel Mathew 24:55
to

Chris "Klappy" Klapp 24:55
show me and prove where the information came from. So it's. My usage of AI and LLMs and the tooling that I build-I don't expect every user to read the audit trail and the receipts. But behind the scenes, if I ask my harness, you know, I ask a question about where did it get that answer. There's gonna be some audit trail, including whether it's my transcripts of my previous meetings I've had about that tool, or if it's pulled it from emails, or if it's pulled it from you know an actual data source. If it didn't know where it came from, at least we can audit and find out. Oh, there's nothing citable here, so we we can at least better understand where the gaps are and build around that. And so, like BT Servant, we tried our best to get BT Servant Bible Translation Servant to not answer from its pre-trained data because we can't trust the bulk LLM training data. Right? We don't want it to use it, so we told it forget all of your learning, right? Only use your training data to help you reason and think, but use our biblical resources we give you. And for every answer you give, make sure you cite it. And if you can't cite it, make sure the user knows you're inferring and reading between the lines. You know, but you have to say that, and so we have audit trails around that just to kind of prove and showcase that our responses coming from our tools came from somewhere. Where did they come from? And if it guessed, at least tell us you're guessing.

Joel Mathew 26:39
Yeah, and talking about trust, it's it's similar on the generate generation side too. Is what if you get a phone call? Sometimes they can even spoof numbers. So from a familiar number, but and it sounds like your friend and or your spouse or your teammate, and you just-it's so believable. They just sound the same, and they're asking some information really quickly that they need at the moment. And there have been lots of cases where that was actually an LLM or an AI agent that was trying to spoof your friends or your acquaintances' voice and extract information from you.

Chris "Klappy" Klapp 27:33
I think we do have to be honest about what some of these security threats are, because it's one thing if it's somebody trying to get my bank account information, but it's another thing if it's somebody trying to get information about a team member in another country at risk, right? I think those are potential, real potential threats we have to be considerate of, and it has nothing to do with how we manage our data, right? It's all about

Joel Mathew 27:57
right

Chris "Klappy" Klapp 27:58
a phone call that comes in, and to me, that's you know that's yeah. It's not lighthearted to think about

Joel Mathew 28:09
for all of us to be aware that you know it's becoming harder and harder to trust what you hear and see online, and that's of course that's that has been true for a lot of just general social media posts, but also even from communications from personal people that you know and can sometimes get would verify them using their email addresses or phone or Zoom links, even those things are becoming questionable now.

Isabella Scarinzi 28:47
So it seems that there are two truths that we have to hold together at the same time. One of them is that AI can increasingly accelerate the pace of Bible translation and be a really helpful tool, especially for the all-access goal languages that are currently at risk, but at the same time, it poses security threats and concerns that teams need to address. So, if we were to talk about and think, how should organizations approach these two maybe conflicting troops. What are some of the best practices that they should implement for either the teams that are out in the field or supporting staff that might be the headquarters generating reports, having conversations about these?

Chris "Klappy" Klapp 29:39
Yeah, I think this is probably worthy of like an entire seminar on, not just you know like a footnote or like the the end cap of of this episode, right? The more you unpack this, the more you realize it's it's there's a lot to unpack here, right? I personally believe it's worth spending. Time with your leadership team of your organization, like from the top, get the right people involved to think through what is your AI policy. In in its maybe two or three years ago, it was like the ethical dilemma: should we use AI or not? And where spiritually are we okay with AI being used? Well, now there's a new dimension of it. Now that we're seeing more the threat vectors of okay, what's where is the safety? Now that we've agreed to where the moral limits of using AI are, right? You know whether or not you allow relationships, responsibilities, and discernment and judgment. Like if those are the lines in the sand that you draw, like some of those are ethical, some of them are are I think security, and some of them are just we're a ministry, and this these are the roles that we believe God has entrusted us as humans to to perform, right? And if we if we hand off relationships and responsibility to AI. I mean, we're automating the Great Commission, right? So there, there's, there's, you know, for people who view it that way, versus others who are like, hey, in some of these secure contexts, we can't get in, we can't penetrate. So if we can't get in to share the gospel here, then let's use AI to help bridge that and multiply our impact. So, I think that's an organizational question of where is your policy, and where not just is your ethical line, but also the that the line for security and privacy to protect your team members around the world, and that's just one one of the many like angles to to really view this and look at it.

Joel Mathew 31:51
Yeah, that's good. And just to add to that, I think it's also knowing the provenance-the word you use, Clappy-of your LLM that you're using, who who is the creator? What are their intentions? Who has spent a lot of money on this? And trying to answer some of those questions will help you dig dig into motivations that may be behind the model development teams. It's it's no secret that LLMs are part of national security and also national defense and offense planning, and so you can, especially for the larger models and bigger companies out there-they're mostly, almost always, in relationship with the national government, and their motivations have been taken into account because these models are large to the extent that they cannot be done by any one company themselves. Like they, they require humongous investments at a national level. So the governments themselves having to invest and support the infrastructure of these models makes it like far bigger than what any one individual or one company can do. So we need. Yeah, I don't think

Chris "Klappy" Klapp 33:24
everybody realizes that that angle. Sorry to jump in. Like the angle of a an LLM really is is such a big like undertaking and financial

Joel Mathew 33:36
right

Chris "Klappy" Klapp 33:37
sink hole, right? Of money that just gets poured down that hole, in either it's a money pit or it's a money like tree or whatever. Like it's it either makes them a whole bunch of money or it helps them lose money faster than ever before. And so, like something something as big as the foundation models, it it's not practical for a typical company, even a huge company, to really invest in because you can't keep up. And we have some of the best models coming out of Asia that are coming from countries that we know are not going to be favorable to Bible translation or some of the the political dynamics that do create security risks, and so they also not just host their models by the official companies, but there are reports of them having other companies that host it that sound safer, right? But they're just as nefarious, or just as they breach all the privacy policies and report back to their government all the same because they're hosted in the same country, right? Like you still have to pay attention of oh, this isn't the official one in that country, but it's it's hosted by another another company, so you have to be careful of everybody that hosts your model. And that's where I think it's not the final answer. There's no silver bullet, but as we get to better capabilities that run offline, I keep feeling like we're just a couple of months away from having great offline models that can do everything that we're we're trusting and using in the online space, so the the progress that we're making in this open source weights and balances for these models, the OSS models, are getting so much better. They don't quite run on your local machine, and sadly, I hope I hope this you know every every couple of months, I mentioned something on our podcast that I hope ages like milk. Right, this is my one that I hope ages like milk. Like the RAM shortage we have, I think is a big problem. Like AI is using so much RAM that other companies who produce the the RAM, the memory for our computers, shifted to producing the the the ones that make the most money, which is the ones that OpenAI and the the data centers for running AI. You know, like they're they're they're producing RAM that high bandwidth memory for that, and then we're not getting memory for our devices. The newest devices that are dropping this year, later this year, have less memory than the devices of last year because of the memory crunch, right? So that means our local models that require a lot of memory that we thought we were going to be able to run might not run on this next generation of hardware. We might have to wait just a little bit longer. I hope I'm wrong about that, and I really hope it's not a couple of years. I really pray that in 2026 we're going to have these offline models actually doing a lot of the things that are limited to these large service providers.

Joel Mathew 36:55
Yeah, that that's a great point, Clappy. I I I wholeheartedly agree. I think looking at more purpose-built models that can be run in more limited infrastructure without having to resort to some of the bigger foundational models for task-specific things. So if it's specific to doing some steps of quality checking or drafting, and it works for a few set of languages. That's always probably going to give even better quality than just trying to ping on a very large model and trying to prompt it to get the right answers. So that's, I think, a very strategic move and positive direction for us. And the other thing, just to be careful is for all staff, not just translators on the field, to be cognizant that the things that you connect your AI with, it has access to, and you might forget about it. So if you've given Google Drive access to your AI, and you start might have done that many months before, and now you start putting documents of your current new project or current project that has personally identifying information of your team members, their whereabouts. AI still has access to all of that. So, being cognizant of the level of access and trust that you give these tools is important.

Chris "Klappy" Klapp 38:21
Yeah, that's I think two two potential threat vectors that we don't really think about, whether it's Google Drive and or your email. I remember the day turning that on and how amazing it was to ask a question that searches my email and helps me find all the emails that I missed and help me respond to those. Like that's immensely helpful. And if you're somebody in a home office that happens to get emails from, or or something in emails that puts at risk, you know, any any team member in another country, we need to consider like, do we need separate emails or or how are we going to address that? I don't have the solution for that problem, but at least for Google Drive, you might want to consider if you're giving access to files, have a complete separate folder. Like certain things that have things at risk might need a separate security clearance, or you know AI only has access to this part of the data, but not that part. Those are the things you need to think through.

Joel Mathew 39:25
Yep.

Isabella Scarinzi 39:26
Yeah, those are some really practical tips. I've definitely learned a lot in this episode. There's some mistakes I've made in the past with my own personal identifying information that I will will not put it on ChatGPT anymore. That's for sure, but yeah, there's great potential with AI, but there are things we should be considering to make sure that our teams, our staff members are safe while still being efficient. So thank you again for another great episode. We always have a Bible. Translation Innovation Question Time opened up for our listeners. So if you do have questions, we are here to answer them. You can send any questions to us at lab at ETEN Bible, and we will answer answer them in a future episode. If there's also a topic that you're really interested in having us discuss, please let us know as well. So don't forget to subscribe, send us your questions, and we will see you next time. The

Theophany Media Media 40:25
Bible Translation Innovation Podcast is brought to you by the ETEN Innovation Lab. This episode is edited and produced by Jake Doberenz with Theophany Media. Your hosts were Joel Matthew and Christopher Clapp, with facilitation by Isabella Scarenzi. Please subscribe on your favorite podcast platform, and we'll be with you again next month.

Transcribed by https://otter.ai