This article is approximately 369
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 459
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
A while ago in my time as a consultant I was working with a mid-size E&P outfit on a document management system. Things were going reasonably well until we tried to include the legal department in the venture. The lawyers had a healthy skepticism of our efforts. They did not trust the security. They feared that secret documents would be visible by people who should not see them. Perhaps they feared that by sharing their work with others in the organization people would see how little they actually did …
Although this was ‘a while ago’, it was not so long that things were not yet ‘digital’. OK, they may not have been ‘digitized’ or ‘digitalized’ as some fatuously taxonomize ‘progress’ in information technology, but I digress. I think that the lawyers probably used their PCs, and shared/exchanged documents via ‘sneakernet*’. Maybe they had some sophisticated legal DMS software that allowed all sorts of version control and secure sharing. I don’t know – they didn’t really talk to us. The point of this reminiscence is that, the stuff in the legal department was totally hidden from view to any DMS, search engine and, looking to the future, large language models.
Speaking of which, an observation from the EAGE Digital event we report on in this issue. When geologists put LLMs through their paces, they might ask questions along the lines of, ‘What is the porosity and permeability of the such and such in such and such an area’ or perhaps more interestingly, ‘What is the best exploration strategy for a startup working in Norway’. The capability of the LLM to answer the first question depends on how much data is visible to the LLM. A good answer to the second question might involve a knowledge of legal agreements between companies and regulators. Some of which may be publicly available, but more probably it will be hidden from view by the legal department.
I’m sure that the LLM vendors do a great job of convincing their corporate clients that the bespoke model that is being developed for them a) requires all the company’s data and documents to work and b) that non of this confidential data will ‘leak’ back into the system and be available in some form or another to future users – public or private. But given that there has been considerable debate as to how an LLM actually works, how difficult it is to ‘explain’ its findings, then it is quite possible that even if the details of some documents are declared as ‘private’, the LLM may be capable of arriving at more general conclusions from your intellectual property. This is especially worrisome if your competitors are using the same LLM vendors that also get to peek at their ‘confidential’ documents.
No I hear you say, ‘confidential’ is ‘confidential’. You have placed your trust in ChatGPT and Microsoft and are sure that they are playing ball. But information is funny stuff. It is actually quite hard to say what is or should be ‘confidential’. Traffic analysis has been used by spies to exploit publicly available information to learn about the enemy’s activity. The US is rumored to have used Soviet railway timetables to see where nuclear silos were being built.
An LLM may learn from the simple fact that you, as a successful operator, have an unusually large (or small) ‘confidential’ repository. There will always be some metadata upstream of the secrets that is potentially exploitable. And so the LLM, if it is really ‘smart’, will be learning something generic from your stuff which will make the model, and subsequent users including your competitors, smarter too.
As I editorialize, Microsoft/Office 365 is checking everything that I write and making helpful (and sometimes not so helpful) suggestions. How much of what we write is creeping back into Copilot or the Office 365 smarts? To find out, you will have to read and understand all the end user legal agreements (you know, those things where you click ‘OK’ just to install an OS or a software package). Well at least this is something that the legal department can get its teeth into. Meanwhile, as AkerBP’s Peder Aursand observed at the EAGE, ‘OpenAI and others are moving into the domain space’.
* Walking from one office to another with a floppy disk or USB stick.
Author: McNaughton, Neil
Source: Oil IT Journal, Volume 29, Number 5, October 2024
©Publisher: The Data Room SAS
DOI: https://doi.org/10.69894/457327
This article is approximately 441
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 210
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 310
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 941
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 155
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 258
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 404
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 714
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 498
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 1225
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 1023
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 846
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 1201
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 485
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 214
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 148
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.
This article is approximately 169
words long. Click here
if you would like to receive a complimentary copy. This offer is discretionary and
limited to one article per month for non-subscribers to Oil IT Journal.