WEBVTT 00:00:00.001 --> 00:00:05.900 Welcome back to Legal Prompting, I'm Nicola Fabiano and this is episode 8. 00:00:06.699 --> 00:00:13.119 In the previous episode, we saw how to integrate legal prompting into corporate compliance workflows 00:00:13.140 --> 00:00:17.340 with structured processes, governance and traceability. 00:00:18.139 --> 00:00:21.639 Today we tackle a topic that underpins everything, 00:00:22.760 --> 00:00:26.700 professional secrecy and the choice of AI infrastructure. 00:00:27.700 --> 00:00:31.659 The principle is simple but often underestimated. 00:00:32.639 --> 00:00:41.080 When a lawyer, a legal consultant or a DPO enters information covered by professional secrecy into an AI system, 00:00:41.819 --> 00:00:46.540 they are making a deontological choice, not just a technical one. 00:00:46.979 --> 00:00:52.880 The infrastructure on which the model runs, the place where the data is processed, 00:00:52.880 --> 00:01:00.240 the parties who can access it, the contractual guarantees in place, the applicable jurisdiction, 00:01:00.779 --> 00:01:04.400 all of this is part of respecting that secrecy. 00:01:05.319 --> 00:01:09.860 No well-written prompt can compensate for inadequate infrastructure. 00:01:10.500 --> 00:01:13.019 Let's look at three concrete applications. 00:01:14.120 --> 00:01:17.440 First application, analyzing a confidential file. 00:01:17.440 --> 00:01:26.639 If I need to ask a model to summarize the documents of a dispute or to compare clauses of a contract covered by an NDA, 00:01:27.199 --> 00:01:29.779 the first step is not writing the prompt. 00:01:30.360 --> 00:01:35.760 It is verifying where that data will go, whether it will be used for training, 00:01:36.300 --> 00:01:40.400 who will have access to the logs, how long it will be retained 00:01:40.400 --> 00:01:46.440 and whether the provider offers guarantees compatible with the duty of confidentiality. 00:01:46.440 --> 00:01:54.900 For highly sensitive data, a model run locally or on European infrastructure with a solid DPA 00:01:54.900 --> 00:02:00.879 and explicit no training clauses is often the only coherent option. 00:02:01.720 --> 00:02:05.720 Second application, privacy or health-related advice. 00:02:06.760 --> 00:02:10.279 Special categories of data require reinforced caution. 00:02:10.279 --> 00:02:13.559 Even when the use case appears generic, 00:02:14.080 --> 00:02:20.259 a single identifying detail can turn the prompt into a processing of sensitive personal data. 00:02:21.139 --> 00:02:28.460 The operational rule is preventative pseudonymization or, where possible, abstraction of the case. 00:02:29.399 --> 00:02:38.399 If the provider does not offer adequate guarantees on extra EU transfers, sensitive data must not leave my perimeter. 00:02:38.399 --> 00:02:45.419 Third application, drafting opinions on extraordinary transactions or criminal proceedings. 00:02:46.539 --> 00:02:49.139 Here technical confidentiality is not enough. 00:02:49.679 --> 00:02:55.820 One must consider the provider's jurisdiction, access requests from foreign authorities, 00:02:56.759 --> 00:03:05.300 exposure to regulations such as the Cloud Act and the implications of Italian Law 132-2025 00:03:05.300 --> 00:03:08.460 and the AI Act regarding infrastructure. 00:03:09.399 --> 00:03:11.699 Three cross-cutting operational rules. 00:03:13.020 --> 00:03:15.679 First rule, classify before prompting. 00:03:16.539 --> 00:03:22.500 Every piece of information entering a model must be classified by confidentiality level. 00:03:23.300 --> 00:03:27.800 Without classification there can be no informed choice of infrastructure 00:03:27.800 --> 00:03:31.440 and every subsequent decision is blind. 00:03:31.440 --> 00:03:34.899 Second rule, document the choice. 00:03:35.520 --> 00:03:41.559 The client file must be able to show which tool was used, with what contractual guarantees, 00:03:42.080 --> 00:03:47.100 on what legal basis and why that choice was proportional to the case. 00:03:48.320 --> 00:03:51.940 Documentation protects the client and protects the professional. 00:03:52.600 --> 00:03:54.820 Third rule, favor abstraction. 00:03:55.820 --> 00:03:59.940 Where possible, replace identifying data with placeholders. 00:04:00.860 --> 00:04:02.580 Work through categories and schemes. 00:04:03.360 --> 00:04:06.779 Reconstruct the context only in your own mind. 00:04:07.600 --> 00:04:09.479 The model helps you reason. 00:04:10.119 --> 00:04:13.880 It does not need to know the identity of the persons involved. 00:04:14.899 --> 00:04:19.679 Professional secrecy is not a constraint that limits the use of AI. 00:04:20.440 --> 00:04:23.279 It is the framework that makes it legitimate. 00:04:23.279 --> 00:04:30.720 Without this framework, every efficiency gain turns into a deontological and disciplinary risk 00:04:30.839 --> 00:04:35.500 and every operational advantage becomes a latent liability. 00:04:36.420 --> 00:04:40.140 In the next episode we will enter the heart of the AI Act, 00:04:40.859 --> 00:04:44.200 which specific obligations fall on the legal professional, 00:04:44.959 --> 00:04:47.140 how human oversight is articulated 00:04:47.140 --> 00:04:52.720 and what transparency in the use of AI systems concretely means. 00:04:53.279 --> 00:04:55.779 To explore these topics further 00:04:55.779 --> 00:05:02.320 and receive weekly reflections on the relationship between law, privacy and technology, 00:05:03.100 --> 00:05:07.579 I invite you to subscribe to the newsletter at nickfab.eu. 00:05:08.440 --> 00:05:09.359 Thank you for listening!