WEBVTT 00:00:09.659 --> 00:00:13.939 Welcome back to Legal Prompting, this is Nicola Fabiano. 00:00:14.500 --> 00:00:21.299 In the previous episode, we discussed professional secrecy and how choosing an AI infrastructure 00:00:21.360 --> 00:00:25.559 is, first and foremost, a deontological decision. 00:00:26.559 --> 00:00:33.159 Today, we shift our gaze from the single tool to the regulatory framework that governs 00:00:33.159 --> 00:00:39.779 its use, the European AI Act and the obligations it places on legal professionals. 00:00:40.020 --> 00:00:44.419 The central principle is easy to state and hard to practice. 00:00:45.020 --> 00:00:52.619 The AI Act does not only regulate those who develop or place AI systems on the market, 00:00:53.299 --> 00:00:59.599 it also regulates those who use them, that is, the deployer, and the legal professional 00:00:59.599 --> 00:01:05.779 who integrates AI into their practice is, in every respect, a deployer. 00:01:06.779 --> 00:01:10.080 This means that choosing a good tool is not enough. 00:01:10.639 --> 00:01:14.599 You must use it within a precise perimeter of obligations. 00:01:15.120 --> 00:01:21.199 To this perimeter, we must add a cross-cutting one already fully applicable since February 00:01:21.199 --> 00:01:29.099 of 2025, the duty of AI literacy for staff set out in Article 4 of the regulation. 00:01:30.180 --> 00:01:35.680 Anyone using AI in a law firm must have acquired adequate competence. 00:01:36.839 --> 00:01:42.540 Anyone who entrusts those tools to collaborators must ensure they have it too. 00:01:42.959 --> 00:01:44.980 Three concrete applications. 00:01:45.620 --> 00:01:49.339 First, classify the risk of the system you use. 00:01:50.099 --> 00:01:58.419 The AI Act distinguishes between prohibited systems, high-risk systems, limited-risk systems, 00:01:58.720 --> 00:02:00.559 and minimal-risk systems. 00:02:01.480 --> 00:02:08.960 Many general-purpose tools used in a law firm fall within the limited-risk category with 00:02:08.960 --> 00:02:16.580 obligations that are mainly about transparency, but when a system is used to assess individual 00:02:16.580 --> 00:02:23.979 profiles, to support decisions that affect fundamental rights, or for activities covered 00:02:23.979 --> 00:02:28.839 by Annex 3 of the regulation, the classification changes. 00:02:29.399 --> 00:02:36.800 The first question is not whether the tool works well, but which risk category your concrete 00:02:36.800 --> 00:02:38.580 use of it falls into. 00:02:39.000 --> 00:02:42.880 Second, notice to the client and to the recipients. 00:02:43.600 --> 00:02:51.000 Article 50 of the AI Act imposes transparency obligations toward those who interact with 00:02:51.000 --> 00:02:54.100 AI systems or receive their output. 00:02:54.899 --> 00:03:00.179 For the legal professional, this translates into clear clauses in the engagement letter, 00:03:00.820 --> 00:03:07.679 in opinions that flag where AI has contributed, and in communications that do not pass off 00:03:07.679 --> 00:03:12.240 as purely human and content generated or reworked with AI. 00:03:12.880 --> 00:03:18.520 It is not a bureaucratic formality, it is how you preserve the trust that grounds the 00:03:19.339 --> 00:03:20.240 professional relationship. 00:03:20.619 --> 00:03:23.179 Third, documenting your use. 00:03:24.179 --> 00:03:31.520 The AI Act requires the deployer to keep the logs of high-risk systems and to monitor how 00:03:31.520 --> 00:03:32.460 they operate. 00:03:33.259 --> 00:03:39.839 Even outside the high-risk category, professional prudence requires you to document which tools 00:03:39.839 --> 00:03:46.779 are used, for which tasks, with which standard prompts, and with which human controls. 00:03:47.360 --> 00:03:53.899 An internal register, even a simple one, is the trail that allows you to answer tomorrow, 00:03:54.539 --> 00:04:00.960 before a client, an authority or a judge, on how a given content was produced. 00:04:00.960 --> 00:04:04.300 Three cross-cutting operational rules. 00:04:05.320 --> 00:04:08.820 First, effective human oversight, not symbolic. 00:04:09.600 --> 00:04:14.740 The AI Act speaks of human oversight, and the phrase is not a slogan. 00:04:15.440 --> 00:04:21.799 It means that the professional must be in a position to understand, verify and, if necessary, 00:04:22.160 --> 00:04:24.100 contradict the output of the system. 00:04:24.640 --> 00:04:29.720 If the output is not verifiable, oversight is only apparent. 00:04:29.720 --> 00:04:33.140 Second, preliminary assessment of tools. 00:04:34.179 --> 00:04:39.440 Before integrating a system into your workflow, you must check the technical documentation 00:04:39.440 --> 00:04:47.420 provided by the supplier, the data processing terms, the conformity guarantees and the consistency 00:04:47.420 --> 00:04:50.279 with your duty of confidentiality. 00:04:50.380 --> 00:04:55.779 A technological choice made in haste becomes a regulatory problem tomorrow. 00:04:55.779 --> 00:04:58.359 Third, continuous updating. 00:04:58.839 --> 00:05:05.880 The AI Act is a regulation that enters into application gradually, with different dates 00:05:05.880 --> 00:05:08.160 for different categories of systems. 00:05:08.880 --> 00:05:16.619 Some prohibitions apply from February of 2025, the obligations on general-purpose AI models 00:05:16.619 --> 00:05:23.459 from the following August, and the rules on high-risk systems come fully into force from 00:05:23.459 --> 00:05:25.380 August of 2026. 00:05:25.779 --> 00:05:33.279 The legal professional must follow its implementation, because their responsibility grows as the 00:05:33.279 --> 00:05:35.040 obligations become binding. 00:05:35.500 --> 00:05:41.239 In the next episode, we will close the journey, a synthesis of the legal prompting method, 00:05:41.739 --> 00:05:47.140 and a look at the perspectives of evolution, both regulatory and technological. 00:05:47.299 --> 00:05:54.540 To explore these themes further, I invite you to subscribe to the newsletter at nickfab.eu, 00:05:54.540 --> 00:06:02.260 where each episode is accompanied by a dedicated issue with references, notes, and additional 00:06:02.260 --> 00:06:02.920 materials. 00:06:03.779 --> 00:06:04.459 See you soon!