Latest / REDefine // Civic Intelligence Podcast / EU in Practice: The AI Act Is Already Here. Just Not All at Once.

Transcript

EU in Practice: The AI Act Is Already Here. Just Not All at Once. · REDefine // Civic Intelligence Podcast

• 0:00
  SPEAKER_00: If you opened a customer service chat today and were told that you were speaking
• 0:03
  SPEAKER_00: to AI, that notice sits inside a part of the EU AI Act that already applies.
• 0:11
  SPEAKER_00: If an employer buys AI today to screen job applications,
• 0:14
  SPEAKER_00: however,
• 0:15
  SPEAKER_00: the full high-risk requirements for that kind of system will not apply
• 0:19
  SPEAKER_00: until December 2027.
• 0:23
  SPEAKER_00: Both situations belong to the same law, but to different points in its timetable.
• 0:28
  SPEAKER_00: Some prohibited AI practices have applied since February 2025.
• 0:33
  SPEAKER_00: Rules for general purpose AI followed later that year.
• 0:38
  SPEAKER_00: Since August 2026, transparency requirements have also become much more visible.
• 0:44
  SPEAKER_00: People should be told in certain situations when they are interacting with AI.
• 0:49
  SPEAKER_00: Deepfakes can require disclosure.
• 0:51
  SPEAKER_00: And generative systems have obligations around making synthetic content detectable.
• 0:56
  SPEAKER_00: But transparency is not a quality certificate.
• 1:00
  SPEAKER_00: Knowing that something was AI generated does not tell you whether it is accurate,
• 1:05
  SPEAKER_00: fair or appropriate.
• 1:07
  SPEAKER_00: The next major phase concerns high-risk AI used in areas such as employment,
• 1:13
  SPEAKER_00: education, essential services and migration.
• 1:16
  SPEAKER_00: Many of those detailed requirements apply from December 2027
• 1:21
  SPEAKER_00: With some product related systems following in 2028.
• 1:25
  SPEAKER_00: That does not mean people are unprotected until then.
• 1:28
  SPEAKER_00: GDPR, employment law,
• 1:31
  SPEAKER_00: equality rules, consumer protection and other existing laws continue to apply.
• 1:37
  SPEAKER_00: The AI Act adds another layer.
• 1:40
  SPEAKER_00: And companies should not treat 2027 as permission to wait.
• 1:45
  SPEAKER_00: If an organisation is developing a recruitment system now,
• 1:48
  SPEAKER_00: decisions about data,
• 1:50
  SPEAKER_00: testing, logging, human oversight and responsibility are already being made.
• 1:55
  SPEAKER_00: Much of the evidence needed later is created during development and deployment.
• 2:01
  SPEAKER_00: It cannot always be reconstructed neatly when the deadline arrives.
• 2:05
  SPEAKER_00: At the same time, the rules are not simply retroactive.
• 2:09
  SPEAKER_00: Some systems already on the market can benefit from transitional provisions,
• 2:14
  SPEAKER_00: particularly where their design remains unchanged.
• 2:17
  SPEAKER_00: So, neither shortcut works.
• 2:20
  SPEAKER_00: Nothing matters until 2027 is wrong,
• 2:23
  SPEAKER_00: but so is every existing AI system will suddenly have to be rebuilt.
• 2:28
  SPEAKER_00: And this week, another part of implementation became concrete.
• 2:33
  SPEAKER_00: Europe cited EN 18286,
• 2:36
  SPEAKER_00: the first harmonized standard under the AI Act,
• 2:39
  SPEAKER_00: covering quality management systems for providers of high-risk AI.
• 2:43
  SPEAKER_00: It does not make an entire AI system automatically compliant.
• 2:48
  SPEAKER_00: What it does is begin translating part of the law into processes
• 2:52
  SPEAKER_00: and evidence organizations can actually implement.
• 2:56
  SPEAKER_00: Other standards are still being developed around areas such as human oversight and bias.
• 3:02
  SPEAKER_00: And that is where the difficult questions start.
• 3:05
  SPEAKER_00: A human being in the loop is not much protection if they lack the information
• 3:10
  SPEAKER_00: or authority to challenge the system.
• 3:13
  SPEAKER_00: And managing bias means very little until we ask,
• 3:18
  SPEAKER_00: against whom, measured how, with which data and in which context.
• 3:24
  SPEAKER_00: So the useful question in 2026 is no longer simply, when does the AI Act start?
• 3:31
  SPEAKER_00: It already has.
• 3:32
  SPEAKER_00: The better question is,
• 3:34
  SPEAKER_00: which rules apply to the system in front of you,
• 3:37
  SPEAKER_00: which are still coming, and what should you already be preparing?
• 3:41
  SPEAKER_00: Read the full EU in practice on REDefine for the citizen explanation and
• 3:46
  SPEAKER_00: the deeper professional guide to the new standard,
• 3:49
  SPEAKER_00: human oversight, bias management, and what organisations should be documenting now.