Transcript
Use with ChatGPT / Claude Auto-scroll A
Los ponentes se presentan, explican el propósito de la sesión y repasan las advertencias legales y organizativas. También introducen la idea de ver la transparencia como una oportunidad de diseño y confianza.
So, good afternoon, uh, good morning, good evening from wherever you're calling in from.Um, my name is Dr.Adrien Burn and I'll be taking you through the kind of the theory side of today's webinar and I'm joined by my colleague Joshua.Uh, Joshua, would you like to introduce yourself, please?Hello everyone.Welcome once again to the art of possible regenerative AI webinar.My name is Joshua and um I'm an AI training researcher within CAD and my role in this webinar today is to do a quick demo after Adrian has spoken on the topic for today.And um if you've been joining our previous webinars, you should be familiar with our different use cases and different tools of generative AI that we've done.And if you've not um caught up with the most recent one that we on copilot for example, you can visit our YouTube page to catch up with that.
Yeah, thank you.
Thanks very much Joshua.Um, so yeah, I should say that uh I'm a AI initiatives analyst working with Joshua and the EDH team and uh we're very happy to be with you today.Um, so I can get this moving.Okay.Yeah, there we go.Right.And I'm going to just move the tiles out of the way.So I just got to get the boring bit out of the way first, which is the disclaimer.It's very important though.So this webinar is provided for generalformational and educational purposes only and does not constitute legal or professional advice.The applicability of the EUAI act requirements will depend on the specific organization AI system role use case and circumstances.The content of this webinar is based on information legislation and guidance available as of the 4th of September 2026 and may be subject to change.It provides a high level overview and is not intended to be exhaustive.
AI use disclosure gender of AI tools were used during the preparation of this presentation to support drafting and research.The final content was substantially reviewed and edited by Cedar.Any examples, scenarios, tools, or platforms referenced are included for illustrative purposes only and should not be relied upon as a determination of legal or regulatory compliance.Responses provided during the webinar Q&A are general in nature and may not address all facts or circumstances relevant to an individual organization or use case.All trademarks, logos, brand names and the property are the property of the respective owners.References to third party companies, products or services are for identification and illustrative purposes only and do not imply endorsement or recommendation by Cedar.Thisformational webinar was prepared by Cedar Ireland Center for Applied AI and the European Digital Digital Innovation Hub for AI with funding from Enterprise Ireland, IDA Ireland and the European Commission.
Views and opinions expressed are those of the authors, myself and Joshua, only and do not necessarily reflect those of the European Union, European Commission, Enterprise Ireland, or IDA Ireland.Neither the European Union nor the granting authorities can be held responsible for them.Now, on to the housekeeping.So, we will have Q&A open.Uh you can type your questions into the Q&A uh function here on Zoom.Uh please submit them there.Joshua will be monitoring the Q&A as we proceed and we will get to questions at the end.uh we will answer the ones that are general and anything that's more bespoke we will say so and there's opportunities to read reach out to us afterwards if you feel um that you want to continue the conversation orget more queries addressed um as I said we will do this at the end of the presentation so given the nature of the EUA AI act some questions may depend on the specific AI system use case organizational context
and we were not going to get into specifics of any individual organization uh that is not the purpose of this webinar which is going to stay general And so the fourth point covers this where a question needs morenuance then we can reasonably cover live.We may follow up with you directly after the webinar where you're happy to do so.So please indicate your consent for us to reach you reach out to you afterwards.And finally please avoid sharing confidential commercially sensitive or personal information in the Q&A.And if you notice along the bottom bar of the presentation this is going to run throughout the the slide deck.So this presentation has been prepared by Cedar for general information only and should not be regarded as legal or professional advice.applicability will depend on the specific circumstances and use case.So we have to stress that time and again and we're all here today I guess because uh of the area that we're operating in which is AI and we have the EU AI act in Europe.
Um so there we go.The information I'm presenting today is current as of uh last Friday the 4th of September.Please uh process this as practical guidance not legal advice.And we shall begin with the art of the possible in terms of look looking at this uh this article 50 and the obligations as an opportunity.Um so this is an opportunity to to bake in trust into your product into your AI system into your operation and rather than just being an at the end process it can be by by a design process.So you can have trusted first contact.It it can help with procurement because the the processes involved here to document what you're doing can help with that.Then it it reinforces brand providence and then of course it can help them with scalability across Europe um as we are in the EU marketplace here.
So I would view these as opportunities.Uh so it's uh you shouldn't just look at this as a burden um but an opportunity actually to improve some processes for your benefit and your customers will thank you for it.Um so that um I've I haven't come across many customers that have put their hands up and said I'd like to be deceived please.Um so this is helping with that in terms of use of AI systems.But I have to say that these are business opportunities.They don't guarantee any legal commercial or reputational outcomes.
Se explica que la transparencia debe incorporarse al diseño y se distinguen la ley vinculante, las directrices de la Comisión, el código de práctica y los iconos opcionales. También se presentan los plazos de entrada en vigor y quién debe cumplir.
So yes transparency should now be factored into design.So this is to uh develop user trust.It's as I said put it into product design not just a label at the end.and then uh you know you can um show the evidence of this of this discipline because you're going to have to create logs.There is no certification for this article.So the best way to proceed is to develop your own in-house logs to have them there for your own records and then if any uh market surveillance authority or national competent authority come looking for evidence that you're adheringto this article then you can do so.Um so there we have it.Um, so bake in transparency into the process into the design process.And the other thing to make clear is that this article is already in in force.It became in force last month on the 2nd of August.
Um, so we're we are up and running.So what is binding and what is an implementation aid?Well, the first thing to note is that we're we're we're talking about the EUAI act and we're specifically talking about article 50 which is to do with the transparency obligations for certain AI systems.Okay?Okay.And that's binding law.So then in support of this law, we have the commission has issued guidelines that were published on the 20th of July.Uh I recommend that you review these guidelines because they are an assistant to uh the seven paragraphs within the AI act.This is a a multi-page document to to walk you through all different facets of those seven paragraphs.Then there's a code of practice that was published in June of this year.It has been deemed adequate according to European Commission and it's specifically there to help you with uh uh provisions number two for uh providers and number four for deployers um and both in reference to paragraph number five which
is to do with the accessibility and the clarity of the message.So I'd also recommend that you look at the uh to the code of practice.And then the EU has also produced a set of icons which are optional which I will introduce to you today.Uh and uh Joshua will also demonstrate them in his section later.Um they're not mandatory.You do not have to go with these but it's just uh uh optional.Uh but the EU has uh produced them and they have user tested them as well across several member states.Um, so if you don't have a better idea, you could uh do worse than start with these EU icons.But just to say that the uh items number two, three and four are considered implementation or support aids for the binding law which is number one.So start in the order of these um would be my recommendation.
uh you you cannot uh uh uh sherk the number one and don't make the mistake of thinking that if you were to adopt the code of practice or even become a signatory to it or just taking the iconfrom number four means compliance in and of itself.Okay.So you have to start with the law and follow from there.Timing.So the ERA act it seems like a long time ago now for some of us but came into force on the 1st of August 2024 and then two years later uh article 50 has come into uh force andthen there is one other thing to be noting and this is in relation to uh article 50 uh clause 2 um and has to do with the marking of synthetic content that for any gender AI system that cameonto the market before the 2nd of August 2026 there is a period of grace of uh 4 months in order to get the marking system up and running.
Um, so any system that's come into uh come onto the uh the market uh after the date of the 2nd of August uh must uh apply straight away.But if you have a system that's been on the market before the 2nd of August, you have a few more months in which to get uh article 2 up and running.Now who must comply?So uh the way in which this is done is it's separated between provider and deployer and you can be both.So provider is someone that develops an AI system and puts it into the EU market.Okay.And that means if you're in that space, then article uh sorry um article 50 sub clause one and two are going to apply to you.Five is also going to apply because that's around clarity of the message and accessibility.And then if you're a deployer, well you you didn't build the system yourself, but you're adopting the system for your own business case and real world use.
And therefore articles uh clauses three and four are going to apply to you as well as five.And the note here is also to say that employees and contractors are not to be considered separate deployers when controlled by the one organization.So you have to first of all determine whether you're a provider, a deployer or both.And then you will see which articles uh which clauses of article 50 apply um to yourself.Also to mention that there is crossber reach here.Uh so AI systems can be developed outside the EU.But the key thing is here is that if they're deployed within the EU, then this article speaks to to that system.And then there's a uh one more thing just on if you're hosting or distributing only um you're not considered a deployer if you don't have authoritarative use over that AI system.Okay, so that's just a little nuance there.
Okay, four obligations which I've briefly mentioned already.One horizontal standard.the horizontal element being that it doesn't matter if your system is high risk or not.Um uh so that's another thing to bear in mind that when the EUA act came out first everybody was focused on what was considered high risk because there was a conformity assessment for this particular article article 50 itsits outside of that chapter.So therefore whether you're high-risisk or not you still need to be aware of this uh this this article.So 50 number one is interaction.So that you have to let the user know that they're in directly interacting with an AI chatbot or AI system.Then in terms of providence, so then you have to let the user know that uh for uh synthetic content uh developed by an AI system or manipulated by an AI system, you have to let the user know that
that is the case.Then number three, exposure.You must tell a user that they been exposed to either an emotion recognition or a biometric uh classification AI system.You must tell them that.And number four, in terms of content, you must disclose to users that if you uh publish if you generated a defake or a text of public interest that it has been done so it has been generated or manipulatedby an AI system.Now, there are boundaries down here at the bottom which kind of uh speak towards exemptions.I will have a a more thorough slide on exemptions coming uh shortly.Um so that's just giving you a little flag there that there are exemptions.
Se detallan las cuatro obligaciones principales: interacción con sistemas de IA, contenido sintético, reconocimiento biométrico y divulgación de deepfakes o textos de interés público. Además, se introducen los requisitos de claridad, accesibilidad y el papel de los proveedores.
And yes, just to reinforce that paragraph 5 in article 50 is talking about being the message being clear, distinguishable and accessible.Um, so do not forget that part of it.Oh yeah, and just also to say that uh when I say that when you're uh interacting with a user, it's the first interaction is the one where this label should be revealed to to the user.Okay, for providers, here's the playbook as we see it.So you have to have an inventory of your AI systems.You have to classify which which uh which systems are triggering uh article 50 obligations.Then you should disclose uh uh which uh which ones are for hu users are interacting with AI systems.Then you should mark and detect for number two.Then you should test to make sure everything is working um as it should in terms of disclosure and detection.
and then you build the evidence base around it to show that it's working uh for the benefit of any authorities that may be asking to check to see that you're uh complying but also gives you an opportunity toimprove uh if uh over time if systems are evolving then the the marking and the the disclosing and the detection may also be hampered or degrade over time and you have an opportunity to improve there and that's justuh saying it in other ways and as I said just the last the last box there the bottom right is basically just saying that there is this period of grace uh for systems that are on the market beforethe 2nd of August you have uh four more months uh in relation to number two which is the the marking and detection for deployers playbook can be as follows so we need to define uh an authorative figure so
who is um deciding on why and how the AI system should be used then you have to classify uh the systems that are are being uh deployed then you have to label uh those systems and their the outputfrom those systems.Then you've got with the label, you have to decide where to place it.And that's important especially uh if you're using across different uh media.Um does it render well across different platforms?That's why you should review it.And then you you have ongoing monitoring.And of course, you're also building an evidence log uh for the benefit of um any authorities they may ask to see uh evidence of compliance.And it can help you improve.So accessibility by design.And then just to reiterate again as I've already said it, employees and contractors acting under organizational control are not to be considered separate deployers.
Okay.So if we go through the each of the uh the paragraphs in the article.So number one is the interaction.So essentially we have an AI system.We have two-way exchange direct interaction with a natural person.Okay.Um and this must be disclosed unless it is abundantly clear and very obvious to a reasonably well-informed observant and circumstant per person that they are interacting like they could tell uh that it's so obvious that they're interacting withAI system.If it's not extremely obvious then you should disclose and there are some uh uh outside of the scope down the bottom there and here's an example of uh of good practice.So straight up front you say you are chatting with NAS system.I can help with product questions.Okay.Uh and then just remind what good practice is first turn in context plain language disclosure.Okay.
You match the modality.So if you're in a chatbot situation you're it's where the user comes into contact.Okay.And then the the final bit is to repeat in sensitive or long interactions.Okay.That's a good that's a good tip for number two in synthetic output.So market for machines and prove it survives.Some of you may be aware that um Claude has already started this.Entropic has made a a start with Claude to to to embed it into their outputs.So there's a general pipeline here for um for how to go about doing this.So you have to layer the controls.So this is uh uh the use cases are going to come through more and more.I've mentioned already that Claude some has has been an early mover in this but we're going to see over time uh more in this space in terms of marking the uh the task force that was behind the code
of practice or they are currently working on uh marking for audio uh outputs uh and they hope to deliver later this year and it's something that we're keeping an eye on and uh we may well do another webinarin future just specifically on number two in this article in relation to the the marking and detection of uh AI outputs um because it's uh it's a fruitful area and I think there's going to be a lot ofuse cases coming forward.So um but just to be clear that when you do a machine readable marker it's not the same as a human facing label.Make that clear.And then yes there is a narrow exception which I'm going to cover again in my uh exemption slide.But yeah, if you're just doing uh editing and you're not doing substantial alterations to the inputs that go into the AI system, then uh this could be exempt from uh synthetic marking marking of synthetic content I should say.
Right.So for emotion recognition and biometric uh identification systems.So you have to inform everybody exposed and uh at at at the first opportunity should do that.And then you have to cover the difference between real time uh real time biometric identification and then uh post review identification.And then uh but this does not uh uh neglect the the the the uh the obligations you may have in running a real real-time biometric identification system uh because that is considered prohibitive under the AI act.So you'd have to double check that you're um compliant in other ways.even though you may be compliant with article 50, you may still be in a high-risisk or in a prohibited AI system um situation according to the AI act.So that final text box that comes in there, uh you're not off the hook.You have to double check with the rest of the act that you're okay to use the AI system.
Se profundiza en los casos de deepfakes y textos de interés público, junto con las exenciones aplicables en contextos artísticos, de revisión humana sustantiva y responsabilidad editorial. También se aclara qué casos siguen requiriendo divulgación.
And defakes number four for deployers.So uh I said so we say there's three ingredients here.So it's a AI generated or manipulated and that goes for image, audio or video.Uh does it resemble a real or plausible subject?And could it falsely appear as authentic or truthful?Okay, this is uh the creation of a deep fake.So what you do in those cases then you disclose the day I was used uh at the first exposure and there's some examples and though today you'll have the benefit of Josh providing examples uh once I've finished thethe theory part.Uh Josh will cover the practical part and uh show some examples of this.And again in relation to some exemptions, uh disclosure is still required in an artistic sense.Um but uh allowances made for not having to have the label so uh visible when it may uh hamper the enjoyment of the artwork created and presented.
So the second part of uh number four in the article 50 is to do you've got defects and you've also got public interest text.Um, so again, have you published something?Um, does it inform the public?Is it of public interest?If so, um, uh, if you've used AI to generate it, you have to disclose that.There's a two-part ex uh, exception.Again, going into the exemption slide, but the the heads up on this is that if you've had substantive human review, and if also a human is prepared to take editorial responsibility for the content, then you're exempt from disclosingin this case.But both conditions must be met.So there's some likely in scope examples and there's some ones that are uh likely to be out of scope.And I should have said at the start, but apologies, but we will be making this video.
It's being recorded and we will make it available on our YouTube channel.So if you find I'm going too quickly or flying through the slides too fast, you please feel free to come back and have a look at the video later uh and dwell upon the slides of interest to you.Finally, in this slide, uh just to say that the the bottom here, these would not be considered substantive reviews.So if you're doing a spell check grammar only or or an AI only review, okay, that does not meet the the the standard for substantive review in the exemptions.So let's do a decision tree.So do you have an AI system it's professional use under your authority?No.Okay.No, no application.Yes.Is it a a deep fake uh image, audio, or video that's being produced by the AI system?No.Then you have to ask the question is it being published a text for to inform the public on a matter of public interest.
So if it's yes to the to the defect one above then you have to label the first exposure and if it's yes to the the the the text for public interest then you ask the question has it beensubstantially reviewed and is there a human editorial responsibility happening?If it's yes to both of them then you do not need to label but if it's um exemption may apply otherwise you'll have to provide disclosureand then a separate question is here that even though uh uh where deployer may not need uh a label so for example uh you've substantially reviewed the text and you're taking editorial responsibility you may still have used anAI system and that may have an embedded uh machine mark okay and uh the the provider still has to go that route and still provide the embedded marking for number two to be satisfied.
So just because uh number four true number four you're exempt does not mean uh through the provider number two is also exempt.Um so as I said do not collapse there.Okay exceptions so personal non-professional use scientific research and development but pre-market.Okay, this is before it goes uh uh released to the market in in Europe.One that's here, and apologies maybe it's not the best place to put it, but I've highlighted in bold, free and open source does not exempt you from article 50 because we're focusing on the use.So if you're using open source to develop deep fakes, you still have to comply with this uh this uh act.Um but this is comes here because people might think or it might be circulating out there that if you're using open source, you're exempt.You're not.So, I've already mentioned this, but obvious interaction, if it's obvious in relation to number one, and you're interacting with a chatbot, you do not need to disclose it with a with with a label.
Uh, but only if it is extremely obvious.Again, for number two, if you're using um an AI system, but you're just using it for standard editing and it's not making any substantial change to the language and the input, then you can say you're exempt from havinga synthetic marker on that to say it was uh um su substantially modified by AI.There are some authorized law enforcement exemptions as I said on the artistic side.Um there like so disclosure is it's not a complete uh uh exemption for disclosure but where appropriate you do not need to disclose if it would hamper the enjoyment of the work.And then finally for number four um if uh you're using an AI system to produce content of public interest but you are substantially reviewing it from a human perspective and you're also assuming editorial controlled so therefore basically you
remain responsible for that content then you're exempt from disclosing um uh that it was AI generated or mod modified.
Se presenta el código de práctica como ayuda para cumplir con el artículo 50 y se explican sus ventajas y límites. Luego se muestran los iconos opcionales, el seguimiento y las posibles sanciones, incluyendo cambios recientes para pequeñas y medianas empresas.
Okay.
So, the EU code of practice, how am I doing for time here?Right, five minutes.So, I should have said as well, I was aiming to do about 30 minutes uh to hand over to Joshua then for about 15 minutes and then uh we would um have some time at the end for Q&A.So, apologies for not making that clear at the start, but I'm hoping to wrap up in another five minutes or so.But the EU code of practice.Okay.So, I mentioned already that this is to help with um uh um the transparency obligations as uh stated in the EUI act.So the commission has deemed it adequate for the purposes of articles 2 4 and five comes along because five is the one that's about clarity and accessibility.So that's for both providers and deployers.Okay.So it provides guidance if you're a provider for number two and how to uh to do the marking and then for number four and how to do the labeling if you're a deployer.
Okay.But just to make clear that like you know there are some pros and cons to to assigning this.So not only can you read you can read it of course you can you download the code of practice and just read it but if you want you can actually become a signatory um as a provider you cansign as a provider or deployer or sign both if you if you happen to be both um and the benefits of doing so is that it's EUI recognized framework um it's seen as a a positive thing in theeyes of the authorities if you've signed up to the code of practice um and it can be considered initiate like in some forms of mitigation um but it's voluntary to sign up to it it's not conclusive proof thatyou are actually compliant and if you don't sign up you do need to have alternative means in which to to show your marking and your um labeling.
Okay.So you don't need to take on board what the code of practice uh offers.Uh but that but you do need to have alternative ways to do to to do your uh to comply with the act.Um and yes that makes the point there.Okay.On the icons.So here's the set that the EU has um provided as uh optional.Um but as I said before, if you um have no better ideas, then you could do a lot worse than start here.And I will mention that um uh according to the task force that worked on the code of practice, these uh these icons were user tested across several member states.Um so they've gone through some um validation already.So you're on good ground to start with these these labels.And as I said, Joshua is going to give examples of using these labels later.
So again, they have to be placed there at first exposure.Uh no obstruction and you have to think about how this looks and renders on different platforms in different media.Okay?So uh uh just be careful about the accessibility of the of these labels if you're going to use them.They have to they have to withstand um the different forms that you would use uh with your content.Um and as I said just on the audio only uh there will be guidance coming um the task force is currently working on this in terms of marking an audio only for this is for um uh number twouh in terms of uh creating uh synthetic content audio only.Okay.So that's just mentioning the user testing and just to remind everybody that you know if you just slap this icon on that does not uh confirm that you are compliant with the uh article 50.
Okay.you have to do more than that.
Okay.So, evaluating, monitoring and reviewing.This is all going to be good for creating own your own logs.As I said, um there is no um standard that comes with article 50.Uh so, basically uh everybody has to do their own in-house um documentation.So here's some advice about like you know think about the design the testing the monitoring build the evidence log that you can then share with the authorities but it can also help you improve your designs and enhance thattrustworthy element of your um systems and your outputs.Um I'm going to talk about uh fines down the bottom there between small medium enterprises and small midcap firms.uh there is the difference in how they are treated and how they're viewed under uh this act.Um external review clock just to mention as well this this legislation will be reviewed.Um so uh they will be looking at this again in 2028.
Um so this is a moving target and what we're presenting to you today as as current as we can.Um but these things do get reviewed.Things change as technology moves on.We have seen this before.Uh so just be aware to keep in touch with the uh the provisions of the act and uh what's happening at uh European level.Um there will be reviews and things do change and I will talk about small midcap uh firms as an example of change happening very recently.So um the headline in terms of the fines for lack of compliance with article 50 is uh so 15 million or 3% of worldwide annual turnover whichever is higher.But I mentioned that the treatment is going to be different formemes and SMC's.So for them uh it's going to be proportionate and it's going to be the lower of 15 million or 3% annual turnover.
Uh for EU institutions it's going to be up to 750,000.I did say I was going to mention uh recent developments.So there was a definition which I actually have in my references about small midcap firms dated last year uh 2025 where it said it was looking at firms that were up to the size of uh 750 employees.However, this summer that's been reviewed and now it's been provisionally approved to increase the scope to include small midcap firms that are up to a thousand employees with an annual turnover of less than up to €200 million eurosand a balance sheet total of up to 172 million.So these numbers have been increased from last year's definition and just goes to show that the EU responds to feedback that it gets from uh from um from all over uh the place uh within Europe and it does
adapt.So uh you should keep in touch to make sure you're uh uh keeping a breast of these ad adaptations.There are things that will affect the the severity of the the fine and as I said it is going to be proportionate for SMMES and SMC's.So they are going to be treated at a lower level than the the full high in the box here on the left.Um they will not be treated at the higher end of the the spectrum when it comes to penalties.
Se describen los organismos de supervisión, la necesidad de documentar excepciones y un plan práctico de 30 días para inventariar, etiquetar y monitorizar. La sección termina con el resumen clave y la transición a la demostración práctica.
Okay.And then again so the in terms of the monitoring of this you've got national competent authorities in Ireland.you have the AI officer to have a limited role in in Europe and then you also have the European data protection um uh at a European level for European uh institutions.Okay.So there's a lot of different stakeholders involved here.Um but the first protocol for us in Ireland would be to go through the national competent authorities here in Ireland.So 60 test for article 50 whether it applies or not.So check your role which trigger applies to your to your system.Do you qualify for an exception?And then think about control in terms of building uh the log of evidence that shows that you're on top of this.Do not treat businessto business, open source, free or using a vendor as a blanket exemption.
Okay?Don't start from that perspective.Document your exceptions.Um so don't just take it verbally.Uh write everything down.next 30 days as a as a kind of suggested way to proceed, okay, is to look at your inventory and then decide the triggers and what applies in terms of your your systems and your outputs.Then implement your marking and your labeling.then prove that it works in deployment and then uh uh monitor the uh the evidence um as you proceed and keep uh keep uh uh the log document the log of all that you do in this regard toshow compliance.You got five executive questions which I'll just leave on the screen here for a momentand the take-home message is uh to make transparency a product uh feature rather than a footer or a last minute label.to recap.So what do we got?
We've got two actors.So that's a either a deployer or a deployer and you could be both.We've got four duties to apply to provider to apply to deployer.We got two layers in terms of like we got the technical providence uh for uh providers and we have the human disclosure for deployers.And we've got one evidence trail that applies for all as I said the documentation that you keep to evidence compliance.We're already up and running with this.Article 50 is live as we speak.It's been live for a month.So, u we are up and running.There's no time like the present to get going since the act is this is already in force.Okay.Remember, you got to be visible and you got to do it at first uh disclosure uh with the user.Don't be hiding it uh anywhere.Don't be putting baking it into your terms and conditions buried in some small print.
It's got to be visible.Um so, so uh stress test that in your organization.And yeah, just to remind SMMES and SMC's that yes, you will have proportionality in terms of the uh potential penalties that could come in.Um uh so you will have reduced penalties, but that doesn't mean that you're exempt from complying with the the legislation.Just don't make that confusion, okay?You still got to comply.Does not remove underlying obligation.Okay?So I'll just leave this on the screen.We've reached the appendix.uh just a couple of minutes over.Um but I'm not going to really speak to these.Uh they could be useful.So when you come back to the YouTube video of this uh you can uh maybe freeze frame the screen and check um uh this could be useful if you're from a provider perspective.Then I'll move over to the deployer responsibility matrix.
This is just a recap um of what I've covered.Hopefully I tried to get across to you in the last 30 minutes.And then finally, the sources.Again, you won't be able to click into these.Uh, in theory, if you had the slides, you could, but um, uh, but anyway, these are the the the references I used to create the the slide deck.Um, oops, I just go back a second here.Just to give you an example, this definition down here, bottom right, is the 2025 definition of SMC.I um expect this to be updated um uh in due course because it's provisionally been uh approved to increase the scope of small midcaps from 750 employees to a thousand.And now I'm ready for I see a hand is raised here from a from a member of the audience but um I believe we're going to save uh Q&A.
If you would type your um comments or questions into the Q&A uh we will uh get to them at the end.I think now I'm going to hand over to Joshua.Uh I'm going to stop sharing my screen so we can move to the demonstration.Um but please do write down your thoughts if you have them now.Write them down now and we will get to them when Joshua concludes his section.Right.I stop.
Joshua muestra cómo etiquetar imágenes generadas o modificadas por IA usando Canva y Gemini. Explica dónde colocar las etiquetas, cómo evitar que se recorten y cómo conservar evidencia del proceso.
Can I mute them?
Uh you're muted Joshua.
Yes.I'll just go ahead and share my screen.
Okay.Um, thank you for staying with us.I would just um go ahead and do a quick demo on how to label AI generated content in your own use cases or workflows.
So um this is a repeatable workflow that I have put together for all four modalities across text, images, video and audio.So when you've created a content the first thing you need to ascertain is has AI been used and if AI has been used then is it um an image content or text audio or video then you need togo ahead and decide if it's a deep deep fake content or if it's a public facing content or a content of public interest and then you need to go forward to disclose if it's um a content whereby Iit contains basic use of AI or if um it's been substantially generated with AI or if it's been modified with AI and then the next thing to do is to sufficiently um put the labeling in the position thatcan be easily seen at the first um look at such content or at the first listening to such um content.
So you for images and video and test as well you need to put the you need to put the label in where it cannot be easily cropped out which is one of the um stipulations that has been statedby the European Commission.And next is you can go ahead and publish and keep evidence.So note that this demo session is not to show how to generate um AI content but just to show you how to do the labeling.
So um we have an interesting um use case here when it comes to labeling and some of us are probably familiar with this picture showing on the screen.So um the picture on the left is an actual um image of a house here in Ireland that was um advertised online and then AI was used in modifying a replica of it in the image on the rightand you know in such a case like this you would look at if you look at the screen you would see the picture on the left that the sky wasn't blue over there but on the right the skybecame blue and on the left as well you can see like um electric wire like passing through the roof of the house and you can see difference in the staircase as well.You can see difference in the in the flower or the grasses as well.
So in this case, AI has been used to modify the original look of the house for advertisement purposes and it is important that it is clearly stated to members of the public that you know AI has been usedin modifying this content in order not to mislead um potential persons that might want to rent such house.So I'll just quickly show us how you could label this for example.Um, so what I would do for example is to use Canva and when I go to Canva, I would just click on the upload button over here and I would just drag whatever image I want to I'll justdrag and drop whatever image I want to label into the upload interface here.and I would click on use in a new design or you could just edit image for example.So the image is now uploaded into Canva and what I would do is to go and get the appropriate icon that be fits this kind of image has been modified with AI and that icon is the AI modified
icon that Adrian showed to us earlier on.So I'll just um drop the icon.So okay just a moment.So click on createcreate and you can choose whatever type of um creation you want to do.So I could do upload, drag the image into Canva
and edit image.
So um the image would usually take time to upload.
So now this is the image.So I would copy the relevant label which is the AI modified label which is it over here.So like I said earlier on you need to place the label somewhere where it can be easily cropped out of the AI modified image.So you can place it somewhere here and you don't have to adhere to this um shape that has been provided by the European Commission.You can you might want to use um square boxes or rectangle rectangular boxes instead.And you might want to change the color of the writing on it.You might want to change it to for example um a blue background and a white um text inside of it.But the essence is having the right um text inside the image and placing the label where it can't be easily cropped out.So this is just a simple example of how you can label for image generation.
And we'll look at another use case when it comes to image generation as well.And in this um context here, so if you look at the slide by the right side of it or the middle part of it, you would see a video playing.So I used um Gemini for example to generate this image on the left.Though this image on the left is a synthetic image as well, but I'm using it for demo purpose to show how you can um use Gemini for example to modify an existing image or a real life image andalso using Gemini to also embed the label rather than going to Canva.So what I did here was to upload the image I want to modify and I also uploaded the appropriate label over here as well and I just imputed this prompt.I said create a version of this image that includes um pedestrianization trees cycle lanes outdoor sitting also label it with the attached AI modified label attached in line with article 50 of the hue hair act.
Now when I did this, Gemini generated an image and automatically imputed the AI modified label into it.Now I found something interesting while I was doing this for those of us that are familiar with Gemini.So Gemini usually has its own custom watermark that it puts somewhere around the bottom right corner of an image or a video.But based on the instruction that I gave to Gemini, what I noticed was that it took off its own watermark and replaced it with the air modified icon that I um provided which is in line with the HUAI hat article 15.So you're going to see the output coming up shortly and you'll see how it transformed this um this supposed um original image.So you can see the screen you can see that has transformed um the image and it has also imputed the label over here as well and the Gemini um Gemini's own watermark isn't showing because of the instruction that
I gave to it and I didn't tell you where to place the image as well to note that out but it placed that image where it would usually place its own watermark and then replace the watermark with themodified image.So this is another way you can do this outside the canva option that I showed earlier on.
Se enseña cómo etiquetar vídeos deepfake o traducidos con sincronización labial, incluyendo el uso de iconos, texto y voz en off. También se explica cómo adaptar el etiquetado a distintos formatos y mantener la claridad para el público.
Now this is another interesting use case when it comes to um deep fake and video generated um content.So yeah we have a use case here.This is a fictional character appearing on the screen, but I'm using this use case as um a supposed real life CEO of a business.Now, the CEO wants to make an announcement of um applications for the new business support scheme which is opening the next Monday and this this initiative is to provide funding and mentorship and it's the video is saying thatum the audience should visit the websites of the organization to learn more.Now next we are going to see the supposed original version of this video that was gen that was um created in English
applications for the new business support scheme open
AI generated
applications for the new business support scheme open Monday.This initiative will provide funding and mentorship.So visit our website to learn more.
Okay.So over here this is the supposed original video of the CEO speaking on this announcement.Now AI has been used in this next slide to generate an Irish version.
Yeah.So um I would like to say that within our team here at CEDA um there's no one very fluent in the Irish language.So you might notice some discrepancy in the ascent or in the way AI has generated um regenerated this person speaking Irish and also in the Irish um language written on the screen as well.So um you might find some inaccuracies in this but this is just for illustrative purpose only.So let's listen to this video one more time and I'll point out some things to us.AI generated voice and lip synchronization scheme.
Now if you listen to that you notice from the beginning there was um a voice over that said AI generated voice and lip synchronization and you can also see the watermark on the screen as well but why isthere a need for a watermark if you know the announcement is supposedly accurate why still there a need to disclose that this has been AI generated and this is simply because um any any audience or any viewer canreally just think that this is the CEO actually very fluent in Irish speaking in Irish whereas that's been um it has been edited and AI has been used to make the CEO speak in Irish and some of usare probably um conversant with some videos on Instagram at times whereby um for example a football press conference has been held in Spanish but then you see that person speaking in English their mouth is sinking and speaking in
English as well.So that has simply been translated with AI.So I'll quickly show us how you can also do this labeling as well within Canva.
So, what I'll do again is to go to Canva and I would use the create button over here and I'll click on upload and I would drag and drop um the Irish um version of the announcement video.
So now this is our upload and in this video you can see that Gemini's um you can see Gemini's watermark appearing at the bottom right over here as compared to the image we generated earlier on whereby Gemini usedthe watermark I provided to replace its own logo.So it's not sufficient according to the European Commission to use this logo alone.You need to put the you know the recommended um logos.You need to put it right in because someone might see this and not understand or might not be conversant that this is a Gemini logo.So you still need to put um um a very obvious logo and test into it.So what I would do now is to add um a voice over to say this is AI generated.So I would click on add audio at the bottom left here and I'll click on AI voice and I'll copy and paste the text I want inside the voice over.
So I would paste a test like this saying AI generated voice and lip um synchronization and you can choose whatever um voice you want to use.If you want to use a female voice or a m voice I'll just use this voice over here
AI generated voice.So now it's generating and you can see that the generation is now at the bottom here and it takes um it takes about 3 seconds.So what I would do is to leave it at the beginning here.And I also one thing I also would do is to add um a test at the beginning of the video that also says that this is AI generated for accessibility needs of those that might not be able toget but just um can be alone.So, I would click on blank over here.And you can change this color of this blank um background here.You can change the color if you want to or just change it to like maybe like a gray color.Yeah.And I'll move this to the beginning.I'll move this to the beginning.So this same test that we used here, you can copy the same test and paste it into this blank page here.
So you can see the test on the blank page.You can change the color if you want to and you get this AI this basic AI icon over here.So in the labels that um Adrian showed earlier on.So you wouldn't see a label like this saying generated voice and lip synchronization.So this is where you use the basic AI icon that's been provided to um replace the AI in this test.So now it's going to be AI generated voice and lip synchronization.So in this case you use this basic icon where you need to modify the labeling to make it very much more obvious for your particular use case.So rather than using the AI modified or the AI generated icon alone, I'm using this basic icon for the AI and then I'm writing the test AI generated voice and lips synchronization.
So I'll drag that to the end of to the very beginning of this um content we're generating in here and I'll make everything to sync with this 3 second um AI generated voice over that we have here.So I'm going to sync this blank page that we edited to it.I'm going to sync this test to be in line with it as well.The 3 second and I'm going to sync this AI icon as well to be in line with it over here as well.Now within this video itself, I'm going to also put the icon
I'm going to put the AI icon that we used earlier on.So, you can see the icon.I'll place it somewhere here.I'll place it somewhere here.And I'll reduce it a little bit.And I'll also place the test that we had at the beginning page over here as well.
And you can reduce the font size if you want to.So you can see um
so you can play around with how it looks in your own place.So you can see that I've placed the AI icon and I'll remove this because the icon is replacing the AI in the test.So now it's AI generated voice and lip synchronization.So you you can keep the Gemini icon if you want to still or you can crop this Gemini icon out if you want to.And now we have the labeling for this video ready.And when we play
AI generated voice and lip synchronization
the fog line.
Now the is done.So you can see some white blank pages.So you just need to keep the content within the length of the actual video and you're done.And you can click on share.You can download it as a video or whatever.You can share the link with um anyone in your team that needs to review this and so on.So that's it with um the video labeling.
Se aborda el caso de textos generados por IA y cómo señalarlos con iconos y notas editoriales cuando no han sido revisados sustantivamente por humanos. Finalmente, se resume un flujo de comprobación antes de publicar contenido.
So um next we have the audio um audio use case.And now um for this use case we are not doing a demo of how to do this um in real time because um the way this can be done has not been explicitly stated yet by the European Commission.So um there's a there's there's a European task force at the moment looking at how marking of audio files can be technically accomplished and once this um guidance is published sed plans to do further webinar on this topicbut you know what you can note is um if you're generating a fully cloned um voice of the executive in an organization the voiceover should start with a statement like this program contains an AI generated through production ofthe managing director's voice for example and also if this audio is being published on a website or on YouTube or whatever channel for example there's also a need to have a test like this the same icon basic AI
icon that I used earlier on so having it and saying AI generated voice reproduction so like I said I'm not doing a demo on this and we'll do a subsequent webinar on this particular uh topic.
So um next we have um the test use case how to label um test generated content.So um Eduan spoke earlier on on when you should label and not label.So as long as a test has been generated solely by an AI and it hasn't been edited or modified in any form, reviewed by a human or owned by a human, then you need to label such um test.So if you look at the screen, this is a um this is a fictional county in Ireland and um the test was 100% generated with um cloud AI and I have labeled it at the top here inside theheader with the icon AI generated and I've also put a footer note saying AI generated information not substantively reviewed by a human editor.And if you look at the screen, you can see the prompt I used in generating this.
Like I said, the purpose of this webinar is not to show teach how to generate, but just to show how to label in your own use case.So I'll just quickly show how I did this labeling as well.So you can see the you can see the tests and what I did was to go into the double click on the top here inside the header and I copied um the AI generated icon.So if you go to um if you go to the Europa um link um the Europa link there you would see these icons and you will be able to download those icons and I believe we'll be sharing we'llbe sharing the links afterwards.So what I would just do is to copy the AI generator icon
and paste it inside of it.So, and what you do is to reduce the size.Just play around with it until you have the size or the shape you would like to see.So, yeah.So, I'll do that and I'll go into the footer here and I would paste this test inside of it.
So the test is simply saying AI generated information not substantively be viewed by a human auditor a human editor beg pattern and that's it with the labeling of the test generated um content.
So um lastly here's a quick publishing test that you can do for yourself when you've generated the content.The first thing is has AI been involved?If yes, is it a deep fake um content?If it's a deep fake content, then is it of a public interest?And also, if you've generated a test, is it a public interest um test?If it is, then has it been reviewed?If the answer is no, but the review of the public interest test and if the content you've also generated, be it image or your video is deep fake as well, then you need to label it using a clearicon.the AI modified or the AI generated or the basic AI icon and the modified test as the case may be.So um the compliance insight in the is shown before the drag and drop.Um that's just a quick um showcase of how you can do this um labeling but this would evolve over time as the European Commission releases more guideline more updates on this topic.
So thank you very much and I'll hand over to Adrian now.
Se comparten códigos QR para el código de práctica, la oficina nacional de IA en Irlanda y un evento próximo. La sesión termina con la indicación de que las preguntas se tratarán fuera de grabación y con una invitación a contactar offline.
Right guys, we're just at time here, but um thank you so much for staying right to the end and I'm going to reward you with a couple of QR codes.Um uh just pop it up the screen.Josh, can you confirm you can see disclaimer screen?
Yes.
Yes.Okay.Uh so yeah, there's the disclaimer again.I'm not going to read it out this time, but just to remind everybody that uh we we had to tread a careful path today in terms of um what we can say and um and so forth.Uh but anyway, um see this works.There we go.Yeah.So, um first QR code for you.Um if you are thinking about signing the code of practice and you want to learn more about it, um please feel free to uh scan the QR code on the screen now um and take that forward.I'll just leave that on for a couple more seconds if anybody is uh got their phone out.And anyway, if you can get the YouTube video later to to do this again if you miss out.Um another thing to introduce to you is that we now have a national AI office in Ireland.
Um if you want to learn more about that, uh please scan the QR code on your screen now.Um so just just as with the uh article 50 coming into force on the 2nd of August uh the national office in Ireland came into force uh came into being on the 2nd of uh August as well thisyear.So what great timing.So if you want to learn more about uh uh Ireland's own uh AI office uh there's the QR code there and please scan and um and have a look.And then from from singer's perspective, we're going to have a half-day event in Crow Park on October 20th.Um, if anybody here on joining the webinar today is interested to to participate in our our half-day event uh uh next month in Crow Park in Dublin, uh please scan the QR code on your screen now um to
learn more details about how you can join the event and uh participate and we'd love to see you there.
And then finally um and I am going to get to the Q&A.There is a question for me which I'm going to answer.Um but I'm going to stop the recording before doing so as as per our protocol which I actually didn't mention to you earlier.But uh we will turn off the recording before addressing any Q&A uh today.Um, but if you feel like you uh didn't get a chance to discuss anything with us or you want to take things forward with us in any way possible, that's absolutely great.Uh, just scan the QR code on your screen now if you want to reach out and if you've got any queries that you want to take offline as it were.