Regulation of Artificial Intelligence Bill 2026: Committee Stage (Resumed) and Remaining Stages
Seanad Éireann — 2026-07-15
============================================================

Alice-Mary Higgins (IND), National University of Ireland

I move amendment No. 55:
In page 34, line 18, after “60(4)” to insert the following:
“and has clearly established how a risk management system referred to in Article 9 will be implemented”.
Section 48 outlines the provisions for the testing of high-risk AI systems in real-world conditions. We are out of the sandbox now and into real-world conditions. We oppose this section. We have also offered a number of significant amendments to it. As I will outline, the provisions in the section are highly concerning.
High-risk systems that might be permitted for testing in real-world conditions under this section include AI systems used in law enforcement, the administration of justice, biometrics and access to education and essential services like healthcare. Paragraph 7(a) of Annex III outlines high-risk AI systems that may be tested in relation to asylum, including polygraphs.
What the section provides for is the approval of the testing of AI lie detectors in real-world conditions. This is one of the most discredited technologies, yet we are going to have AI operate it. It is an experimental mechanism that gets everything wrong all the time. AI hallucinates and has sycophantic tendencies. These are all things about AI that make it one of the most problematic and discredited mechanisms to operate in real-world conditions on vulnerable people, including asylum seekers. It is dystopian.
What this section provides for is the testing of these systems. This could have a massive impact on freedoms and civil liberties. "Real-world conditions" is a weird way of describing reality. It is not "real-world conditions". It is the real world, with testing on real people. It is important to be very clear that this is about testing not just any AI systems, but those identified as high-risk AI systems, on real people. We are back to the "move fast and break things" piece. They are moving fast, trying stuff out and breaking things, but in these cases, the things that may be broken are people, their lives and their life chances. That is the experimental space we are in. That is the space being facilitated by this section, and it is why we will be opposing it.
As well as opposing it, we have also proposed a few amendments to try to safeguard against the mass of potential damage of this kind of reckless provision. Again, this is where the tension comes in. Are we experimenting in these real-world conditions? Will that be done for the sake of AI innovation and to promote it? A lot of what we are doing that affects society seems not to be about AI helping us to do what we want to do as a society, but how society can help AI as an incredibly overleveraged new economic experiment, with the money flowing to a very few individuals, predominantly. It is about how society can help to make that work. That is the fear I have. It is real-life experimenting with high-risk AI systems, not just AI helping people to find their way or provide maps or food services. These are identified high-risk areas.
Amendments Nos. 55 and 56 seek to ameliorate the damage by requiring that the approval of any application for the testing of high-risk AI systems in real-world conditions would be subject to a clearly established risk management system, as referred to in Article 9 of the EU AI Act, being presented to the relevant market surveillance authority making the approval decision. It is the minimum that a risk management system would be presented and be part of the process, and that is what is envisaged in Article 9 of the EU law. Clearly establishing these risk management systems should facilitate the relevant market surveillance authorities in making better, safer and more informed decisions with regard to approving any high-risk AI system for training in real-world conditions, in the real world, on real people. The risk management measures outlined under Article 9 of the EU AI Act include detailing how identified risks can be eliminated or reduced, including detailing the risks that cannot be eliminated. This information will be extremely useful for the relevant market surveillance authorities when making an informed decision on permitting the testing of high-risk AI systems in real-world conditions. That is what we are trying to ensure as part of the process.
Amendment No. 58 is an important amendment that seeks to ensure that fundamental rights considerations are an explicit factor in decisions on the granting of permissions for the testing of high-risk AI systems in the real world, with particular reference to the rights of vulnerable persons. Under the Bill, providers must apply to relevant market surveillance authorities for permission to conduct testing of their high-risk AI systems. The Bill does not provide sufficient information as to what are the aspects of a system, or the potential impacts that the market surveillance authority will be considering when making the decisions. We know that they have to apply, but we do not really know what is getting properly considered or how it is getting considered.
The real-world use of experimental systems, or systems that are being tested, can expose individuals to very serious risks and harms before those systems have been demonstrated to be safe or, crucially, non-discriminatory. We know there is a very real situation of discrimination and prejudice with AI systems. The excellent research done by Abeba Birhane during her time in UCD, and in her work at MIT, has shown this. It is a known fact that there are huge issues of prejudice and discriminatory applications. This is an opportunity to try to comply with fundamental rights and privacy protections.
The amendment seeks to ensure that any decision to permit or facilitate testing is informed by an assessment of the potential impacts, shifting the focus beyond purely technical or commercial considerations, and requiring regulators to assess the human consequences of high-risk systems. The amendment specifically highlights the fundamental rights of vulnerable persons who are likely to be disproportionately affected by AI system errors, bias, discrimination or misuse, and, therefore, require special consideration. We use the term "vulnerable persons" because there is precedent for that term in Irish law. The intention is to capture within it the wide range that we know of. I have sat in on the AI committee, and I substitute regularly for Senator Ruane. We have heard from older people, younger people, people with disabilities, people with mental health difficulties, people with limited literacy or digital literacy, and people from marginalised communities and backgrounds. A lot of different vulnerable communities are being captured here. I point to the obligation to look at special measures for vulnerable persons that is in the original EU AI Act. We want to know how that is being applied to the crucial decision to let an AI model test itself on real people in the real world. I should say that within that "vulnerable persons" category are those from ethnic minority backgrounds who have been shown to have been mistreated and impacted.
With regard to testing in real-world conditions, there is talk of testing it on law enforcement, on the administration of justice, which could be on an asylum application, which is life or death for some people, on access to education, which is the absolute core foundation of people's life chances, and on essential services like healthcare, where there is vulnerability upon vulnerability. These are serious points. Everybody is vulnerable when they are seeking to engage with education, the justice system or healthcare. Then there are people who are even more vulnerable.
On amendment No. 59, witnesses and stakeholders who have spoken to the Joint Committee on Artificial Intelligence have expressed real concern as to how their data will be processed in the development of high-risk AI systems and whether their fundamental rights will be imperilled. This is a genuine concern on the ground from representative bodies of all kinds. They want the primary legislation to give effect to the EU AI Act, which acknowledges the need for fundamental rights and privacy, and speaks about vulnerable persons. Amendment No. 29 provides that in making a determination to permit the testing of a system in the real world, the market surveillance authorities shall have due regard to the potential impact of such testing on climate, biodiversity, environmental protection and our obligations under the Paris Agreement.
There is an intrinsic link between the use of AI and energy consumption. Ireland is an absolutely absurd outlier, with a quarter of our electricity going to this one industry. It is wild. We are far away from anything like what the norm is anywhere else in relation to this matter. There is the energy issue, but there is also the water issue.
We have seen in the United States huge issues in relation to the environmental impact on water of these AI systems. When you test or train a system that is all energy usage and it is also water usage. I am sure the Minister of State is hearing from people who get frustrated when they are told to turn off their taps when they are brushing their teeth and then they see data centres down the road guzzling water. We know Uisce Éireann is not currently even documenting properly what is happening with water consumption.

---

Source: Houses of the Oireachtas. Licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). The Official Report is revised after first publication; the fetch timestamp below identifies the version quoted.

Record URI: https://data.oireachtas.ie/akn/ie/debateRecord/seanad/2026-07-15/debate/main
Retrieved: 2026-08-27T06:52:16+00:00
Sitting date: 2026-07-15