Brussels, 06 October 2022
WK 13423/2022 INIT
LIMITE
TELECOM
WORKING PAPER
This is a paper intended for a specific community of recipients. Handling and
further distribution are under the sole responsibility of community members.
CONTRIBUTION
From:
General Secretariat of the Council
To:
Working Party on Telecommunications and Information Society
Subject:
Artificial Intelligence Act - DE comments on 2nd part of 3rd compromise proposal
(document 12549/22; Title IA, Arts 30-85, Annexes V-IX)
Delegations will find in the Annex the DE comments on 2nd part of 3rd compromise proposal on
Artificial Intelligence Act (document 12549/22; Title IA, Arts 30-85, Annexes V-IX).
WK 13423/2022 INIT
LIMITE
EN

---

DE comments on second part of third compromise proposal on AIA (document 
12549/22; Title IA, Arts 30-85, Annexes V-IX) 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
 
 
 
Please note that the following views are 
preliminary as we are still examining the 
proposal. We reserve the right to make 
further comments. 
 
We also refer to the comments that we 
handed in on 14th September 2022 and 
within the CWP on 29 September 2022. 
 
Germany is of the opinion that the 
specific characteristics of the public 
administration (and in particular those of 
the security, migration and asylum 
authorities, as well as the tax and customs 
authorities, including the FIU) can be 
better accommodated in a separate, 
specific technology act or in a separate 
section in the Regulation (referred to in 
this document as “separate regulation”). 
The provisions in the separate regulation 
should be exhaustive. For details, we 
refer to our Paper on the separate 
regulation. 
 
We would like to state as well that double 
regulation should be avoided. 
Recital 61 
(61)  Standardisation should play a key  (61)  Standardisation should play a key role to 
Addition 1: Art. 41 does not allow for 
role to provide technical solutions to 
provide technical solutions to providers to ensure 
CS anymore if existing hENs are 
providers to ensure compliance  with 
compliance  with  this  Regulation,  in  line  with 
“insufficient”. 
 
this  Regulation,  in  line  with  the 
the  state  of  the  art.  Compliance  with 
Addition 2: copied from the Nonpaper 
state  of  the  art.  Compliance  with 
harmonised  standards  as  defined  in  Regulation 
CS from the Commission to clarify the 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
 
harmonised  standards  as  defined  in 
(EU)  No  1025/2012  of  the  European Parliament  conditions for CS as described in Art. 41. 
 
Regulation  (EU)  No  1025/2012  of  the  and of the Council25 should be a means for 
Alternatively, Recital 40 from Machinery 
European Parliament and of the Council25  providers to demonstrate conformity with the 
regulation (as also stated in the Appendix 
should be a means for providers to 
requirements of this Regulation. However, the 
of the Nonpaper) could be copied here:  
demonstrate conformity with the 
Commission could adopt common technical 
(40) The current EU standardisation framework 
requirements of this Regulation. 
specifications  in  areas  where  no  harmonised 
which is based on the New Approach principles 
However, the Commission could adopt 
standards  exist and a standardization request 
and on Regulation (EU) No.1025/2012 represents 
common technical  specifications  in 
following Art. 10 of Regulation (EU) No 1025/2012  the framework by default to elaborate standards 
areas  where  no  harmonised  standards 
has not been successful for the reasons described in  that provide presumption of conformity with the 
relevant essential health and safety requirements 
exist  or  where  they  are insufficient.  An  Art.41 of this regulation  or  where  they  are 
of this Regulation. European standards should be 
appropriate  involvement  of  small 
insufficient. Concerning these reasons, it  is  to  be 
market-driven, take into account the public 
and  medium  enterprises  in  the 
noted  that  Regulation  (EU)  1025/2012  provides 
interest, as well as the policy objectives clearly 
elaboration of standards supporting 
that  the  European  standardisation  organisation 
stated in the Commission’s request to one or 
the implementation of this Regulation  should adopt a standard within the deadline set out  more European standardisation organisations to 
is essential to promote innovation and  in the request of the Commission. However, 
draft harmonised standards, within a set deadline 
competitiveness in the field of 
possible delay may occur in the delivery of such 
and be based on consensus. However, in the 
absence of relevant references to harmonised 
artificial intelligence within the  Union.  harmonised standards due to the technical 
standards, the Commission should be able to 
Such  involvement  should  be 
complexity of that standard. Such possibility 
establish, via implementing acts, common 
appropriately  ensured  in  accordance  should be considered thoroughly before having 
specifications for the essential health and safety 
with Article 5 and 6 of Regulation 
recourse to the adoption of common specification,  requirements of this Regulation as an exceptional 
1025/2012. 
which should remain an exceptional measure. 
fall back solution to facilitate the manufacturer’s 
An  appropriate  involvement  of  small  and 
obligation to comply with the health and safety 
medium  enterprises  in  the 
requirements, when the standardisation process 
elaboration of standards supporting the 
is blocked or when there are delays in the 
implementation of this Regulation is essential to  establishment of an appropriate harmonised 
promote innovation and competitiveness in the 
standard. If such delay is due to the technical 
field of artificial intelligence within the  Union. 
complexity of the standard in question, this 
should be considered by the Commission before 
Such  involvement  should  be  appropriately 
contemplating the establishment of common 
ensured  in  accordance  with Article 5 and 6 of 
specifications. 
Regulation 1025/2012. 
  

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
Art. 3 
We recommend referring to the 
(28) 
‘common specifications’  means  (28) 
‘common specifications’  means a set of 
definition in Regulation (EU) 1025/2012.  
a set of technical specifications 
technical specifications document as defined in 
 
document, other than a 
point 4 of Article 2 of Regulation (EU) No 
Art. 41/4 of this Regulation states in line 
standard,  containing solutions, providing  1025/2012, other than a 
with that producers may use technical 
a mandatory means to, comply with 
standard,  containing solutions, providing a 
certain requirements and obligations 
mandatory means to, comply with certain the 
solutions that are equivalent to those 
established under this Regulation; 
essential requirements and obligations established 
outlined in the CS to prove conformity. 
under this Regulation; 
Thus, CS are voluntary and not 
mandatory. This choice reflects the fact 
that harmonised standards are also 
voluntary, unless the Regulation in 
question defines them to be mandatory 
– this is not the case for AI.  
 
As harmonized standards and following 
the rules of NLF, CS can only cover 
essential requirements. Not all 
requirements of the Regulation in 
question or those that the regulator 
wishes to be covered.     
Article 4b(1) 
 
 
Article 4b refers to implementing acts 
regarding specifying and adopting 
requirements established in title III, 
Chapter 2. From our point of view - even 
though these requirements may need 
some more examination and assessment - 
we would prefer the conditions for 
specifying and adopting these 
requirements would be formulated 
directly in the AI act. 
 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
Article 4b(1) 
…When fulfilling those requirements, 
including as reflected in relevant harmonised 
Otherwise delegated acts might be an 
the  generally  acknowledged  state  of 
option. 
standards or common specifications. 
the  art  shall  be  taken  into  account, 
 
including  as  reflected  in  relevant 
 
 
harmonised  standards  or  common 
In general, the application of harmonised 
specifications. 
standards and common specifications is 
voluntary. By pointing out harmonised 
 
standards and common specifications in 
this context it sounds like a required 
mandatory application. 
Article 4b(5) 
In order to ensure uniform conditions 
 
If not regulated directly in the AI Act - 
for the implementation of this 
what we would prefer- we would suggest 
Regulation as regards the information to 
guidelines on the proposed cooperation of 
be shared by the providers of general 
providers. 
purpose AI systems, the Commission 
may adopt implementing acts in 
accordance with the examination 
procedure referred to in Article 74(2).  
 
Article 35 
2. 
The Commission shall make 
The Commission shall make publicly available the 
Regardless of the drafting suggestions the 
publicly available the list of the bodies 
German Federal Government is 
list of the bodies notified under this Regulation, 
notified under this Regulation, including 
concerned that the publication of the list 
the identification numbers that have been  including the identification numbers that have been  of notified bodies in the area of law 
assigned to them and the activities for 
enforcement, as provided for in Article 
assigned to them and the activities for which they 
which they have been notified. The 
35(2) of the AI Regulation, could 
Commission shall ensure that the list is 
have been notified accessible to the public in 
encourage illegal influence on or research 
kept up to date. 
of such bodies, for example by foreign 
NANDO. The Commission shall ensure that the list  services. Do the Commission or other 
is kept up to date.  
Member States share this view? Should 
an exemption clause be included to 
 
enable the Member State concerned, 
 
under conditions to be defined in more 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
detail, to refrain from publication in 
individual cases if and to the extent that 
interests in the protection of secrets 
conflict with this? DEU security 
authorities are in favour of this. 
Article 40 
 
 
In the area of law enforcement, there are 
and 41 
specific requirements for IT security, 
confidentiality, protection of fundamental 
rights and data protection as well as 
specific technical requirements. In DEU, 
it is being discussed whether it can be 
ensured that such specific requirements 
of the security sector can be taken into 
account within the framework of the 
standards according to Article 40 and the 
specifications according to Article 41 of 
the Draft Regulation. How do the COM 
and the other Member States see this? 
Should the amendments introduced in 
Article 41(2) ensure this for "commen 
specifications"? 
Art. 41 
Following the NLF, harmonized standards 
1.  
The Commission is 
1.  
The Commission is empowered to 
and hence CS can only deal with 
empowered to adopt,  after consulting 
adopt,  after consulting the AI Board referred to  essential requirements. Not all 
the AI Board referred to in Article  56,  in Article  56,  implementing  acts  establishing 
requirements of the Regulation or those 
implementing  acts  establishing 
common  technical  specifications  for  the 
that the regulator wishes to be covered.   
common  technical  specifications  for 
requirements set out in Chapter 2 of this Title, 
the requirements set out in Chapter 2 
or, as applicable, with requirements set  out  in 
 
of this Title, or, as applicable, with 
Article  4a  and  Article  4b,  where  the 
Implemting acts follow comitology rules: 
requirements set  out  in  Article  4a 
following  conditions  have  been fulfilled: 
The practical rules and general principles 
and  Article  4b,  where  the  following 
to be followed on comitology are laid 
conditions  have  been fulfilled: 
down in Regulation (EU) No 182/2011. 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
Accordingly, the Commission consults a 
(a) 
no  reference  to  harmonised 
(a) 
no  reference  to  harmonised  standards 
committee composed of representatives 
standards  covering  the  relevant 
covering  the  relevant  essential 
of all Member States and chaired by the 
essential  safety or fundamental right 
r e q u i r e m e n t safety or fundamental right 
Commission on draft implementing acts. 
concerns is published in the Official 
concerns is published in the Official Journal of 
These committees use 2 types of 
Journal of the European Union in 
the European Union in accordance with 
accordance with Regulation (EU) No 
Regulation (EU) No 1025/2012; 
procedures: 
1025/2012; 
-  examination and 
(b) 
the  Commission  has  requested  one 
-  advisory. 
(b) 
the  Commission  has 
or  more  European  standardization 
requested  one  or  more  European 
organisations to draft a harmonised standard for  These procedures differ in their voting 
standardization organisations to draft 
the requirements set out in Chapter 2 of this 
rules and in the way their votes influence 
a harmonised standard for the 
Title; 
the Commission’s possibilities to adopt 
requirements set out in Chapter 2 of 
the implementing act in question. The 
this Title; 
(c)  
the  request  has  not  been  accepted  by 
choice of procedure for a given act is 
any  of  the  European  standardization 
made by the EU legislator, and depends 
(c)  
the  request  has  not  been 
organisations or the standard is not delivered 
on the nature of the implementing 
accepted  by  any  of  the  European 
within the deadline 
powers that are laid out in the basic act 
standardization organisations or the 
(regulation, directive or decision). 
standard is not delivered within the 
Those implementing acts shall be adopted in 
 
deadline 
accordance with the examination procedure 
referred to in Article 74 (2) of this regulation. 
CS are to be a temporary safety-net if 
1a. 
Before  preparing  a  draft 
and only if harmonized standards do not 
implementing  act,  the  Commission 
1a. 
Before  preparing  a  draft 
exist. The Regulation has to clear about 
shall  inform  the committee referred 
implementing  act,  the  Commission  shall 
what happens if standards are then later 
to in Article 22 of Regulation EU (No) 
inform  the committee referred to in Article 22 
cited in OJEU and also about what 
1025/2012 that it considers that the 
of Regulation EU (No) 1025/2012 that it 
happens if the CS carry some problems 
conditions in paragraph 1 are fulfilled.  considers that the conditions in paragraph 1 are 
with respect to other Regulations or 
fulfilled. 
standards touched upon by the CS. 
2. 
In  the  early  preparation  of 
Hence there need to be more practical 
the  draft  implementing  act 
2. 
In  the  early  preparation  of  the  draft 
rules for CS in this Regulation. All copied 
establishing  the  common 
implementing  act  establishing  the  common 
from the Nonpaper’s Appendix.  
specification,  the  Commission  shall 
specification,  the  Commission  shall  fulfil  the 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
fulfil  the  objectives  referred  to  in 
objectives  referred  to  in  Article  40(2) and 
Article  40(2) and gather the views of 
gather the views of relevant bodies or expert 
relevant bodies or expert groups 
groups established under relevant sectorial 
established under relevant sectorial 
Union  law.  Based  on  that  consultation,  the 
Union  law.  Based  on  that 
Commission  shall  prepare  the draft 
consultation,  the  Commission  shall 
implementing act. 
prepare  the draft implementing act. 
The Commission, Wwhen preparing the common 
The Commission, Wwhen preparing the 
specifications referred to in paragraph 1, 
common specifications referred to in 
the Commission shall fulfil the objectives referred 
paragraph 1, 
of  Article  40(2)  and  gather  the  views  of  relevant 
the  Commission  shall  fulfil  the  bodies  or  expert  groups  established  under  relevant 
objectives referred of Article 40(2) and  sectorial Union law. 
gather  the  views  of  relevant  bodies  or   
expert  groups  established  under  relevant  3. 
High-risk AI systems or general purpose 
sectorial Union law. 
AI  systems  which  are  in  conformity  with  the 
 
common  specifications  referred  to  in  paragraph  1 
3. 
High-risk AI systems or general  shall  be  presumed  to  be  in  conformity  with  the 
purpose  AI  systems  which  are  in  requirements  set  out  in  Chapter  2  of  this  Title 
conformity 
with 
the 
common  or,  as  applicable,  with requirements set out in 
specifications  referred  to  in  paragraph  1  Article  4a  and  Article  4b,  to  the  extent  those 
shall  be  presumed  to  be  in  conformity  common specifications cover those requirements. 
with  the  requirements  set  out  in 
Chapter  2  of  this  Title  or,  as  4a. 
Where providers do not comply with the 
applicable,  with requirements set out  common specifications referred to in paragraph 1, 
in Article 4a and Article 4b, to the extent  they shall duly justify in the technical 
those common specifications cover those  documentation referred to in Article 11 that they 
requirements. 
have adopted technical solutions that are at least 
equivalent thereto. 
4. 
Where providers do not comply 
with the common specifications referred 
4b. When references of a harmonised standard are 
to in paragraph 1, they shall duly justify 
published in the Official Journal of the European 
in the technical documentation 
Union, implementing acts referred to in this Article, 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
referred to in Article 11 that they have 
or parts thereof which cover the same essential 
adopted technical solutions that are at 
requirements shall be repealed. 
least equivalent thereto. 
4c. When a Member State considers that a common 
 
specification does not entirely satisfy the essential 
 
reuqirements set out in this Regulation as stated 
above, it shall inform the Commission thereof with 
a detailed explanation and the Commission shall 
assess that information and, if appropriate, amend 
the implementing act establishing the common 
specification in question. 
 
Article 42 
 
3. For high-risk AI systems where the provider is a 
As the entities regulated by Directive 
credit institutions regulated by Directive 
2013/36/EU, Directive 2009/138/EC, 
2013/36/EU or an entity regulated by Directive 
Directive (EU) 2016/2341, Directive 
2009/138/EC, Directive (EU) 2016/2341, Directive  2014/65/EU, Directive (EU) 2015/2366, 
2014/65/EU resp. Directive (EU) 2015/2366, 
Directive 2009/65/EG resp. Directive 
Directive 2009/65/EG and Directive 2011/61/EU, 
2011/61/EU already follow highest 
conformity is assumed when these entities fulfill the  standards and double regulation has to be 
requirements following Directive 2013/36/EU, 
avoided, conformity of high-risk AI 
Directive 2009/138/EC, Directive (EU) 2016/2341,  systems provided by them should be 
Directive 2014/65/EU, Directive (EU) 2015/2366, 
assumed when they fulfill the respective 
Directive 2009/65/EG resp. Directive 2011/61/EU 
requirements to the extent that those 
to the extent those Directives cover the 
requirements cover the requirements set 
requirements set out in this Regulation. 
out in this Directive. 
Article 43 
 
set out in Chapter 2 of this Title this Regulation, 
Providers should demonstrate compliance 
with the requirements of the entire 
regulation in the conformity assessment 
procedure. The reference (only) to 
chapter 2 would, for example, omit the 
quality management system or the post-
market surveillance system. 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
Article 46(4) 
The obligations referred to in 
4. The obligations in this provision only apply as far  Information from the notified body can 
paragraphs 1 to 3 shall be complied 
as secrecy obligations do not conflict. 
only be transmitted as far as transmission 
with in accordance with Article 70. 
does not interfer with serecy obligations 
especially regarding operative scenarios.  
This may already be covered by Art. 70, 
but we would still like to emphasize that 
this is an important issue for us, wich 
may be addressed here as well. 
Article 47 
 
 
We suggest to add a paragraph to 
empower the EU-COM to extend the 
validity of an authorisation to the 
territory of the Union for a limited period 
of time by means of implementing acts. 
 
 
Article 
In a duly justified situation of urgency 
 
Germany is of the view that the 
47(1a) 
for exceptional reasons of public security 
presidency’s supplementary proposal is 
or in case of specific, substantial and 
generally reasonable. However, Germany 
imminent threat to the life or physical 
considers that the requirements of the 
safety of natural persons, law 
derogation regulation in Article 47 (1a) 
enforcement authorities or civil 
of the proposal are too unspecific. In 
protection authorities may put a specific 
particular, it remains unclear exactly 
high-risk AI system into service without 
what “duly justified situation of urgency 
the authorisation referred to in 
for exceptional reasons of public 
paragraph 1 provided that such 
security” means. Discussions are under 
authorisation is requested during or 
way in Germany regarding whether, from 
after the use without undue delay, and if 
the point of view of protection of 
such authorisation is rejected, its use 
shall be stopped with immediate effect. 
fundamental rights, provisions should be 
included for safeguards and an 
arrangement for the legal consequences 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
of violations of the rule. Moreover, 
discussions are also under way in 
Germany on whether the market 
surveillance authorities should be 
informed before the provisional putting 
into service in such cases, to enable 
verification of the criteria. From an 
operational point of view, the proposal 
also does not yet answer the question of 
the usability of intelligence obtained from 
the deployment of a non-certified AI 
system in urgent cases. Germany sees a 
need for this question to be addressed in 
the proposed regulation. As a whole, 
regulation of AI should include 
provisions to balance fundamental rights 
aspects with operational aspects in these 
urgent cases, providing legally secure 
certification that ensures that data and 
information from these systems can be 
used; in doing so, the AI Act should in 
particular stipulate the conditions under 
which the certification procedure affects 
the legitimacy of measures based on the 
provisional deployment of AI systems. 
Please refer to the separate position paper 
handed in, proposing necessary diverging 
regulations for public administration 
(especially LEAs and migration 
authorities) „[Regulation of AI – taking 
greater account of the specific 
characteristics of the public 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
administration, particularly in the fields 
of security and migration]“. 
Article 51 
 
 
There is a fear that the disclosure of all the 
law enforcement agencies' AI applications 
in operation or development will facilitate 
the assessment of an overall picture of the 
operational capabilities of the respective 
agencies. This database could potentially 
be used to identify capability gaps or to 
create thematic profiles of individual 
countries. This in itself could pose a 
security risk and affect the capabilities of 
the authorities. Do the Commission or 
other Member States share this view? The 
newly added para 2 is not sufficient to 
adress this issue.  
Please also refer to the separate position 
paper handed in, proposing necessary 
diverging regulations for public 
administration (especially  
LEAs and migration authorities) 
„[Regulation of AI – taking greater 
account of the specific characteristics of 
the public administration, particularly in 
the fields of security and migration]“ . 
 
Article 51(2) 
Before using a high-risk AI system, 
2. Before using an AI system, the relevant public 
We thank the presidency for introducing 
users of high-risk AI system that are 
authorities shall register the system used by the 
an obligation for public authorities to 
public authorities, agencies or bodies, 
public authority in the EU database referred to in 
with the exception of law enforcement,  Article 60a.   
register ths use of AI systems in a public 
border control, migration or asylum 
database. However, we would like to ask 
authorities, shall register themselves in 
the EU database referred to in Article 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
60 and select the system that they 
for some changes and refer to our 
envisage to use. 
comment regarding Art. 60a (new).  
Article 52 
 
NEW (1) The Commission is empowered to adopt 
In order to accommodate the AI-specific 
delegated acts in accordance with Article 73 to 
environmental and sustainability aspects, 
appropriate changes should be made. DE 
amend this Regulation by establishing, after having  proposes laying down horizontal 
consulted relevant stakeholders, a common Union 
transparency rules in Art. 52a in order to 
scheme for describing and rating the environmental  enable providers and users to lower the 
energy and resource consumption caused 
sustainability of AI systems placed on the market or  by the development and the application 
put into service in its territory. The scheme shall in 
of AI systems and to contribute to reach 
a first step, by one year after the entering into force  the goal of carbon neutrality. 
 
of this Regulation, establish a definition of AI 
This proposal of a horizontal 
systems’ sustainability and set out a small number 
transparency requirement aims at 
of easy-to-monitor indicators related, for example, 
reporting a limited number of easy-to-
to good practice through energy-efficient 
monitor sustainability indicators of AI 
systems. These might entail simple, 
programming or to data centre resource efficiency. 
binary statements on whether AI 
In a second step, by 2027, the scheme shall set up a  providers follow a good practice 
lean methodology to measure and rate AI systems 
regarding energy-efficient programming 
(‘green coding’) or whether the 
based on the indicators. The indicators and 
computing power originates from 
methodology shall be updated in light of technical 
certified data centres that, for example, 
progress. The scheme shall only concern direct 
generate own renewable energy, obtain 
green electricity, use waste heat or 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
environmental impacts of AI systems and may 
employ more sustainable cooling 
allow for exceptions for SMEs. 
techniques.  
The definition of sustainability indicators 
 
is best left to an expert committee; 
therefore, no pre-determinations should 
be made. This committee should also take 
into account the ease of monitoring and 
reporting to minimize burdens for AI 
providers, particularly SME providers. It 
should be composed of a broad range of 
experts from science, business, civil 
society and standardization organisations. 
It is conceivable that the AI Board may 
be involved or take over (a part of) the 
function of the expert committee. 
While we emphasize that many AI 
products lead to major environmental 
benefits, our goal is to ensure that the 
positive environmental outcome of an AI 
system is not, as an undesirable side 
effect, partially negated by poor energy 
and resource efficiency. The proposed 
reporting requirement firstly aims at 
creating incentives for AI providers to 
raise their sustainability ambitions and in 
the medium term, increase the demand 
for more sustainable computing power 
provision.  

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
Even though energy prices spike and 
chips become scarce, we do not witness 
significant changes in AI development 
practices, as other performance criteria 
than energy efficiency predominate in 
design and sourcing choices. In addition, 
it is often not transparent for AI 
developers, how much energy their 
models and programs actually consume, 
due to time-based, flat rate pricing of 
cloud services providers. Thus, as price 
signals do not sufficiently incentivize 
more sustainable practices, transparency 
requirements present a necessary 
additional and unrestrictive incentive.  
To highlight credible sustainability 
information also offers advantages in the 
marketing of AI systems. It is only 
through such information that a 
distinctiveness, ideally a unique selling 
proposition, can be established, which 
gives European providers a competitive 
advantage in the long run ("sustainable 
AI made in Europe"). 
Due to the pace of development and 
many crossroad decisions on the direction 
of AI development underway, the AI Act 
offers a timely and flexible option to 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
address sustainability issues of AI 
systems, in contrast to the long, complex 
and detailed procedures under the 
Ecodesign Regulation.  
Further clarifications are given in recital 
70. 
Article 52(3) 
However, the first subparagraph shall 
or it is necessary for the exercise of the right to 
The deleted sentence should be reinserted 
not apply where the use is authorised 
freedom of expression and the right to freedom 
(in suggested Titel IV A).  
by law to detect, prevent, investigate 
of the arts and sciences guaranteed in the 
 
and prosecute criminal offences or it is  Charter of Fundamental Rights of the EU, 
necessary for the exercise of the right 
to freedom of expression and the right 
to freedom of the arts and sciences 
guaranteed in the Charter of 
Fundamental Rights of the EU, and 
subject to appropriate safeguards for 
the rights and freedoms of third 
parties. 
 
 
TITLE IVA  
Please also move Art. 52(1) and (2) to 
INFORMATION TO BE PROVIDED TO 
this title as new Art. 52a. 
NATURAL PERSONS 
 
 
One of the primary reasons why AI is 
Art. 52b 
being regulated at all is to protect 
Information to be provided for high-risk AI 
individuals from the risks generated by 
systems 
AI systems to fundamental rights and to 
 
create trust. In this context, the need for 
1. Users of High Risk-AI systems shall provide 
transparency is one of the main factors 
the person affected by a decision at least 
explicitly addressed by the AIA. The 
partially determined by the output of the AI 
GDPR already grants certain rights to 
system (“Affected Person”) with standardized 
information to natural persons/data 
information about 
subjects. However, the GDPR does not 
 
sufficiently cover constellations where AI 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
(a) the fact that an AI system has been used 
systems are involved. For example, 
within the context of the decision-making 
Articles 22, 13 (2) (f), 14 (2) (g) and 15 
process; 
(1) (h) address automated processing, but 
 
these provisions only cover cases where 
(b) a reference to the EU-data base as referred to  natural persons are directly exposed to 
in Art 51, 60 and Annex VIII; 
automated decision making. This would – 
 
at least according to the wording of the 
(c) the general role and purpose of the AI system  GDPR - not cover the constellation that 
in the decision-making-process; 
AI is used to prepare a decision 
 
ultimately made by a human (for 
(d) the relevant data categories of the input data;  example, an AI might provide a credit 
 
rating score that a bank employee uses to 
(e) information provided to the user pursuant to  decide on the granting of a loan to a 
Article 13 paragraph 3 letters b and d; and 
natural person).This constellation may 
 
have consequences for the natural person 
(f) the right to receive an explanation upon 
that can be just as serious as where the 
request according to paragraph 3. 
natural person is directly exposed to 
 
automated processing, and gives 
The information shall be provided at the time 
therefore rise to a similar need for 
the decision is communicated to the affected 
protection. It is necessary for an affected 
natural person. 
natural person to understand the risks 
 
which they are being subjected to in order 
2. Where a high-risk AI system is used for 
to be able to seek redress.  
automated individual decision-making, including  Therefore, we consider it necessary to 
profiling, within the meaning of Article 22 of 
include an obligation of the user to 
Regulation (EU) 2016/679, information 
provide the affected natural person with 
according to Articles 13 (2) (f), 14 (2) (g) and 15 
standardized information on the use and 
(1) (h) of Regulation (EU) 2016/679 shall also 
general function of the AI system and to 
comprise information according to paragraph 
include a substantive right for affected 
(1) (b), (d), (e) and (f). 
persons to request further information on 
 
the input data and the relevant data 
categories, in constellations, where an AI 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
3. Users of high-risk AI systems shall provide the  system is used to prepare a human 
affected natural person upon his or her request 
decision.  
in addition to the standardized information 
We also consider it necessary to 
provided according to paragraph 1 with concise,  supplement the existing information 
complete, correct and clear explanation of the 
requirements in the GDPR with some 
individual input data relating to the affected 
further information that seem necessary 
natural person and the relevant data categories 
specifically in the context of AI systems 
on the basis of which the decision was made. 
in order to provide natural persons with 
 
all relevant knowledge to understand 
4. Paragraph 1 (e),2 insofar as it refers to 
their situation.  
paragraph 1 lit. (e) and paragraph 3 shall not 
With the suggested Article 52b, we aim 
apply to the use of AI systems that are 
to avoid any duplication or overlapping 
authorised by law to detect, prevent, investigate 
with existing rights under the GDPR, but 
and prosecute criminal offences or prevent of a 
merely to supplement them only to the 
threat to critical infrastructure, life, health or 
extent necessary, as it is important to 
physical safety of natural persons. 
avoid legal uncertainty regarding the 
 
relation of this Regulation to the GDPR. 
5. Paragraph 1 to 3 shall not apply to the use of 
 
AI systems 
In addition to the implementation of Art. 
 
52b new, the following sentence should 
(a) for which exceptions from, or restrictions to, 
be added to Recital 43: „Natural persons 
the obligations under this Article follow from 
affected by decisions at least partially 
Union or Member State law (such as a 
determined by high-risk AI systems (this 
prohibition or restriction to disclose information  includes decisions that were made after a 
covered by paragraph 1 and 2 to the affected 
high-risk AI system provided a 
person), which lays down appropriate other 
recommendation for the decision)  placed 
safeguards for the affected person's rights and 
on the EU market or otherwise put into 
freedoms and legitimate interests when such an 
service should be informed in an 
exception or restriction respects the essence of 
appropriate, easily accessible and 
the fundamental rights and freedoms and is a 
comprehensible manner about the use of 
necessary and proportionate measure in a 
the AI system, the role and purpose of the 
democratic society; or 
AI system in the decision-making 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
(b) that have only minor influence within the 
process, the logic involved and the main 
decision-making process. 
parameters of decision making. Such 
 
information could be provided in 
6. Information according to paragraph 1 to 3 
electronic form, for example, when 
shall be given in a concise, transparent, 
addressed to the public, through a user‘s 
intelligible and easily accessible form 
website while providing the link to this 
appropriate to different kinds of disabilities, 
website at the time the decision is 
using clear and plain language. 
communicated to the affected person. For 
 
this purpose, with regard to the 
Article 52c 
standardized information to be provided 
Relation to Title III 
under para. 1 and 2, the user should be 
 
able utilise the information received from 
Obligations under this Title shall not affect the 
the provider according to article 13 
requirements and obligations set out in Title III 
paragraph 3 letters b and d. With regard 
of this Regulation. 
to the individual explanation according to 
 
para. 3, the affected natural person must 
be provided with the individual input data 
relating to the affected natural person and 
the relevant data categories that serve as 
the main parameters on the basis of 
which the output was given.” 
 
Furthermore, the term “affected natural 
person” should be defined in Art. 3 AIA.  
 
The rights of the persons affected are 
limited by the wording to individual 
persons. This does not include the 
protection or representation of collective 
interests. This means that particularly 
vulnerable groups or groups at risk of 
discrimination can exercise their rights 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
less effectively. Possibilities for 
collective enforcement still need to be 
examined within the federal government. 
 
Generally, it has to be made sure that 
Union or Member State law containing 
prohibitions of disclosure or relevant 
restrictions on the affected person’s right 
of access to the information covered by 
Art. 52a (new) remains unaffected, 
especially in the area of law enforcement.  
For example: In case of suspicion of 
money laundering, the competent 
authority (Financial Intelligence Unit, 
“FIU”) is prohibited to reveal information 
to the affected person (based on Art. 41 
para. 4 EU-act ). Therefore, we suggest to 
add para. 4 or a similar provision inspired 
by Art. 23 GDPR saying that the 
obligations or rights granted under Art. 
52 a (new) can be restricted by Union or 
Member State law that e.g. prohibits or 
restricts the user of the AI system to 
reveal the information, provided that such 
a prohibition or restriction respects the 
essence of the fundamental rights and 
freedoms and is a necessary and 
proportionate measure in a democratic 
society. 
 
We are still discussing details on this 
provision. 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
 
Paragraph 2 only covers situations that 
are already covered by automated 
processing in accordance with Article 22 
of the GDPR (i.e., constellations where a 
natural person is exposed directly to an 
AI system). In these constellations, 
information obligations under the GDPR 
are extended to certain further, AI 
specific information according to 
paragraph 1. 
 
We are still discussing if AI systems used 
for the prevention of a threat to critical 
infrastructure, life, health, physical safety 
of natural persons or public safety should 
also be excluded from the obligations of 
par. 1-3 and may comment on this later. 
 
This corresponds to the current Article 
52(4) and should apply to the entire title. 
 
 
Article 53 
1a. 
National competent authorities 
1a. 
National competent authorities may 
Add “under the direct supervision, 
may establish AI regulatory sandboxes  establish AI regulatory sandboxes for the 
guidance [..] by the national competent 
for the development, training, testing 
development, training, testing and validation of 
authority”. The key element of 
and validation of innovative AI 
innovative AI systems under the direct 
supervision and guidance by the national 
systems, before their placement on the 
supervision, guidance and support by the 
competent authority was deleted by 
market or putting into service. Such 
national competent authority, before their 
deleting the whole article 53 (1) and 
regulatory sandboxes may include 
placement on the market or putting into service.  should be returned. 
testing in real world conditions 
Such regulatory sandboxes may include testing 
 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
supervised by the national competent 
in real world conditions supervised by the 
Add “support”: Especially for start-ups it 
authorities. 
national competent authorities. 
is very important that competent 
authorities – within their legal 
possibilities – act as supporters in 
ensuring compliance, e.g. through 
mentoring, personal exchange or 
customized guidance. The impressive 
examples of data regulatory sandboxes by 
the French CNIL and the British ICO also 
explicitly “support” the projects. The 
term “support” is also used in EU 
Commission’s Better Regulation Toolbox 
Tool #69 on regulatory sandboxes (page 
597). 
Article 
d) 
enhance authorities’ 
d) 
enhance authorities’ understanding of the  To prevent diverging approaches, the key 
53(1b) 
understanding of the opportunities and  opportunities and risks of AI systems as well as 
objective of regulatory learning with its 
risks of AI systems as well as of the 
of the suitability and effectiveness of the 
different facets (better understanding of 
suitability and effectiveness of the 
measures for preventing and mitigating those 
opportunities and risks, contribution to 
measures for preventing and 
risks; 
effective implementation and 
mitigating those risks;  
development of standards and 
specification) should be returned. The 
Council conclusion on regulatory 
sandboxes (para 10) as well as the 
Commission’s Better regulation toolkit 
(page 595) highlight regulatory learning 
as crucial feature of regulatory 
sandboxes. 
 
e) 
contribute to the uniform and 
e) 
contribute to the uniform and effective 
see comment on d) 
effective implementation of this 
implementation of this Regulation and, where 
Regulation and, where appropriate, its  appropriate, its evidence based swift adaptation, 
swift adaptation, notably as regards 
notably as regards the techniques in Annex I, the 
the techniques in Annex I, the high-

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
risk AI systems in Annex III, the 
high-risk AI systems in Annex III, the technical 
technical documentation in Annex IV; 
documentation in Annex IV; 
 
f) 
contribute to the development 
f) 
contribute to the development or update 
see comment on d) 
or update of harmonised standards 
of harmonised standards and common 
and common specifications referred to  specifications referred to in Articles 40 and 41 
in Articles 40 and 41 and their uptake 
and their uptake by providers. 
by providers. 
 
 
g)     contribute to the possible future evidence-
Add new para: The Council conclusion 
based advancement of this Regulation and, 
on regulatory sandboxes (para 10) as well 
where appropriate, of other Union and Member 
as the Commission’s Better regulation 
States legislation through regulatory learning. 
toolkit (page 595) highlight regulatory 
learning as crucial feature of regulatory 
sandboxes. Regulatory sandboxes should 
contribute to resilient and relevant 
legislation through facilitating regulatory 
learning. 
Article 
 
Participation in the AI 
 
Participation in the AI regulatory 
The requirements to the specific plan 
53(2a) 
regulatory sandbox shall be based on a  sandbox shall be based on a specific plan 
should be returned. The lessons from 
specific plan referred to in paragraph 
referred to in paragraph 6 of this Article that 
sandboxes are only comparable if there is 
6 of this Article that shall be agreed 
shall be agreed between the participant(s) and 
a common framework. Harmonizing the 
between the participant(s) and the 
the national competent authoritie(s) or the 
rules concerning this specific plan of 
national competent authoritie(s) or the  European Data Protection Supervisor, as 
participation helps regulatory learning as 
European Data Protection Supervisor,  applicable. The plan shall contain as a minimum  well. It is important to have a clear 
as applicable. The plan shall contain as  the following: 
objective in mind when operating a 
a minimum the following: 
regulatory sandbox.  
If the context of participation is 
documented well, it is easier to compare 
the results of the sandbox with sandboxes 
that have taken place under the 
supervision of other national competent 
authorities 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
 
a) 
description of the participant(s)  a) 
description of the participant(s) involved 
See comment above. 
involved and their roles, the envisaged 
and their roles, the envisaged AI system and its 
AI system and its intended purpose, 
intended purpose, and relevant development, 
and relevant development, testing and 
testing and validation process; 
validation process; 
 
b) 
the specific regulatory issues  at  b) 
the specific regulatory issues  at stake and  See comment above; add “support”: 
stake and the guidance that is expected  the guidance and support that is expected from 
Especially for start-ups it is very 
from the authorities supervising the AI  the authorities supervising the AI regulatory 
important that competent authorities – 
regulatory sandbox;  
sandbox;  
within their legal possibilities – act as 
supporters in ensuring compliance, e.g. 
through mentoring, personal exchange or 
customized guidance.  
 
 
bb) the novelty of the specific regulatory issue, 
Additionally we propose a new provision 
compared to the annual reports referred to in 
2a(bb). Note that this does not require 
Article 53(5), and whether analyzing this 
regulatory issue in the regulatory sandbox 
participants to have a novel regulatory 
contributes to the objectives of Article 53(1b)(c) 
issue in order to participate in the 
and (d); 
sandbox. Whether a regulatory issue is 
novel can also become clear during the 
sandbox. 
 
 
c) 
the specific modalities of the 
c) 
the specific modalities of the 
See comment above. 
collaboration between the 
collaboration between the participant(s) and the 
participant(s) and the authoritie(s), as 
authoritie(s), as well as any other actor involved 
well as any other actor involved in the 
in the AI regulatory sandbox; 
AI regulatory sandbox; 
 
d) 
a risk management and 
d) 
a risk management and monitoring 
See comment above. 
monitoring mechanism to identify, 
mechanism to identify, prevent and mitigate any 
prevent and mitigate any risk referred  risk referred to in Article 9(2)(a); 
to in Article 9(2)(a); 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
 
 
(dd)     obligations for the participants to provide  An evaluation on the basis of current and 
the authority with information needed for the 
accurate data is crucial in order to 
authority’s evaluation of the project. 
enhance authorities’ understanding and to 
allow for regulatory learning. COM’s 
better regulation toolkit p. 597 stresses 
that the main evaluation criteria (and that 
includes also the data and data source) 
should be established ex ante. 
 
e) 
the key milestones to be 
e) 
the key milestones to be completed by the  See comment above. 
completed by the participant(s) for the  participant(s) for the AI system to be considered 
AI system to be considered ready to 
ready to exit from the regulatory sandbox. 
exit from the regulatory sandbox. 
 
 
(2b) After an AI regulatory sandbox has ended, 
In various national regulatory sandboxes, 
it  is common practice to issue an exit 
the participant(s) and the national competent 
report after the sandbox has concluded. 
authoritie(s) or the European Data Protection 
We propose to include this practice in the 
AI Act as well. The exit reports focus 
Supervisor, as applicable, shall draw up an exit 
more specifically on the case at hand, 
report. This exit report shall contain as a 
instead of the more vaguely drafted 
‘annual reports’ (which also focus on the 
minimum the following: 
implementation of sandboxes).  
a) 
The plan referred to in paragraph 2a of 
In order to truly utilize lessons learnt, 
they must first be defined. The national 
this Article;  
competent authorities are in the best 
b) 
An evaluation of the specific regulatory 
position to do this, right after a sandbox 
has ended.  
issues that were at stake during the AI 
Under paragraph 5a, the exit reports will 
regulatory sandbox, including a problem 
then be used by the AI Board and 
Commission to improve interpretation, 
definition and proposed solutions; 
guidance, communication and 
amendments regarding this Regulation.   

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
c) 
Whether the key milestones referred to in 
paragraph 2a(e) of this Article have been 
completed; 
d) 
A conclusion on the lessons learnt, 
specified in the following categories of use:  
1.  An improved understanding on the 
implementation of the AI regulatory 
sandboxes; 
2.  Improved methods of supervision by 
national competent authorities; 
3.  A revised or novel interpretation of this 
Regulation. 
 
 
3. 
The participation in the AI 
with the objective of supporting innovation in AI in  The former formulation should be 
regulatory sandboxes shall not affect the 
the Union. Any significant risks to health and safety  reinserted and amended since significant 
supervisory and corrective powers of the 
and fundamental rights identified during the 
risks to health and safety and 
competent authorities supervising the 
development and testing of such systems shall 
fundamental rights require immediate 
sandbox. Those authorities shall 
result in immediate mitigation and, failing that, in 
mitigation (and failing that suspension). 
exercice their supervisory powers in a 
the suspension of the development and testing 
The authorities shall support the 
flexible manner within the limits of the  process until such mitigation takes place. The 
participants in developing and 
relevant legislation, using their 
authorities shall cooperate with the participants of 
implementing the mitigation. 
discretionary powers when 
the sandbox to develop and implement a mitigation   
implementing legal provisions to a 
plan to enable a resumption of the testing process 
 
specific AI sandbox project., with the 
without undue delay. 
objective of supporting innovation in 
AI in the Union Any significant risks to 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
health and safety and fundamental rights 
identified during the development and 
testing of such systems shall result in 
immediate mitigation and, failing that, in 
the suspension of the development and 
testing process until such mitigation takes 
place.  
 
 
However, pProvided that the 
 
We welcome the amendment. However, 
participant(s) respect the sandbox plan 
in order to ensure the protection of 
and the terms and conditions for their 
fundamental rights, we remain critical of 
participation as referred to in 
the fact that fines should continue to be 
paragraph 6(c) and follow in good faith 
excluded even in the case of 
the guidance given by the authorities, 
infringements that lead to high risks for 
no administrative enforcement action 
the rights and freedoms of natural 
shall be taken fines shall be imposed by 
persons. 
the authorities for infringement of 
 
applicable Union or Member State 
Legal Council Service should verify if 
legislation, including the provisions of 
with regard to the financial market a 
this Regulation.  
sector specific clarification is necessary. 
The rules and provisions for participation 
in a sandbox program should not 
contradict the harmonized financial 
market regulation. We would therefore 
advise to consult with COM (DG 
FISMA) on this specific financial sector 
related question.  
 
 
 
5a. 
To ensure that sandboxes will deliver 
more than vaguely defined annual 
1. After an AI regulatory sandbox has ended, the  reports, this paragraph requires the AI 
national  competent  authority  shall  share  the  Board and Commission to utilize the exit 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
exit report of that sandbox with the AI Board  reports that have been drawn by the 
national competent authorities.  
and the Commission.  
 
2. The  AI  Board  shall use  the  annual  reports  of  As these exit reports may contain 
sensitive information that should be kept 
paragraph 5 of this Article and the exit reports  confidential, an explicit reference to 
it  recieves  according  to  paragraph  1  in  the  Article 70 has been made. This also 
prevents a situation in which participants 
exercise of its tasks as listed in Article 58.  
may be reluctant to participate in 
3. The Commission shall  use the annual reports  sandboxes because they are afraid that 
their trade secrets or other sensitive 
of  paragraph  5  of  this  Article  and  the  exit  information will be made public. 
reports it recieves according to paragraph 1 in 
the  exercise  of  its  tasks  in  Articles  4,  7,  11(3) 
and 58a.  
The exit reports shall be shared on a confidential 
basis and in accordance with Article 70.  
Art. 53(6) 
6. 
The detailed modalities and the 
[..] with the examination procedure referred to in 
Add “These modalities and conditions 
conditions for the establishment and of  Article 74(2). These modalities and conditions shall  shall foster innovation and shall take into 
the operation of the AI regulatory 
foster innovation and shall take into account 
account particularly the special 
sandboxes under this Regulation, 
particularly the special circumstances of 
circumstances of participating small and 
including the eligibility criteria and the 
participating small and medium-sized enterprises. 
medium-sized enterprises”: The 
procedure for the application, selection, 
objectives of the regulatory sandboxes 
participation and exiting from the 
should be to foster AI innovation (recital 
sandbox, and the rights and obligations of 
71). In order to promote innovation, it is 
the participants shall be set out in 
important that the interests of small-scale 
implementing acts. Those implementing 
providers are taken into particular 
acts shall be adopted through 
account (recital 73. Both must be 
implementing acts in accordance with 
reflected in the regulatory sandboxes’ 
modalities and conditions. 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
the examination procedure referred to in 
Article 74(2). 
Art. 53(6) 
 
(bb) provisions for a possible subsequent 
Add “provisions for a possible 
introduction into permenant operation 
subsequent introduction into permanent 
operation”: In order to provide innovators 
with transparent and reliable investment 
conditions, perspectives for scaling the 
AI systems outside the regulatory 
sandbox should be set up. 
 
 
 
(cc) The modalities for the evaluation of the 
As emphasized in recital 72, one 
sandbox and the transfer of results into 
objectives of the regulatory sandboxes is 
legislative process; 
to enhance the competent authorities’ 
oversight and understanding. EU 
Commission’s Better Regulation Toolkit 
(page 595 and 597) and the Council 
Conclusions on Reg. Sandboxes (para 10) 
also stress the objective of advancing 
regulation through regulatory learning. 
To achieve this overarchingly, clear rules 
shall be set up. 
 
7. 
When national competent 
7. When national competent authorities consider  In order to avoid fragmentation, national 
authorities consider authorising testing  authorising testing in real world conditions 
competent authorities should not draw up 
in real world conditions supervised 
supervised within the framework of an AI 
their own terms and conditions but 
within the framework of an AI 
regulatory sandbox established under this 
should ensure the compliance with the 
regulatory sandbox established under 
Article, they shall ensure that the testing in real 
relevant requirements of Article 54a and 
this Article, they shall specifically 
world conditions takes place according to the 
54b when allowing testing in real world 
agree with the participants on the 
requirements of Articles 54a and 54b specifically  conditions in an AI regulatory sandbox. 
terms and conditions of such testing 
agree with the participants on the terms and 
and in particular on the appropriate 
conditions of such testing and in particular on 
safeguards. Where appropriate, they 
the appropriate safeguards. Where appropriate, 
shall cooperate with other national 
they shall cooperate with other national 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
competent authorities with a view to 
competent authorities with a view to ensure 
ensure consistent practices across the 
consistent practices across the Union. 
Union. 
Art. 54(1) 
(ii) 
public safety and public health, 
(ii) 
public safety, long-term care and public 
In the current draft, AI in the field of care 
including disease prevention, control and  health, including disease prevention, control and 
cannot be trained and supported in 
treatment of disease and improvement 
treatment of disease and improvement of health 
regulatory sandboxes. Therefore we 
of health care systems; 
care systems; 
propose the addition of “term long-term 
care”. 
 
(v) 
a high level of efficiency and 
(v) 
a high level of efficiency and quality of e-
“public administration and public 
quality of public administration and 
governmentpublic administration and public 
services” seems to be far too vague and 
public services.; 
services.; 
should be replaced by “e-government”. 
According to the GDPR, a legal basis for 
the processing of personal data should be 
clear and precise. The inclusion of the 
public sector could be ensured by our 
proposal “e-government”. 
 
 
(vii) ensuring or increasing data protection and data  We suggest to add this purpose to the list. 
security in AI systems or other technology; 
There are risks for data protection and 
security regarding technology in general, 
but also AI (such as membership attacks), 
and counter-measures are currently still 
under research. It would be in the public 
interest to foster such research as well by 
providing regulatory sandboxes. This 
could also help providers increase legal 
certainty (Art. 53 (1b) (c)) and reduce 
risks and costs by defining and 
implementing appropriate mitigation 
measures. 
Art. 54a(5) 
5.  
Any subject of the testing in 
 
Since "informed consent" is a different 
real world conditions, or his or her 
form of consent than in Art. 6 (1) lit a, 7 
legally designated representative, as 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
appropriate, may, without any 
GDPR we might suggest clarifying this in 
resulting detriment and without having 
a recital.  
to provide any justification, withdraw 
Further, we suggest that the possibility to 
from the testing at any time by 
revoke consent be included directly in the 
revoking his or her informed consent. 
definition in Art. 3 (51). 
The withdrawal of the informed 
 
consent shall not affect the activities 
already carried out and the use of data 
obtained based on the informed 
consent before its withdrawal. 
Art. 56 
Article 56 
 
WE CONTINUE TO SUGGEST AN 
Establishment and structure of the 
ORIENTATIONAL DEBATE 
European Artificial Intelligence Board 
REGARDING THE AI BOARD. THIS 
DEBATE SHOULD INCLUDE THE 
GENERAL ALIGNEMENT OF THE 
BOARD AND THE SCOPE, ITS 
MEMEBERS AND POSSIBLE RULES 
OF PROCEDURE. WE WOULD LIKE 
TO RESERVE THE OPPORTUNITY 
TO MAKE FURTHER COMMENTS 
AFTER THAT PROPOSED DEBATE. 
Art.  60 
Article 60 
 
The German Federal Government is 
EU database for stand-alone high-risk AI 
concerned that public access to the EU 
systems listed in Annex III 
database of high-risk AI applications 
provided for in Article 60 (5) of the AI 
Regulation could clash with justified 
security interests of the Member States. 
There are fears that even the publication 
of all AI applications that are operated or 
under development by the security 
authorities would make it easier to gain 
an overall picture of the operational 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
capabilities of the authorities in question. 
Using this database, potential gaps in 
capabilities could be identified, or 
profiles of the focus areas of individual 
countries could be compiled. This could 
represent a security risk in itself and 
could impact the capabilities of the 
authorities. Do the Commission or the 
other Member States share this 
assessment? Should an exception be 
included in this regard in Articles 60 and 
61 enabling the Member State in question 
under certain conditions, which must first 
be defined in more detail, to refrain in 
individual cases from publishing AI 
applications where security interests are 
at odds with this? German security 
authorities are in favour of this. Please 
also refer to the separate position paper 
handed in, proposing necessary diverging 
regulations for public administration 
(especially LEAs and migration 
autoritiess) „[Regulation of AI – taking 
greater account of the specific 
characteristics of the public 
administration, particularly in the fields 
of security and migration]“. 
 
 
Article 60a  
Due to the unique role and responsibility 
public authorities bear, the sensitive 
EU database for stand-alone AI systems used by 
personal data they have access to, the 
public authorities 
consequentional effects their decisions 
have on individuals, and thus their 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
1. The Commission shall, in collaboration with the  primary obligation to respect, protect and 
fulfil fundamental rights, public 
Member States, set up, maintain and manage a EU  authorities should be subject to more 
database  to  enable  the  public  to  be  adequately  stringent transparency requirements when 
using AI systems. Hence, any 
informed about AI systems placed on the market and  deployments of AI systems – regardless 
containing  information  referred  to  in  paragraph  2  of their level of risk – by or on behalf of 
public authorities should be registrered 
concerning AI systems used by the public authorities  within a separate EU database in addition 
registered in accordance with Article 51 (2). 
to the registration as High Risk AI in the 
database referred to in Article 60.  
 
 
2.  The  data  listed  in  Annex  VIIIb  shall  be  entered  We refrain from drafting up an Annex 
VIIIb for this comment. However, the 
into the EU database by the public authorities. The  data base should include the name of the 
Commission shall provide them with technical and  AI system and a short description of its 
intended purpose as well as the name, 
administrative support.  
address and contact details of the public 
 
authority by whom or on whose behalf it 
is used. However, in the field of law 
3. Art. 60 par. 3-5a shall apply accordingly.   
enforcement, the possible security risk 
 
arising from the database must also be 
considered. Please also refer to the 
comment above (no. 130) and the 
separate position paper handed in, 
proposing necessary diverging 
regulations for public administration 
(especially LEAs and migration 
authorities) „[Regulation of AI – taking 
greater account of the specific 
characteristics of the public 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
administration, particularly in the fields 
of security and migration]“. 
 
We are also still discussing this topic 
under aspects such as operating expenses, 
especially how to avoid exceeding 
operating expenses, as well as how often 
the database should be updated. 
 
Art. 61(2) 
In order to allow the provider to 
 
There is concern that the mere disclosure 
evaluate the compliance of AI systems 
of all law enforcement agencies' AI 
with the requirements set out in Title 
applications in operation or development 
III, Chapter 2 throughout their life cycle, 
will facilitate an assessment of an overall 
Tthe post-market monitoring system 
picture of the relevant agencies' 
shall actively and systematically collect, 
operational capabilities. Should an 
exception be included in Articles 61, 
document and analyse relevant data, 
which would allow the Member State 
which may be provided by users or 
concerned, under conditions to be defined 
which may be collected through other 
in more detail, to refrain from publication 
sources on the performance of high-risk 
in individual cases if and to the extent 
AI systems. This obligation shall not 
that confidentiality interests conflict with 
cover sensitive operational data of 
this? DEU security authorities are in 
users of AI systems which are law 
favor of this. 
enforcement authorities. throughout 
their life time and allow the provider to 
evaluate the continuous compliance of 
AI systems with the requirements set out 
in Title III, Chapter 2. 
Art. 61(4) 
For high-risk AI systems covered by the 
 
It remains unclear whether the AI Act 
legal acts referred to in Annex II, Section 
poses additional requirements for entities 
A, where a post-market monitoring system 
already regulated by comprehensive 
and plan is already established under that 
financial sector regulation. Please specify 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
legislation, the elements described in 
how the AI Act does avoid double 
paragraphs 1, 2 and 3 shall be integrated 
regulation for the highly regulated 
into that system and plan as appropriate the 
financial sector while also avoiding 
post-market monitoring documentation 
regulation gaps. 
as prepared under that legislation shall 
be deemed sufficient, provided that the 
template referred to paragraph 3 is used.  
The first subparagraph shall also apply 
high-risk AI systems referred to in point 
5(b) of Annex III placed on the market or 
put into service by credit financial 
institutions that are subject to 
requirements regarding their internal 
governance, arrangements or processes 
under Union financial services 
legislation.regulated by Directive 
2013/36/EU. 
Art. 62(4) 
For high-risk AI systems which are 
For high-risk AI systems which are safety 
In the Regulation (EU) 2017/745 and 
safety components of devices, or are 
Regulation (EU) 2017/746 there are 
components of devices, or are themselves 
themselves devices, covered by 
serious incidents that directly or 
Regulation (EU) 2017/745 and 
devices, covered by Regulation (EU) 2017/745 
indirectly led into death of patient, user 
Regulation (EU) 2017/746 the 
or other persons. It is important that 
and Regulation (EU) 2017/746 the notification of 
notification of serious incidents shall be 
every notification of serious incident is 
limited to those referred to in Article 
serious incidents shall be limited to those 
reported to the competent authority for 
3(44)(c) and be made to the national 
Medical Devices or IVD. If the serious 
referred to in Article 3(44)(c) and be made to the 
supervisory competent authority 
incident is limited to those referred to in 
chosen for this purpose by of the 
national supervisory competent authority chosen  Article 3(44)c the national competent 
Member States where that incident 
authority for Medical Devices or IVD 
for this purpose by of the Member States where 
occurred. 
forward the notification of serious 
that incident occurred under this legislation. If 
incidents to the competent authority for 
this purpose. 
the serious incidents is limited to those referred 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
to in Article 3(44)c the national competent 
authority referred to in the precending sentence 
shall be forward the notification of serious 
incidents to the competent authority for this 
purpose.  
Art. 62a 
 
CHAPTER 2A 
Regarding the data access for vetted 
researchers, in the field of law 
 
enforcement, the possible security risk 
DATA ACCESS FOR VETTED RESEARCHERS 
arising from data access must also be 
considered. Therefore, there is a need for 
 
excemptions from the requests laid down 
Article 62a 
in article 62a. These excemptions are still 
under discussion. We may provide 
Data Access for vetted Researchers 
further comments. 
 
1 Upon a reasoned request from a public or private 
body to be determined by each member state, 
providers shall within a reasonable period, as 
specified in the request, provide access to training, 
validation and test-ing datasets used by the provider 
to vetted researchers who meet the requirements in 
para-graph 2 of this article for the sole purpose of 
conducting research that contributes to the 
development, training, validation and testing of AI 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
systems within the existing legal framework, in 
particular with regards to bias monitoring, detection 
and correction of such systems and that is related to 
a public interest. Access to personal data shall be 
provided in anonymised or at least pseudonymised 
form as long as this is possible without jeopardizing 
the research pur-pose. 
 
2 Upon a duly substantiated application from 
researchers, the responsible body shall award them 
the status of vetted researchers and issue data access 
re-quests pursuant to paragraph 1, where the 
researchers demonstrate that they meet all of the 
following conditions: 
 
(a) researchers shall be affiliated to a research 
organisation as defined in Article 2 (1) of Di-rective 
(EU) 2019/790 of the European Parliament and of 
the Council 
(c) the application submitted by the researchers 
justifies the necessity and proportionality for the 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
purpose of their research of the data requested and 
the timeframes within which they re-quest access to 
the data, and they demonstrate the contribution of 
the expected research results to the purposes laid 
down in paragraph 1, 
() the planned research activities will be carried out 
only for the purposes laid down in para-graph 1,  
(f) shall commit and be in a capacity to preserve the 
specific data security and confidentiali-ty 
requirements corresponding to each request. In 
particular, a protection concept shall be provided 
with the request, containing a description of the 
research purpose, the intended use of the 
information, measures taken to protect the interests 
of the data subject and tech-nical and organisational 
measures taken to protect personal data.  
 
3 The provider may refuse the requested 
information, if trade secrets are affected and the 
public interest in the research does not outweigh the 
interest in confidentiality. The provid-er may refuse 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
access to personal data, if the legitimate interests of 
the data subject outweigh the public interest in the 
research. Where the data holder claims 
compensation for making data available, such 
compensation shall not exceed the technical and 
organisational costs incurred to comply with the 
request including, where necessary, the costs of 
anonymisation and of technical adaptation. 
 
4 The public or private body that awarded the status 
of vetted researcher and issued the access request in 
favour of a vetted researcher shall issue a decision 
terminating the ac-cess if it determines, following 
an investigation either on its own initiative or on the 
basis in-formation received from third parties, that 
the vetted researcher no longer meets the condi-
tions set out in paragraph 2. Before terminating the 
access, the body shall allow the vetted researcher to 
react to the findings of its investigation and its 
intention to terminate the access. As soon as the 
vetted researcher no longer meets the conditions set 
out in paragraph 2, the vetted researcher shall report 
this circumstance to the market surveil-lance 
authority. 
Art. 64 
5. 
Where the documentation referred  5. Where the documentation referred to in 
 
to in paragraph 3 is insufficient to 
paragraph 3 is insufficient to ascertain whether a 
ascertain whether a breach of obligations  breach of obligations under Union law intended to 
under Union law intended to protect 
protect fundamental rights, including the right to 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
fundamental rights has occurred, the 
non-discrimination, has occurred, the public 
public authority or body referred to 
authority or body referred to paragraph 3 may make 
paragraph 3 may make a reasoned request  a reasoned request to the market surveillance 
to the market surveillance authority to 
authority to organise testing of the high-risk AI 
organise testing of the high-risk AI 
system through technical means. The market 
system through technical means. The 
surveillance authority shall organise the testing with 
market surveillance authority shall 
the close involvement of the requesting public 
organise the testing with the close 
authority or body within reasonable time following 
involvement of the requesting public 
the request. 
authority or body within reasonable time 
following the request.  
Art. 70 
 
 
In DEU, although lit (e) has been 
included in Art. 70 (1) and references to 
Article 70 (2) have been inserted at 
various points in the text, there is still 
discussion on whether further 
requirements on secrecy and the 
guarantee of confidentiality and data 
protection are necessary. For example, 
Art. 70(2) of the AI Regulation provides 
only few concrete requirements for 
confidentiality and only for the case that 
authorities from the law enforcement 
sector are themselves developers or 
providers of AI applications. We refer to 
the separate position paper handed in, 
proposing necessary diverging 
regulations for public administration 
(especially LEAs and migration 
authorities) Regulation of AI – taking 
greater account of the specific 
characteristics of the public 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
administration, particularly in the fields 
of security and migration]. 
Art. 70(2) 
Without prejudice to paragraph 1, 
(1) 
shall not be disclosed without the prior 
Mere consultation in Art. 70 (2) is not 
information exchanged on a confidential 
sufficient. 
consultation approval (…) 
basis between the national competent 
authorities and between national 
 
competent authorities and the 
Commission shall not be disclosed 
without the prior consultation of the 
originating national competent authority 
and the user when high-risk AI systems 
referred to in points 1, 6 and 7 of Annex 
III are used by law enforcement, border 
control, immigration or asylum 
authorities, when such disclosure would 
jeopardise public and national security 
interests. This obligation to exchange 
information shall not cover sensitive 
operational data in relation to the 
activities of law enforcement, border 
control, immigration or asylum 
authorities. 
Art. 73 
 
"(new [NR])  
 
Before adopting a delegated act, the Commission 
shall consult experts designated by each Member 
State in accordance with the principles laid 
down in the Interinstitutional Agreement of [13 
April 2016] on Better Law-Making." 
Art. 82a 
 
Article 82 new 
The AI Act must allow representative 
Amendments to Directive 
actions to be used to defend natural 
person’s rights collectively. This should 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
2020/1828/EC on Representative Actions for the 
apply in the case of illegal commercial 
Protection of the Collective Interests of Consumers 
practices, or in obtaining compensation in 
case of harm suffered by a group of 
 
natural persons. Natural persons must be 
3.The following is added to Annex I of the Directive 
able via authorised organisations to 
jointly bring a court case to obtain 
2020/1828/EC on Representative Actions for the 
compensation for damages arising from 
Protection of the Collective Interests of Consumers: 
the same source (e.g. multiple consumers 
 
harmed by the same non-compliant AI 
“Regulation xxxx/xxxx of the European Parliament 
system or practice). In the absence of 
and of the Council laying down harmonised rules on 
adding the AI Act to the RAD Annex I, 
artificial intelligence (artificial intelligence act) and 
consumers would have no way of 
amending certain union legislative acts” 
exercising their rights collectively. 
Art. 83 
 
 
DEU further suggests to exclude large-
scale IT systems established by the legal 
acts listed in Annex IX from obligations 
of users of high-risk AI systems set forth 
in Art. 29 (in connection with Art. 12 and 
Art. 11) regardless of the date the 
systems have been placed on the market 
or put into service, since these systems 
are already regulated with regard to those 
obligations and the obligations laid down 
in the AI act may conflict with the 
obligation laid down in the existing 
legislation.  

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
If the amendment of those legal acts leads 
to a significant change in the design or 
intended purpose of the AI system, it then 
should be considered as a question of 
legal technique if any obligations of users 
of high-risk AI systems under the AI Act 
should be implemented directly within 
the legal acts listed in Annex IX itself.  
Furthermore, the suggested exemption is 
without prejudice to Art. 83 (2) of the 
Commission’s proposal, according to 
which the requirements laid down in this 
Regulation shall be taken into account, in 
the evaluation of each large-scale IT 
systems established by the legal acts 
listed in Annex IX to be undertaken as 
provided for in those respective acts. 
Please also refer to the separate position 
paper handed in, proposing necessary 
diverging regulations for public 
administration (especially LEAs) 
„[Regulation of AI – taking greater 
account of the specific characteristics of 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
the public administration, particularly in 
the fields of security and migration]“. 
Art. 83(3) 
Without prejudice to paragraphs 1 and 
Deletion 
We ask for the deletion of the new Art. 
2, high-risk AI systems that have been 
placed on the market before [date of 
83(3), which goes much too far and 
application of this Regulation referred to 
changes the approach of the AI act (as 
in Article 85(2)], may continue to be 
made available on the market until 3 
regulation regarding the future). We need 
years after [date of application of this 
a sufficient grandfathering rule for high-
Regulation referred to in Article 85(2)]. 
risk AI systems that came onto the 
market before the regulation came into 
force. It cannot be that systems in 
operation have to be taken off the market 
after 3 years and can not be made 
available after a certain target date. For 
example, the acquisition of further 
licenses for AI systems which are already 
in use must remain possible on a 
permanent basis. It is highly unlikely, 
that the implication of the requirements 
of the AI act can occure retroactively for 
AI systems that are already developed 
and on the market. However, new 
development of certain AI systems can 

---

 
Reference 
Third compromise proposal 
Drafting suggestion 
Comment 
take much longer then 3 or maybe even 
10 years, if at all possible. The 
implications of this change to Art. 83 are 
huge and may lead to problems for entire 
sectors.  
We will timely submit concrete examples 
to further elaborate on the problems we 
assume. 
 
 
 
 
 
 
 
Kindly indicate the Member State you are representing in the Title and when renaming the document. For specifying the relevant provision, please 
indicate the relevant Article or Recital in 1st column and copy the relevant sentence or sentences as they are in the current version of the text in 2nd 
column. For drafting suggestions, please copy the relevant sentence or sentences from a given paragraph or point into the 3rd column and add or 
remove text. Please do not use track changes, but highlight your additions in yellow or use strikethrough to indicate deletions. You do not need to 
copy entire paragraphs or points to indicate your changes, copying and modifying the relevant sentences is sufficient. For providing an explanation 
and reasoning behind your proposal, please take use of 4th column.  

---

Document Outline

• Coverpage.pdf
• DE Comments on 2nd part of 3rd compromise on AIA.docx

---