AI Model Cards: State of the Art and Path to Automated Use

Ali Mehraj

1 a

, An Cao

1 b

, Kari Syst

1 c

, Tommi Mikkonen

2 d

, Pyry Kotilainen

2 e

David H

astbacka

1 f

and Niko M

akitalo

2 g

Faculty of Information Technology and Communication Sciences, Tampere University, Tampere, Finland

Faculty of Information Technology, University of Jyv

askyl

a, Jyv

askyl

a, Finland

ﬁ ﬁ

Keywords:

Model Cards, AI Model, Privacy, Transparency, Ethical AI, Machine Learning, ML, Artiﬁcial Intelligence, AI.

Abstract:

In software engineering, the integration of machine learning (ML) and artiﬁcial intelligence (AI) components

into modern web services has become commonplace. To comply with evolving regulations, such as the EU AI

Act, the development of AI models must adhere to the principles of transparency. This includes the training

data used, the intended use, potential biases, and the risks associated with these models. To support these

goals, documents named Model Cards were introduced to standardize ethical reporting and allow stakeholders

to evaluate models based on various goals. In our ongoing research, we aim to automate risk analysis and

regulatory compliance checks in software systems. We envision that model cards can serve as useful tools to

achieve the goal. Given the evolving format of model cards over time, we conducted a state-of-the-art review

of the current state and practice of model cards by analyzing 90 model cards from four model repositories

to assess their relevance to our vision. The study’s contribution is a thorough analysis of the model cards’

structure and content, as well as their ethical reporting. Our study reveals the variance in information reporting,

the loose structure, and the lack of ethical reporting in the model cards. Based on the ﬁndings, we propose a

uniﬁed model card template that aims to enhance the structure, promote greater transparency, and establish a

foundation for future machine-interpretable AI model cards.

1 INTRODUCTION

Rapid advancement of Artiﬁcial Intelligence (AI) has

signiﬁcantly impacted the engineering of Internet and

Web systems. Machine Learning (ML) models, espe-

cially Large Language Models (LLMs), have demon-

strated remarkable capabilities in a variety of appli-

cations such as text generation, text translation, text

comprehension, and image generation. Such models

have a vast number of parameters and are trained with

large datasets, which can introduce privacy and se-

curity concerns and may lead to privacy and security

breaches of personal data. With emerging regulations

such as the EU AI Act (Edwards, 2021), the Data Act

(Perarnaud et al., 2022) and the existing General Data

Protection Regulation (GDPR) (Voigt and Von dem

https://orcid.org/0009-0006-7563-1706

https://orcid.org/0009-0005-6460-6567

https://orcid.org/0000-0001-7371-0773

https://orcid.org/0000-0002-8540-9918

https://orcid.org/0000-0002-4645-074X

https://orcid.org/0000-0001-8442-1248

https://orcid.org/0000-0002-7994-3700

Bussche, 2017), these AI models will need to adhere

to the guidelines set forth by the regulations. It is cru-

cial for developers and organizations involved in de-

veloping these AI models to be transparent regarding

the training data, intended use, potential biases, and

the risks associated with their models.

Model Cards, introduced by Google in 2019

(Mitchell et al., 2019) are documents that represent

ML models to standardize ethical reporting. These

documents allow stakeholders to evaluate models

based on inclusiveness, fairness, ethics, and tradi-

tional metrics reported in the model cards (Mitchell

et al., 2019). While model cards are becoming an in-

tegral part of model documentation, their format is

continuously evolving (Kotilainen et al., 2024). We

have utilized the concept of these model cards in our

previous research to ensure safe and ethical orchestra-

tion, and deployment of ML models using metadata

cards (Kotilainen et al., 2025).

In many systems, ML is part of a larger system.

Therefore, to identify potential risks in a system, it is

important to analyze the impact associated with them

to the overall system. In addition, the risks depend on

Mehraj, A., Cao, A., Systä, K., Mikkonen, T., Kotilainen, P., Hästbacka, D. and Mäkitalo, N.

AI Model Cards: State of the Art and Path to Automated Use.

DOI: 10.5220/0013706600003985

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 21st International Conference on Web Information Systems and Technologies (WEBIST 2025), pages 71-82

ISBN: 978-989-758-772-6; ISSN: 2184-3252

the data used by the system and who has access to the

overall system. Therefore, when analyzing the entire

system where ML models are deployed, it is essen-

tial to utilize the information associated with the ML

models provided by developers in the form of model

cards. An example of such an approach has been de-

scribed in our earlier work (Kotilainen et al., 2025).

In this paper, we analyze the structure and con-

tent of different model cards for their applicability as

building blocks for our future work on the automa-

tion of some steps in software architecture design and

analysis. With all software development tasks moving

towards automation, we aim to integrate automated

regulatory compliance into the software development

process and ensure that software systems are safe and

regulatory compliant. In addition, we want to en-

sure that ML models deployed in software systems

are regulatory compliant and do not cause unintended

side effects. We believe that model-related informa-

tion provided by the developers as model cards can be

useful tools for designing and developing regulatory

compliant software and enable the safe deployment of

ML models in software systems. To validate this plan,

we assess the current state-of-the-art in the implemen-

tation of model cards in practice. This will facilitate

a more comprehensive understanding of the model

cards’ content and their relevance in designing and

developing regulatory-compliant software. This state-

of-the-art review provides insight into how different

developers utilize the concept of model cards and the

differences in the model card structure. The review

also investigates the applicability of model cards for

realizing our vision of automating regulatory compli-

ance in software systems.

The structure of the paper is as follows. In Sec-

tion 2, we brieﬂy describe the background and mo-

tivation for our research, mainly discussing motivat-

ing factors and existing research incorporating model

cards. In Section 3, we present the methodology and

the process for selecting and analyzing model cards

for the research. In Section 4, we state the ﬁndings

from our analysis of the selected model cards. In Sec-

tion 5, we discuss and analyze the results. We also

establish a uniﬁed model card template to incorporate

our regulation-aware software architecture vision. Fi-

nally, in Section 6 we present some ﬁnal conclusions.

2 BACKGROUND AND

MOTIVATION

Model cards offer a systematic approach to documen-

tation of essential aspects and factors of ML mod-

els, facilitating comprehension, assessment, and repli-

cation of results for researchers (Dan and Seeuws,

2024). Model Cards were introduced to improve

transparency and responsible use of AI. It is achieved

by outlining metrics such as accuracy in different

datasets, potential biases, and possible ethical prob-

lems (Mitchell et al., 2019). The goal of the model

cards is to increase transparency by providing detailed

information about the intended use, performance met-

rics, identiﬁed limitations, and other considerations

taken into account by the developers (Nunes et al.,

2022). Model cards are the ﬁrst step in standardizing

ethical practice and reporting for AI models and al-

lows users or stakeholders to compare models before

deployment for their speciﬁc use case, thereby facili-

tating decision-making, reducing expenses, and help-

ing users identify the right product in the AI market

(Wadhwani and Jain, 2020). Model cards have gained

signiﬁcant traction in the industry, particularly among

major organizations engaged in developing ML mod-

els, such as Google, Meta, Nvidia, and OpenAI.

The model cards presented in (Mitchell et al.,

2019) have 9 sections. The sections and the content

of the sections of a model card are listed in Table 1.

Among the sections, the Ethical Considerations sec-

tion is the primary source of information for our on-

going research on identifying the potential risks posed

by ML models in software systems and potential reg-

ulatory compliance.

In practice, model cards have gained widespread

acceptance. Hugging Face, a prominent repository for

model cards, currently houses over 1.6 million models

and their corresponding model cards. Organizations

involved in the development of AI/ML models such

as, OpenAI, Nvidia, Google, Anthropic, Meta have

also adopted the concept and provide model cards for

their corresponding ML models. Amazon’s AI Ser-

vice Cards

bear similarity to the model cards and ap-

pear to draw inspiration from the concept of the model

cards. In addition, there are templates and toolkits for

creating model cards provided by Google

, Tensor-

Flow

, PyPi

, and Hugging Face

The concept of model cards has been used in sev-

eral domains. A study in the medical ﬁeld proposed a

Checklist for Artiﬁcial Intelligence in Medical Imag-

ing (CLAIM) (Mongan et al., 2020). The checklist

https://aws.amazon.com/blogs/machine-

learning/introducing-aws-ai-service-cards-a-new-resource-

to-enhance-transparency-and-advance-responsible-ai/

https://research.google/blog/introducing-the-model-

card-toolkit-for-easier-model-transparency-reporting/

https://github.com/tensorﬂow/model-card-toolkit

https://pypi.org/project/model-card-toolkit/

https://huggingface.co/docs/hub/model-card-

annotated

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

Table 1: Sections of a Model Card (Mitchell et al., 2019).

Model Card sec-

tion

Section content

Model Details Basic information regarding the model version, type and other details

Intended Use Information regarding the intended use of this model, as well as its intended users

Factors Relevant factors related to the performance of the model

Metrics The metrics for the performance of the model

Evaluation Data Information regarding evaluation datasets, selection rationale, and pre-processing steps

of the data prior to model assessment

Training Data Information related to model training datasets and possible training process.

Quantitative

Analyses

The results of the evaluation of the model according to the selected metrics

Ethical Consider-

ations

Information regarding privacy, safety, and risks associated with the model. Identiﬁes

potential ethical challenges and the solutions that have been proposed for stakeholders

Caveats and Rec-

ommendations

Information regarding additional concerns that were not addressed in the preceding

sections

corresponds to the information required in a model

card. In our previous research, we have success-

fully employed the model card concept, describing

safe and ethical orchestration and deployment of ML

models using metadata cards (Kotilainen et al., 2025).

Previous research has been conducted on incorporat-

ing empirically derived sensor fusion recurrent neural

network (RNN) performance and cost measurement

data into machine-readable model cards (Booth and

Ghosh, 2023). One study employs extensive and stan-

dardized terminology and scientiﬁc rigor as promoted

by biomedical oncologists and establishes a method

to make model cards machine-readable using seman-

tic web technology (Amith et al., 2022).

As demonstrated by the examples above, model

cards offer a variety of uses and there are attempts

to make model cards machine interpretable. How-

ever, we have not found any work on making model

cards useful for tools and methods used for system-

level design. Our long-term objective is to incorpo-

rate machine-interpretable model cards into the sys-

tem design and analysis process. To achieve this, we

must enhance our understanding of the content and

structure of existing model cards and observe the evo-

lution of model cards in research and practice. In

addition, it is essential to assess the applicability of

these model cards in achieving our long-term objec-

tives. Although our primary focus is on ethical and

regulatory compliant software architecture design and

risk analysis, we believe that the insights gained will

be beneﬁcial to individuals with different interests re-

lated to model cards.

3 RESEARCH APPROACH

In this section, we describe the research questions,

model card search strategy, and data extraction from

the selected model cards.

3.1 Research Questions

In consideration of the research objectives, three re-

search questions (RQs) were formulated.

RQ1. What kind of information is provided

in different model cards? Our primary objective

is to understand to which extent the information in

the current model cards serves the needs of the envi-

sioned regulatory-compliant architecture design and

risk analysis. This includes checking model cards for

ethical considerations information.

RQ2. What are the key similarities and differ-

ences in the structure of the model cards? Diverse

structure requires more adaption and conversion, but

can be solved with proper tooling. Discrepancies in

critical sections of the content could compromise the

viability of analyzing multiple systems with similar

tools and approaches. Diverse structures add com-

plexity to automated extraction of information from

model cards and future machine-interpretability.

RQ3. How well are the ethical and regulatory

aspects covered in the model cards? As we aim to

assess risk and regulatory compliance in software sys-

tems, it is essential to evaluate the extent to which

these aspects are reported in the model cards to en-

sure their suitability for our purposes.

AI Model Cards: State of the Art and Path to Automated Use

3.2 Model Card Search

The search for model cards was carried out

through publicly accessible model repositories be-

tween 04.04.2025 and 10.04.2025. The selection cri-

terion for a model repository was a minimum of 100

models hosted. We found four ML model reposi-

tories that hosted a large collection of ML models

and had dedicated model cards for all of their mod-

els. These model repositories are Hugging Face

Kaggle

, Nvidia NGC Catalog

, and Google Model

Garden

. In contrast, our search results revealed that

other model repositories, with the exception of these

four, had very few models hosted, and the models

were not accompanied by a model card.

Our ﬁrst search was conducted in the largest pub-

licly available model repository, Hugging Face. At

the time of conducting the research, Hugging Face

hosted over 1.6 million ML models and their respec-

tive model cards. The second search was conducted

on another large model repository called Kaggle. At

the time of conducting the research, Kaggle hosted

more than 19,000 models. The third search was con-

ducted using Nvidia’s NGC Catalog, which at the

time of the research contained more than 800 models.

The ﬁnal search was conducted in Google’s Model

Garden. Google’s Model Garden showcases over 100

models developed by the company.

The selection of model cards from each model

repository varies. This is due to the number of mod-

els hosted by each repository. As Hugging Face is the

largest model repository in our selection, we decided

to collect the majority of our sample of model cards

from Hugging Face and selected 50 model cards.

Kaggle was the second largest model repository in

our selection. Given the signiﬁcantly lower number of

model cards in Kaggle compared to Hugging Face, we

decided to select 20 model cards from Kaggle. The

third and fourth largest model repository in our search

were Nvidia’s NGC Catalog and Google Model Gar-

den respectively. We selected 10 model cards from

each of these repositories. In total, 90 model cards

were selected from the repositories.

In order to investigate the state of the art of model

cards, we have selected the top downloaded, liked,

popular, or trending model cards. This selection was

based on the sorting mechanisms available in the

repositories. As model sorting was available in Hug-

ging Face for both the top downloaded and top liked

https://huggingface.co/models

https://www.kaggle.com/models

https://catalog.ngc.nvidia.com/models

https://console.cloud.google.com/vertex-ai/model-

garden

models, we decided to select the top 30 downloaded

and the top 20 liked models’ cards from the reposi-

tory. The sorting function in Kaggle was limited to

top downloads. Therefore, we selected the top 20

downloaded model cards from Kaggle. The only sort-

ing options in NGC Catalog and Google Model Gar-

den were most popular and trending model cards, re-

spectively. Therefore, we have selected top ten popu-

lar and top ten trending model cards from NGC Cat-

alog and Google’s Model Garden, respectively. The

selected model cards are listed in Table 2.

3.3 Data Extraction

Data extraction was conducted by reading the full

content of each selected model card. We collected

statistical data for the model cards, including num-

ber of sections in each model card and model card

sources. In addition to collecting statistical data, we

collected data related to our research questions. Re-

garding RQ1, the different types of information pro-

vided in each model card were collected. For RQ2,

we collected information related to the differences

and similarities in the structure of each model card.

For RQ3, we collected ethical and regulatory data re-

ported in the selected model cards.

4 RESULTS

4.1 RQ1. What Information Is Provided

in Different Model Cards?

The different types of information reported in the

model cards are illustrated in Figure 1. As illustrated

in Figure 1, essential information, with the excep-

tion of the ethical considerations of the original model

card concept outlined by (Mitchell et al., 2019) is rep-

resented in blue. Information related to different eth-

ical considerations discovered in our selected model

cards is represented in red. It can be observed from

Figure 1 that 85 of the 90 selected model cards pro-

vide basic details about the model. This includes a

detailed description of the model as well as the type

of the model. This information is further supported

by speciﬁcs about the model architecture in some of

the model cards. The remaining ﬁve models lacking

model details are primarily empty model cards. These

ﬁve model cards were obtained from Hugging Face.

Information related to the intended use of the

model was present in 69 model cards in various de-

grees. There was also variability in the information.

There was more emphasis on how to use the model

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

Table 2: Selected model cards from different model repositories.

Hugging Face

mobilenetv3 smal-

l 100.lamb in1k

adetailer distilbert-base-

uncased

DeepSeek-R1 whisper-large-v3

xlm-roberta-large electra-base-

discriminator

speaker-diarization-

3.1

black-forest-labs Kokoro-82M

nsfw image dete-

ction

esmfold v1 wav2vec2-large-xlsr-

53-chinese-zh-cn

stable-diffusion-v1-4 stable-diffusion-

2-1

all-MiniLM-L6-

resnet50.a1 in1k roberta-base stable-diffusion-xl-

base-1.0

Meta-Llama-3-

8B-Instruct

fairface age ima-

ge detection

roberta-large xlm-roberta-base Meta-Llama-3-8B Llama-3.1-8B-

Instruct

bert-base-

uncased

wespeaker-

voxceleb-

resnet34-LM

vit-base-patch16-

224-in21k

bloom OrangeMixs

chronos-t5-small clip-vit-base-

patch32

1 stable-diffusion-3-

medium

DeepSeek-V3

clip-vit-large-

patch14

vit-face-

expression

wav2vec2-large-xlsr-

53-japanese

Mixtral-8x7B-

Instruct-v0.1

ControlNet-v1-1

all-mpnet-base-

gpt2 paraphrase-

multilingual-

MiniLM-L12-v2

Llama-2-7b-chat-hf Mistral-7B-v0.1

phi-2-GGUF segmentation-3.0 wav2vec2-large-xlsr-

53-portuguese

Llama-2-7b ControlNet

Kaggle Nvidia NGC Catalog Google Model Garden

blazeface toxicity TrafﬁcCamNet Gemma 2

face detection handskeleton BodyPoseNet Gemma 3

iris handdetector DashCamNet TxGemma

mobilenet v2 handpose 3d VehicleTypeNet Gemini 2.5 Flash Preview

facemesh blazepose 3d VehicleMakeNet Gemini 2.5 Pro Preview

universal-sentence-

encoder

mobilenet v3 FaceDetectir Gemini 2.0 Flash Thinking

Mode

universal-sentence-

encoder

deeplab FaceDetect Gemini 1.5 Flash

movenet bert License Plate Recognition

(LPRNet)

Gemini 1.0 Pro Vision

mobilenet v1 selﬁe segmentation Facial Landmark Estimator

(FPENet)

Gemini 2.0 Flash-Lite

face landmarks -

detection

yamnet EmotionNet Gemini 2.0 Flash

rather than on the actual intended use cases of the

model. Other than the intended use, 15 model cards

reported misuses of their corresponding models.

We observed the lack of information related to

training data reported in the model cards as only 39

model cards documented training data-related infor-

mation. In particular, NGC Catalog model cards

documented detailed descriptions of the training data

along with the training process of the models. Sim-

ilarly, factors and metrics-related data were largely

missing from the model cards, reported only in 18

model cards. ML model evaluation and performance-

related data were present in 36 model cards.

The different ethical considerations information

we discovered from the model cards are related to

bias, privacy, safety, security, risk, misuse, and lim-

itations. Bias-related data primarily provided infor-

mation related to potential biases in the model’s out-

put and attempts to mitigate the bias, if any, and was

present in 23 model cards. Privacy-related informa-

tion in the model cards primarily discuss the measures

taken to protect privacy of the trained data. This in-

formation was followed up with privacy policies in

the model cards developed by Google and Nvidia.

AI Model Cards: State of the Art and Path to Automated Use

Figure 1: Information reported in different model cards.

Figure 2: Variability of information in model cards grouped by model repositories.

Privacy-related information was reported in 8 model

cards. Safety-related information primarily discuss

safety policies and the safety of the generated output

of the models. Safety-related information was bun-

dled with security in the model cards from Nvidia as

Safety and Security. Safety and security-related infor-

mation was reported in 13 and 7 model cards respec-

tively. Risk-related information primarily consisted

of identiﬁed and potential risks with harms, and the

possible measures taken to mitigate them, and was re-

ported in 13 model cards. The limitations of the mod-

els were explicitly reported in 38 model cards.

The variance of the information of the model cards

in different model repositories is illustrated as a heat

map in Figure 2. As illustrated in Figure 2, the model

cards from Hugging Face generally include model

details, intended use, evaluation data, training data,

quantitative analysis, and limitations-related informa-

tion. For the model cards in Kaggle, only the model

details and the information related to its intended use

are primarily available. All other types of informa-

tion are mostly missing. Nvidia’s NGC Catalog in-

cludes more comprehensive model cards that thor-

oughly report essential information, along with ad-

ditional ethical considerations. These considerations

include information related to privacy, safety, secu-

rity, and risks. Finally, for the Google Model Garden,

the only consistent information on the model cards

was the details of the model. Five of the model cards

from the ten selected Google model cards had their

intended use documented. It is noteworthy that al-

though Google pioneered the model card concept and

template, their own 7 model cards from the Gem-

ini family appear to deviate from their recommended

template and provide less information compared to

their 3 model cards from the Gemma family.

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

4.2 RQ2. What Are the Key Similarities

and Differences in the Structure?

Model cards from different repositories have similar-

ities and differences in how they are documented in

the repositories. While Nvidia’s NGC catalog and

Google’s Model Garden repositories host internally

developed models and some partner models, Hugging

Face and Kaggle are open to the public. Developers

can upload their own ML models and are also respon-

sible for maintaining and updating the model cards

for their respective models. A dedicated page for

model cards is available for Hugging Face and Kag-

gle, which is the ﬁrst page that users will see when

browsing a model in the repository. For Nvidia’s

NGC Catalog, the model card is divided into multiple

pages rather than being contained within a single doc-

ument. For instance, the TrafﬁcCamNet model card’s

bias, explainability, privacy, safety, and security are

reported under the Model Card++ page, while rest of

the information is reported under the Overview page.

For Google Model Garden, the model card informa-

tion is provided under a single page.

Furthermore, we have observed variations in the

titles of different sections and the presentation of re-

lated information in different model cards and listed

the information in Table 3. The ﬁrst column of the

table represents the original model card sections and

the new sections that were identiﬁed from the model

cards for RQ1. The Same section name column indi-

cates the number of model cards that share the same

section name. The Similar section name column rep-

resents the number of model cards that have a sim-

ilar section name or header. For example, instead

of Model Details, the information is reported under

Model description section. The section name is con-

sidered similar if it matches partially with the sec-

tion names in the ﬁrst column. Otherwise, the sec-

tion name is considered different and placed under

the fourth column Other section name. If the model

card did not have a structure and the information was

reported as free form text, then it is also considered

under the Other section name column. The ﬁfth col-

umn represents the different section names where the

corresponding information was reported other that the

original section name.

It can be observed from Table 3 that with the

exception of the Factors, Metrics, Evaluation Data,

and Quantitative Analyses sections, the majority of

the other sections either had the same or a similar

name to represent corresponding information in the

model cards. We discovered that the metrics, evalu-

ation data, and quantitative analyses related informa-

tion was reported together as Evaluation or Evalua-

tion Results or Performance in different model cards.

Hardware and software factors were only reported in

the model cards without specifying the Factors sec-

tion header. The section name Quantitative Analyses

was completely missing from all the model cards, re-

placing it with Evaluation Results, Benchmark, and

Performance in different model cards.

In addition, we observed the different ways in

which information related to ethical considerations

was presented in the model cards. Only 13 model

cards reported information related to ethical consid-

erations in a dedicated section or one with a similar

name. However, majority of the model cards that re-

ported these concerns split information related to eth-

ical considerations, such as privacy, safety, security,

risks, and misuses, into dedicated sections or sub-

sections. We believe that the developers wanted to

put more emphasis on the potential ethical concerns,

which is why they decided to have a dedicated section

for this information in the model cards.

4.3 RQ3. How Well Are the Ethical and

Regulatory Aspects Covered?

Ethical and regulatory aspects in model cards are

listed in Table 4. The ethical and regulatory data has

been collected from designated sections and subsec-

tions from the model cards. Risk factors are con-

sidered to be any type of risk posed by the model.

This includes potential risks, severity of the risks, and

misuses of the model. While risk factors were re-

ported in 34 model cards, mitigation of these risks

were only reported in 4 model cards. Similarly, poten-

tial bias was reported in 22 model cards, while only

ﬁve model cards reported measures to mitigate the re-

ported bias. Limitations were reported in 38 model

cards, which highlighted the known limitations of the

model. Privacy concerns and measures taken to ad-

dress the concerns were reported in 11 model cards.

These privacy concerns report information related to

safeguarding of data and identity. 9 model cards fol-

lowed up with a privacy policy stating how the devel-

oping organization collects, uses, shares, and protects

personal data. Safety measures were reported in 13

model cards. These safety measures primarily report

implementation of safety features to reduce the sever-

ity of potential risks. Security measures were reported

in seven model cards, along with the safety measures.

These measures were intended to report both security

and safety measures, in an effort to reduce risk sever-

ity. Safety or security policies were not reported in

any of the selected model cards.

Explicit regulatory compliance was not reported

in any of the selected model cards. However, six

AI Model Cards: State of the Art and Path to Automated Use

Table 3: Section name variation in different model cards.

Section name Same

section

name

Similar

section

name

Other

section

name

Section names (if not the same section name)

Model details 38 26 21 Model description (12), Model Overview (10), Overview (6),

Model Summary (2), Model Facts (1), Description / About /

Model summary (1), Free form text (15)

Intended use 17 49 6 Example use (18), Usage (11), Intended uses & limitations (9),

Use cases (4), Uses (4), How to use this model (4), Intended

Usage (2), How to use (1), Free form text (2)

Factors 1 0 17 Hardware and Software (6), Software Integration (6), How to

use this model (4), Implementation Information (1)

Metrics 3 8 6 Performance Metrics (6), Performance (4), Evaluation Metrics

(2), Evaluation (2)

Evaluation

data

5 7 22 Evaluation results (7), Performance (6), Evaluation (5), Pre-

training (4), Benchmarks (3), Evaluation Approach (2), Eval-

uation dataset (1), Benchmark (1)

Training Data 16 20 3 Training (12), Training Dataset (2), Data (2), Dataset and

training (1), Training Details (1), Free form text (1)

Quantitative

Analyses

0 29 9 Evaluation results (13), Performance (12), Test result (3),

Benchmarks (3), Evaluation (3), Model Comparison (1),

Benchmark (1), Free form text (2)

Ethical Con-

siderations

6 7 0 Ethical Considerations and Limitations (5), Ethical Consider-

ations and Risks (2)

Caveats and

Recommen-

dations

1 0 1 Broader Implications (1)

Limitations 17 21 0 Technical Limitations (6), Ethical Considerations and Limita-

tions (5), Intended uses & limitations (4), Limitations and bias

(4), Performance and Limitations (1), Risks and Limitations

(1)

Bias (and/or

Fairness)

11 6 6 Limitations and bias (5), Ethical Considerations (4), Ethical

Considerations and Risks (2), Unintended bias evaluation data

(1)

Privacy 6 0 2 Ethical Considerations and Risks (2)

Safety 1 11 0 Safety and security (6), Responsibility & Safety (3), Ethics

and safety (2)

Security 1 6 0 Safety and security (6)

Risk 0 13 0 Potential Known Risks (6), Critical risks (3), Ethical Consid-

erations and Risks (2), Risks and Limitations (1), Risks iden-

tiﬁed and mitigations (1)

Misuse 3 4 9 Safety and security (6), Misuse, Malicious Use, and Out-of-

Scope Use (3), Ethical Considerations and Risks (2), Misuse

and Out-of-scope Use (1), Out-of-Scope Use (1)

Nvidia model cards demonstrate adherence to some

aspects of GDPR. These models explicitly provide in-

formation related to data compliance, correction, and

removal. These model cards state compliance with

privacy laws related to data labeling. However, they

do not explicitly declare compliance to GDPR.

Our observations indicate a lack of ethical report-

ing in the model cards that we reviewed. Risk fac-

tors and limitations were reported in less than half of

the selected model cards. In addition, privacy, safety,

and security aspects were largely missing from the

model cards. Privacy policies were reported in some

model cards from Meta, Google, and Nvidia, indi-

cating that major organizations involved in AI/ML

model development tend to provide transparency re-

lated to data privacy regarding their models. Regula-

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

Table 4: Ethical and regulatory aspects in models cards.

Ethical and regula-

tory aspect

Description Number

of model

cards

Risk factors Factors related to any kind of risk reported in the model card including

miuses

Risk mitigation Measure taken to mitigate the risks reported in the model cards 4

Bias and/or fairness Potential bias and/or fairness reported in the model cards 22

Bias mitigation Measure taken to mitigate the bias reported in the model cards 5

Privacy concerns and

measures

Privacy concerns and measures taken to address the privacy concerns 11

Safety and/or secu-

rity measures

Safety and/or security concerns and measures taken to address the safety

and/or security concerns

Safety and/or secu-

rity policy

Privacy policies of the developing organization 0

Limitations Known limitations of the model reported in the model cards 38

Adherence to regula-

tory requirements

Adherence to speciﬁc regulatory requirements reported in the model

cards

Regulatory compli-

ance declaration

Declaration of adherence to regulations reported in the model cards 0

tory compliance declarations were not reported in any

of the model cards. Given the implementation of the

GDPR and EU AI Act, it is unexpected that there has

been minimal reporting on regulatory compliance for

the models in their corresponding model cards.

5 RESULTS ANALYSIS AND

MODEL CARD SYNTHESIS

The adoption of model cards by various organiza-

tions demonstrates considerable potential in promot-

ing transparency and ethical reporting for stakehold-

ers. This study explores the similarities, differences,

and evolution of these model cards over the years. We

analyzed different model cards from different model

repositories and discovered key similarities and dif-

ferences across multiple model cards. It is evident that

most model card developers include information on

model details, intended use and training data related

to the models. While the speciﬁc details of training

data vary between different model cards, there is suf-

ﬁcient information for the readers to comprehend the

nature of the model’s training dataset.

Furthermore, we observed similarities and differ-

ences in the structure of the model cards. The major-

ity of the model cards from the same organization or

model family demonstrate a high level of similarity in

terms of representation methods and overall design.

Furthermore, we have observed variations in the titles

of different sections and how the related information

is presented in different model cards. This observa-

tion is of particular importance to our research since

the different structural formats and section headings

of the model cards increase the complexity of auto-

matically extracting information from them.

We observed that the reporting of most model

cards was done in an informal way, and only Nvidia

and Google have made attempts to provide further in-

formation in their model cards. There is also varia-

tion in model cards from the same organization. For

example, the Gemma family model cards are more

complete than the Gemini family model cards. The

review of the model cards indicates that the model

cards generally have a loose structure. Some devel-

opers, such as Nvidia and Google, have created their

own templates, while others have opted not to follow

any structure and instead organize relevant informa-

tion in the model card according to their preferences.

To facilitate risk analysis and develop regulatory

compliant software, we need information regarding

the ML model’s ethical considerations, such as pri-

vacy, security, risks, misuse, and bias. In some cases,

the information may not be explicitly stated or the re-

lated information may be available under other sec-

tions or subsections. This adds further complexity

to the process of extracting speciﬁc information from

different model cards. Furthermore, model cards with

bare minimum data such as the model cards from

Kaggle are not suitable for risk analysis and regula-

tory check as these models do not provide enough

data about the model to ensure risk-related informa-

AI Model Cards: State of the Art and Path to Automated Use

tion and compliance to certain regulations.

We discovered that model cards could be a suit-

able solution for our research, provided that the rele-

vant ethical considerations are reported in the model

cards. In addition, model cards containing risk-

related information could be useful tools for facilitat-

ing risk analysis and regulatory compliance checks.

However, it is noteworthy that the majority of the

model cards do not currently provide the required

risks and regulation-related information for our work.

The potential risks associated with some of the mod-

els remain uncertain, as the developers have not pro-

vided regulatory compliance information in their re-

spective model cards. Therefore, it is essential that

developers clearly state compliance-related informa-

tion in the model cards to provide transparency. The

implementation of the EU AI Act is expected to en-

courage developers to share regulatory compliance

data in their model cards, enhancing transparency and

ensuring compliance with the regulations. In addition,

the structure and data of model cards vary widely,

making it difﬁcult to automatically extract informa-

tion from a document that does not have a consistent

structure for section titles and text content.

To facilitate our future work for automating in-

formation extraction from model cards, we recom-

mend an updated model card template developed from

the ﬁndings of our data collection and analysis. Our

proposed model card template is similar to the origi-

nal model card introduced by (Mitchell et al., 2019).

However, based on our ﬁndings from analyzing the

content and structure of the model cards in this review

we propose several changes and added new sections

and subsections to our proposed template. The tem-

plate is represented in Table 5. We have also taken in

the account the requirements of Annex IV

of the EU

AI Act (Edwards, 2021) related to AI transparency

obligations in our proposed model card template. Al-

though the requirements in Annex IV are intended for

AI systems as a whole, we have addressed those ap-

plicable to AI models. Table 6 represents the com-

parison of our proposed model card template with the

EU AI Act Annex IV transparency requirements.

Starting from the Model Details section, instead of

free form text, we propose subsections for reporting

model provider, model version, model type and pur-

pose, model architecture, model description, and li-

cense. We added two additional subsections to the In-

tended Use section called Primary Intended Use and

Secondary Intended Use. These subsections have fur-

ther subsections that report the intended users, the us-

age domain, and the use cases. The usage domain and

https://eur-lex.europa.eu/legal-

content/EN/TXT/PDF/?uri=OJ:L 202401689#page=130

Table 5: Proposed Model Card Template.

Model Details

- Name of the provider

- Model version

- Model type

- Model architecture

- Model description

- License

Intended Use

Primary Intended Use

- Primary intended users

- Domain

- Use case(s)

Secondary Intended Use

- Secondary intended users

- Domain

- Use case(s)

Factors

- Hardware

- Software

Evaluation

- Evaluation Metrics

- Evaluation Factors

- Evaluation Datasets

- Evaluation Results / Performance

Training Data

- Training Algorithm

- Training Dataset

- Training Procedure

Ethical Considerations

Bias

- Type of bias

- Bias mitigation

Privacy

- Privacy concerns

- Privacy policy

Safety and security

- Safety measures

- Security measures

- Safety and security policy

Risk

- Risk factors

- Risk severity

- Risk mitigation

Misuse

- Unintended user(s)

- Prohibited domain(s)

- Misuse case(s)

Regulatory compliance

- Adherence to regulatory requirements

- Regulatory compliance declaration

Limitations

Caveats and Recommendations

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

Table 6: Comparison of proposed model card template with EU AI Act Annex IV transparency requirements.

EU AI Act Annex IV transparency requirements Requirements covered by the pro-

posed Model Card template

Intended purpose, the name of the provider and the version of the

system reﬂecting its relation to previous versions.

Model details and Intended Use

How the AI system interacts with, or can be used to interact with,

hardware or software, including with other AI systems, that are not

part of the AI system itself, where applicable. Description of the

hardware on which the AI system is intended to run.

Factors (Hardware and Software)

The data requirements in terms of datasheets describing the training

methodologies and techniques and the training data sets used

Training Data (Training Algorithm,

Training Dataset, Training Procedure)

Instructions for use for the deployer, and a basic description of the

user-interface provided to the deployer, where applicable.

Intended Use (Use cases)

Information about the validation and testing data used and their main

characteristics; metrics used to measure accuracy, robustness and

compliance with other relevant requirements. A description of the

appropriateness of the performance metrics for the speciﬁc AI system

Evaluation (Evaluation Factors, Eval-

uation Dataset, Evaluation metrics,

Evaluation Results / Performance)

Cybersecurity measures put in place Ethical considerations (Security mea-

sures)

A detailed description of the risk management system Ethical consideration (Risk)

A copy of the EU declaration of conformity Regulatory compliance (Regulatory

compliance declaration)

use cases are expected to provide further clariﬁcation

regarding the speciﬁc domain and use cases in which

the model should be implemented.

Since we only found relevant hardware and soft-

ware factors for the Factors section of the model

cards, two subsections, Hardware and Software, are

proposed under the Factors section. Changes are pro-

posed to Evaluation Data section as well by combin-

ing evaluation metrics, evaluation factors, evaluation

datasets, and evaluation results under the section. Ad-

ditional subsection in Training Data section named

Training procedure was added as we found that some

model developers reported the training procedures of

the ML models in their model cards.

Considering that almost half of the model cards

feature a designated limitations section, we have

added a dedicated Limitations subsection under Ethi-

cal Considerations section to our proposed template.

In addition, we added subsections for Ethical Con-

siderations section such as, Bias, Privacy, Safety, Se-

curity, Risk, and Misuse based on the different types

of ethical considerations data that we found from the

various model cards. The developers of the model

cards can state regulatory compliance data concern-

ing privacy, safety, and security in accordance with

the established regulations under these subsections.

We expanded the Bias subsection to include subsec-

tions that specify the type of bias and any available

mitigation strategies. Privacy subsection contains pri-

vacy concerns and policy-related information. We de-

cided to combine the safety and security aspects, as

most model cards with security-related information

include both aspects. The subsection Safety and Secu-

rity contains information related to safety and security

measures taken by the developers and also contains

the safety and security policy. The Risk subsection

provides information about probable risk factors, the

severity of these risks, and measures taken to mitigate

them. Similar to the Intended Use section, the Misuse

subsection report the unintended users, prohibited do-

mains, and potential misuse cases. Finally, we added

one subsection to the Ethical Considerations section

called Regulatory compliance declaration. Under this

subsection, the developers of the ML model can ex-

plicitly state compliance to certain regulations to help

the users have further clariﬁcation of the compliance

and adhere to compliance requirements when deploy-

ing the model in their software systems.

We believe that our proposed model card template

would provide further transparency of the model in

terms of risks and regulatory compliance. Model de-

velopers are expected to provide all the necessary in-

formation in the proposed template when document-

ing and publishing their models. We believe that the

EU AI Act will strongly enforce the requirement to

provide this information in the near future. We be-

lieve that a well-designed model card can facilitate

automated information extraction and enhance its us-

ability for automated regulatory checks in software

systems in the future. While our proposed template is

not yet machine-interpretable and would require fur-

ther reﬁnement and proper tooling to automate the ex-

AI Model Cards: State of the Art and Path to Automated Use

traction of information, a well-structured model card

is undoubtedly the ﬁrst step in this direction.

THREATS TO VALIDITY

To ensure transparency in our study selection and data

extraction, we shared all of our data and ﬁndings in

a spreadsheet

that we used throughout the research

process. Despite our best efforts to cover all major

model repositories in our research, it is possible that

some repositories might have been overlooked. The

decision to select the number of model cards from

each repository was justiﬁed by the number of models

hosted by the corresponding model repository, as dis-

cussed in Section 3.2. Unfortunately, all the selected

repositories did not have a common sorting mecha-

nism to select the same set of model cards for com-

parison. This may affect the overall ﬁndings.

6 CONCLUSIONS

In this paper, we reviewed a total of 90 model cards

to investigate the state-of-the-art in practice. The ﬁnd-

ings of this study demonstrate the evolution of model

cards and the applicability of model cards in the in-

dustry. The study also identiﬁes notable similarities

and differences in different model cards from differ-

ent model repositories. The differences in the con-

tent of different model cards among different organi-

zations are also highlighted in the study. In addition,

based on the results of the model cards examined in

this study, a new model card template is proposed.

The ﬁndings of the study are avenues of further

research. Future research can also address the short-

comings of this study by conducting an analysis of

the state of model cards with a larger pool of model

cards. In addition, the proposed model card template

is the initial foundation for our future research, which

aims to automate the extraction of risk and regulatory

information from model cards. Further research can

be conducted to determine the quality levels of these

model cards in terms of transparency and ethical re-

porting. We believe that the ﬁndings of the research

can contribute to future research related to safe de-

ployment of ML models in software systems, explain-

able AI, and assistance in regulatory compliance in

the ﬁeld of web and software engineering.

https://doi.org/10.6084/m9.ﬁgshare.29634044.v1

ACKNOWLEDGEMENTS

This work has been supported by Business Finland

(project LiquidAI, 6GSoft) and FAST, the Finnish

Software Engineering Doctoral Research Network.

REFERENCES

Amith, M. T., Cui, L., Zhi, D., Roberts, K., Jiang, X., Li,

F., Yu, E., and Tao, C. (2022). Toward a standard for-

mal semantic representation of the model card report.

BMC bioinformatics, 23(Suppl 6):281–281.

Booth, T. M. and Ghosh, S. (2023). Machine learning model

cards toward model-based system engineering analy-

sis of resource-limited systems. In Signal Process-

ing, Sensor/Information Fusion, and Target Recogni-

tion XXXII, volume 12547, pages 188–203. SPIE.

Dan, J. and Seeuws, N. (2024). Enhancing reproducibil-

ity of machine learning-based studies in clinical jour-

nals through model cards. Developmental Medicine &

Child Neurology, 66(2):144–145.

Edwards, L. (2021). The eu ai act: a summary of its sig-

niﬁcance and scope. Artiﬁcial Intelligence (the EU AI

Act), 1.

Kotilainen, P., Mehraj, A., Mikkonen, T., and M

akitalo, N.

(2024). The programmable world and its emerging

privacy nightmare. In International Conference on

Web Engineering, pages 255–262. Springer.

Kotilainen, P., M

akitalo, N., Syst

a, K., Mehraj, A., Waseem,

M., Mikkonen, T., and Murillo, J. M. (2025). Al-

locating distributed ai/ml applications to cloud–edge

continuum based on privacy, regulatory, and ethi-

cal constraints. Journal of Systems and Software,

222:112333.

Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman,

L., Hutchinson, B., Spitzer, E., Raji, I. D., and Gebru,

T. (2019). Model cards for model reporting. In Pro-

ceedings of the conference on fairness, accountability,

and transparency, pages 220–229.

Mongan, J., Moy, L., and Kahn Jr, C. E. (2020). Checklist

for artiﬁcial intelligence in medical imaging (claim):

a guide for authors and reviewers.

Nunes, J. L., Barbosa, G. D., de Souza, C. S., Lopes, H., and

Barbosa, S. D. (2022). Using model cards for ethical

reﬂection: a qualitative exploration. In Proceedings

of the 21st Brazilian Symposium on Human Factors in

Computing Systems, pages 1–11.

Perarnaud, C., Fanni, R., et al. (2022). The EU Data Act:

Towards a new European data revolution? Technical

report, Centre for European Policy Studies.

Voigt, P. and Von dem Bussche, A. (2017). The EU gen-

eral data protection regulation (gdpr). A Practical

Guide, 1st Ed., Cham: Springer International Pub-

lishing, 10(3152676):10–5555.

Wadhwani, A. and Jain, P. (2020). Machine learning model

cards transparency review : Using model card toolkit.

In 2020 IEEE Pune Section International Conference

(PuneCon), pages 133–137.

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies