GPAI Ledger › FIBO (Bria AI) › Capture 11 Aug 2026

FIBO — capture 20260811T103334Z

Provider	Bria AI
Target	AIAL archived copy — https://aial.ie/research/gpai-training-transparency/archive/FIBO_2026_08_03.pdf
Fetched (UTC)	2026-08-11T10:33:34Z
Stored file	99e1acc080c60af53f82397ad0d15774a8cf64148f1c5c0721987d1029b03ee9.pdf (299,517 bytes)
SHA-256	`99e1acc080c60af53f82397ad0d15774a8cf64148f1c5c0721987d1029b03ee9`
OpenTimestamps proof	99e1acc080c60af53f82397ad0d15774a8cf64148f1c5c0721987d1029b03ee9.pdf.ots (calendar-attested; anchored in bitcoin over time)
Wayback	not saved
Prior capture of this target	`— first capture of this target`
Notes	text_sha256 recorded 20 Aug 2026 from the extracted.txt stored at capture time (bootstrap captures predate this field); raw bytes unchanged

Verify: sha256sum 99e1acc080c60af53f82397ad0d15774a8cf64148f1c5c0721987d1029b03ee9.pdf must equal the hash above (the filename IS the expected hash); ots verify 99e1acc080c60af53f82397ad0d15774a8cf64148f1c5c0721987d1029b03ee9.pdf.ots -f 99e1acc080c60af53f82397ad0d15774a8cf64148f1c5c0721987d1029b03ee9.pdf (opentimestamps.org) proves the capture time (fresh proofs report 'pending' until bitcoin-anchored, typically within a day).

Extracted text

Machine-extracted text (layout may be lost; the authoritative content is the stored file above).

Bria Artiﬁcial Intelligence Ltd.
Public Summary of Training Content

FIBO, FIBO Lite and FIBO Edit

Required by Article 53(1)(d) of Regulation (EU) 2024/1689 (AI Act), drawn up according to the Template
annexed

Commission

Communication

C(2025)

5235

ﬁnal

July

2025.

Version of the Summary Version 1.0. No previous versions.
Last update 02/08/2026
1.

General

information

1.1.

Provider

identiﬁcation

Provider name and contact details
Bria Artiﬁcial Intelligence Ltd. 10 Kremenisty St., Tel Aviv, Israel legal@bria.ai https://www.bria.ai
Authorised representative name and contact details
Vered Horesh Chief AI Strategy Ofﬁcer legal@bria.ai
1.2.

Model

identiﬁcation

Versioned model name(s) FIBO, FIBO Lite and FIBO Edit. This Summary covers all three, whose training content is identical. FIBO Model Card: https://huggingface.co/briaai/FIBO FIBO Lite Model Card: https://huggingface.co/briaai/Fibo-lite Fibo Edit Model Card: https://huggingface.co/briaai/Fibo-Edit
Model dependencies FIBO is the product of Bria's own pre-training run and is not a modiﬁcation or ﬁne-tune of another provider's general-purpose AI model. Bria discloses on a voluntary basis that the model incorporates two third-party open-weight components which Bria did not train and does not modify, and which are used with their published weights held frozen: • Wan2.2-VAE, provided by the Wan Team at Alibaba Group under the Apache License 2.0, used as the variational autoencoder. Source: https://github.com/Wan-Video/Wan2.2. Wan2.2 was placed on the market before 2 August 2025 and falls within the transitional period in Article 111(3) AI Act; its provider has not published a summary of training content.
1
• SmolLM3-3B, provided by Hugging Face under the Apache License 2.0, used as the text encoder. Source: https://huggingface.co/HuggingFaceTB/SmolLM3-3B. Pretrained by its provider on 11.2 trillion tokens of web, code, mathematics and reasoning data. Its provider has published a summary of training content on the model card, to which reference is made. The data used to train these components lies outside Bria's content licensing programme and Bria makes no representation as to its provenance. Statements elsewhere in this Summary concerning the exclusive use of commercially licensed content and the absence of crawling relate to the data on which Bria trained the model.
Date of placement of the model on the Union market
FIBO on or about 29 October 2025; FIBO Lite on or about 1 December 2025; FIBO Edit on or about 16 January 2026.
1.3.

Modalities,

overall

training

data

size

and

other

characteristics
Reported for the data on which Bria trained the model. Data used by the third-party providers of the
components

identiﬁed

Section

1.2

not

included

and

not

known

Bria.

Modality Training data size Types of content
Image 479 million images Fully licensed images provided by Bria data partners under transactional commercial licensing agreements with rightsholders globally.
Text Up to 120 billion tokens Captions describing the licensed images in the training set. Fewer than 100 million images carry long structured JSON captions covering attributes including subject, composition, lighting, colour and camera parameters; the remainder carry shorter captions. All captions are generated as metadata to enrich the licensed image data.
Audio / Video / Other
Not selected. N/A

Latest date of data acquisition / collection for model training
June 2025. The model is not continuously trained on new or dynamic data after that date.
Description of the linguistic characteristics of the overall training data
The structured JSON captions are in English. The SmolLM3-3B text encoder identiﬁed in Section 1.2 natively supports English,
2
French, Spanish, German, Italian and Portuguese, which bears on the model's effective prompt language coverage.
Other relevant characteristics of the overall training data
Bria's model weights are trained exclusively on content obtained through commercial licensing agreements with data partners globally, ensuring diverse representation across cultures, ethnicities, ages, genders and geographical locations, with balanced coverage across domain categories and content from multiple international markets. All such training data consists of human-created content with explicit commercial use releases, deliberately excluding public ﬁgures, harmful material and copyrighted ﬁctional characters. This description relates to the data Bria licensed and trained on. It does not extend to the unmodiﬁed third-party components identiﬁed in Section 1.2.
2.

List

data

sources

2.1.

Publicly

available

datasets

Have you used publicly available datasets to train the model?
No. All training data is sourced exclusively through transactional commercial licensing agreements that expressly authorise the use of licensed content for generative AI model training. Each agreement includes warranties and representations from the licensor conﬁrming its full legal right, title and authority to license the data objects and to grant the rights necessary for AI training applications.
2.2.

Private

non-publicly

available

datasets

obtained

from

third

parties

2.2.1 Have you concluded transactional commercial licensing agreement(s) with rightsholder(s) or their representatives?
Yes.
Modality(ies) covered Image.
2.2.2 Have you obtained private datasets from third parties that are not licensed as described in Section 2.2.1?
No.
General description All content is sourced from rightsholders or their authorised representatives, who have obtained all necessary model
3
releases, property releases and intellectual property clearances.
2.3.

Data

crawled

and

scraped

from

online

sources

Were crawlers used by the provider or on their behalf?
No. Bria does not engage in web crawling or scraping. Accordingly no crawler identiﬁers, collection period or domain name list falls to be disclosed.
2.4.

User

data

Was data from user interactions with the AI model used to train the model?
No.
Was data collected from user interactions with the provider's other services or products used to train the model?
No.
2.5.

Synthetic

data

Was synthetic AI-generated data created by the provider or on their behalf to train the model?
No. Note: in accordance with the AI Ofﬁce's instructions, this Section does not refer to the use of AI models to clean or enrich data, for example AI-generated metadata to enrich or modify a dataset such as creating text descriptions of images. Bria does use captions generated with third-party AI models for enrichment of its data.
2.6.

Other

sources

data

Have data sources other than those described in Sections 2.1 to 2.5 been used to train the model?
No. No ofﬂine or self-digitised sources, no datasets labelled by human subjects commissioned by the provider, and no human-generated data obtained through reinforcement learning have been used.
3.

Data

processing

aspects

3.1.

Respect

reservation

rights

from

the

TDM

exception

limitation

Are you a Signatory to the Code of Practice for general-purpose AI models
Yes. Bria is a signatory to the General-Purpose AI Code of Practice.
4
that includes commitments to respect reservations of rights from the TDM exception or limitation?
Describe the measures implemented before model training
Bria performs no crawling or scraping and does not rely on the text and data mining exception or limitation for the acquisition of training data. All training content is acquired under transactional commercial licensing agreements with rightsholders or their authorised representatives, each including warranties as to the licensor's right, title and authority to license the content for generative AI model training. Bria maintains legal veriﬁcation of all licensing agreements, detailed data cataloguing systems, and regular audits of data source compliance. Bria nonetheless acknowledges and respects reservations of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790 and maintains systems to identify and comply with such reservations where applicable. Bria's Copyright Policy under Article 53(1)(c) AI Act is published here.
3.2.

Removal

illegal

content

General description of measures taken
Bria's training data sourcing methodology inherently limits the inclusion of illegal content through the exclusive use of commercially licensed datasets obtained from veriﬁed professional data partners and established image bank providers. Each licensing agreement requires the licensor to warrant that the licensed content is original and human-created, does not infringe third-party copyright, trademark or other intellectual property rights, does not feature public ﬁgures, harmful material or copyrighted ﬁctional characters, and is subject to all necessary model and property releases where identiﬁable persons or property are depicted. Bria operates an automated cataloguing pipeline that records each asset with a clear link to its origin, applies content moderation ﬁltering to exclude harmful material, and subjects data partners to periodic compliance audits and remediation procedures. Training datasets are fully anonymised through removal of naming and identifying metadata.
3.3.

Other

information

(optional)

Other relevant information about data processing
Bria applies C2PA Content Credentials to model outputs and holds a root on the C2PA trust list. Bria operates a patented attribution technology that measures each data licensor's
5
contribution to generated content and provides recurring compensation on that basis. Rightsholders may submit precise and substantiated complaints to legal@bria.ai; this mechanism complements and does not limit any legal remedies available under Union or national copyright law.

6
Bria Artiﬁcial Intelligence Ltd.
Public Summary of Training Content

Bria 3.2

Required by Article 53(1)(d) of Regulation (EU) 2024/1689 (AI Act), drawn up according to the Template
annexed

Commission

Communication

C(2025)

5235

ﬁnal

July

2025.

Version of the Summary Version 2.0. Supersedes the version ﬁrst published with effect from 1 September 2025 and last updated 6 January 2026.
Last update 02/08/2026
1.

General

information

1.1.

Provider

identiﬁcation

Model

identiﬁcation

Versioned model name(s) Bria 3.2 (briaai/BRIA-3.2). https://huggingface.co/briaai/BRIA-3.2
Model dependencies Bria 3.2 is the product of Bria's own pre-training run and is not a modiﬁcation or ﬁne-tune of another provider's general-purpose AI model. Bria discloses on a voluntary basis that the model incorporates two third-party open-weight components which Bria did not train and does not modify, and which are used with their published weights held frozen: • T5 version 1.1 XXL, provided by Google under the Apache License 2.0, used as the text encoder. Bria uses the t5-v1_1-xxl variant. As documented by its provider, T5 version 1.1 was pretrained on the Colossal Clean Crawled Corpus (C4), a dataset derived from Common Crawl web data. T5 was placed on the market well before 2 August 2025 and falls within the transitional period in Article 111(3) AI Act; its provider has not published a summary of training content.
7
• An open-source variational autoencoder with an f8 spatial compression factor, used for encoding images to and decoding images from latent representations. The data used to train these components lies outside Bria's content licensing programme and Bria makes no representation as to its provenance. Statements elsewhere in this Summary concerning the exclusive use of commercially licensed content and the absence of crawling relate to the data on which Bria trained the model.
Date of placement of the model on the Union market
10 June 2025.
1.3.

Modalities,

overall

training

data

size

and

other

characteristics
Reported for the data on which Bria trained the model. Data used by the third-party providers of the
components

identiﬁed

Section

1.2

not

included

and

not

known

Bria.

Latest date of data acquisition / collection for model training
June 2025. The model is not continuously trained on new or dynamic data after that date.
Description of the linguistic characteristics of the overall training data
The captions are in English.
Other relevant characteristics of the overall training data
Bria's model weights are trained exclusively on content obtained through commercial licensing agreements with data partners globally, ensuring diverse representation across cultures, ethnicities, ages, genders and geographical locations. The dataset maintains balanced coverage across domain categories while incorporating content from multiple international markets. All such training data consists of human-created content with explicit commercial use releases,
8
deliberately excluding public ﬁgures, harmful material and copyrighted ﬁctional characters. This description relates to the data Bria licensed and trained on. It does not extend to the unmodiﬁed third-party components identiﬁed in Section 1.2, whose training data lies outside Bria's licensing programme and is not known to Bria.
2.

List

data

sources

2.1.

Publicly

available

datasets

Private

non-publicly

available

datasets

obtained

from

third

parties

2.2.1 Have you concluded transactional commercial licensing agreement(s) with rightsholder(s) or their representatives?
Yes.
Modality(ies) covered Image.
2.2.2 Have you obtained private datasets from third parties that are not licensed as described in Section 2.2.1?
No.
General description All content is sourced from rightsholders or their authorised representatives, who have obtained all necessary model releases, property releases and intellectual property clearances.
2.3.

Data

crawled

and

scraped

from

online

sources

2.4.

User

data

Synthetic

data

Other

sources

data

Data

processing

aspects

3.1.

Respect

reservation

rights

from

the

TDM

exception

limitation

Are you a Signatory to the Code of Practice for general-purpose AI models that includes commitments to respect reservations of rights from the TDM exception or limitation?
Yes. Bria is a signatory to the General-Purpose AI Code of Practice.
Describe the measures implemented before model training
Bria performs no crawling or scraping and does not rely on the text and data mining exception or limitation for the acquisition of training data. All training content is acquired under transactional commercial licensing agreements with rightsholders or their authorised representatives, each
10
including warranties as to the licensor's right, title and authority to license the content for generative AI model training. Bria maintains legal veriﬁcation of all licensing agreements, detailed data cataloguing systems, and regular audits of data source compliance. Bria nonetheless acknowledges and respects reservations of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790 and maintains systems to identify and comply with such reservations where applicable. Bria's Copyright Policy under Article 53(1)(c) AI Act is published here.
3.2.

Removal

illegal

content

Other

information

(optional)

Other relevant information about data processing
Bria applies C2PA Content Credentials to model outputs and holds a root on the C2PA trust list. Bria operates a patented attribution technology that measures each data licensor's contribution to generated content and provides recurring compensation on that basis. Rightsholders may submit precise and substantiated complaints to legal@bria.ai; this mechanism complements and does not limit any legal remedies available under Union or national copyright law.
11
Bria Artiﬁcial Intelligence Ltd.
Copyright Policy

Article 53(1)(c) of Regulation (EU) 2024/1689

Version Version 2.0. Supersedes the version dated September 2025.
Last update 02/08/2026
Models covered All general-purpose AI models placed on the Union market by Bria, including the Bria 3.x family and the FIBO family.
1.

Compliance

Framework

Bria

Our commitment to copyright compliance is fundamental to our responsible AI development. We
recognise

that

true

accountability

requires

rigorous

adherence

intellectual

property

rights

and

legal

standards.

Our copyright policy reﬂects our core principles:
• Respecting rightsholders • Ensuring full legal compliance • Maintaining the highest standards of ethical AI development • Providing equitable compensation to all data licensors through our proprietary and patented
algorithmic

attribution

technology

that

measures

and

rewards

each

contributor's

impact

generated

content
This document outlines our comprehensive approach to copyright compliance for our general-purpose
AI

models

placed

the

market.

We have developed this policy to demonstrate our commitment to:
• Fully complying with laws and regulations on copyright and related rights, speciﬁcally Article
53(1)(c)

the

Act,

and

Article

4(3)

Directive

(EU)

2019/790

and

rights

the

Digital

Single

Market.
• Identifying and respecting the rightsholder's reservations of rights. • Utilising state-of-the-art technologies for copyright protection, ensuring that the model weights
Bria

trains

are

trained

only

fully

licensed

training

data.

Where

Bria

model

incorporates

unmodiﬁed

third-party

open-weight

component

that

Bria

did

not

train,

that

component

identiﬁed

the

Public

Summary

Training

Content

for

the

model

concerned,

and

section

4.2

below

sets

out

the

controls

Bria

applies

such

components.
• Being committed to advancing AI technology while adhering to principles that prioritise societal
beneﬁt

and

accountable

development.
Our approach goes beyond mere legal compliance. It represents an intrinsic commitment to integrity,
transparency,

and

responsible

innovation

our

model

development

and

deployment

processes.

By successfully building high-quality models through sustainable data partnerships, Bria demonstrates
that

responsible

innovation

and

respect

for

intellectual

property

are

not

only

possible

but

also

12
commercially viable. This evidence is crucial for policymakers, regulators, and courts, showing there is
no

need

choose

between

fostering

progress

and

protecting

creators.

Training

Data

Compliance

and

Veriﬁcation

Licensed Data Use Policy Bria exclusively uses commercially licensed training data, explicitly authorised for generative AI model
training,

for

the

model

weights

trains.

maintain

comprehensive

documentation

and

records

all

our

licensing

agreements

and

the

authorisations

provided

our

data

partners.
Data Source Veriﬁcation Our data licensing process includes:
• Legal veriﬁcation of all licensing agreements conﬁrming that: Bria has entered into valid and enforceable agreements with all data licensors that expressly
authorise

the

use

licensed

data

for

generative

model

training

purposes
Each licensing agreement includes comprehensive warranties and representations from data
licensors

conﬁrming

their

full

legal

right,

title,

and

authority

license

the

data

objects

and

grant

the

rights

necessary

for

training

applications
• Maintenance of detailed data cataloguing systems • Regular audits of data source compliance

No Web-Crawling Policy Bria does not and will not engage in web-crawling activities, and does not use publicly accessible online
content

training

data

for

the

model

weights

trains.

This

approach

substantially

reduces

the

risk

associated

with

unlicensed

training

content.
Where a Bria model incorporates an unmodiﬁed third-party open-weight component, the data used by
that

component's

provider

train

lies

outside

Bria's

licensing

programme

and

not

known

Bria.

Such

components

are

identiﬁed

the

Public

Summary

Training

Content

for

the

model

concerned,

together

with

any

summary

training

content

published

their

provider.
Rights Reservation Compliance While our business model does not involve web-crawling, we acknowledge and respect rightsholder
reservations

expressed

pursuant

Article

4(3)

Directive

(EU)

2019/790

and

maintain

systems

identify

and

comply

with

such

reservations,

applicable.

Data

Partner

Compensation

and

Attribution

Framework

Fair Compensation We compensate partners based on their measured contribution to generated content, providing a
recurring

revenue

stream

data

partners.

Bria

rejects

one-time

licence

deals

for

training

data,

instead

implementing

fair

and

transparent

system

ongoing

compensation

that

ensures

creators

receive

the

long-term

value

their

work

generates

existing

and

future

applications.

Attribution Technology
13
Our patented attribution technology enables Bria to bridge the gap between demand for synthetic
content

and

supply

authentic

training

data.

Each

synthetic

output

generated

our

models

attributed

the

original

content

creators

that

had

the

greatest

impact

its

generation.

Transparency Bria's attribution technology offers transparency to both data contributors and customers.
4.

Due

Diligence

for

Third-Party

Sources

and

Components

4.1

Content

obtained

through

third-party

data

providers
Bria maintains comprehensive due diligence procedures for any content obtained through third-party
providers.

require

all

third-party

data

providers

to:

• Warrant their full legal right and authority to license content for AI training purposes • Provide comprehensive warranties that all data licensed is original and human-created content • Ensure all data is subject to legal releases covering commercial use and AI training applications
where

identiﬁable

persons

property

are

depicted
• Provide representations conﬁrming non-infringement of third-party copyrights, trademarks, or
other

intellectual

property

rights
• Provide warranties that content does not feature public ﬁgures, harmful materials, or copyrighted
ﬁctional

characters
• Present all applicable compliance certiﬁcations regarding adherence to all applicable copyright
laws

and

regulations
• Maintain detailed records and chain-of-title documentation • Be subject to periodic compliance audits and remediation procedures
4.2 Third-party model components Bria distinguishes between third-party data sources, addressed above, and third-party model
components

incorporated

into

Bria

model.
Certain Bria models incorporate open-weight components developed by third parties, such as text
encoders

and

variational

autoencoders.

Where

Bria

does

so,

those

components

are

used

with

their

published

weights

held

frozen

and

are

not

modiﬁed

Bria.

They

are

made

available

their

providers

under

open-source

licences,

which

characteristically

disclaim

all

warranties.

The

due

diligence

framework

section

4.1,

which

depends

contractual

warranties

from

data

licensors,

cannot

therefore

applied

them.
For such components Bria applies the following controls instead: • Use is restricted to components released under licences permitting the intended commercial use
and,

where

applicable,

redistribution.
• The licence of each component is reviewed before adoption, and the applicable notice, attribution
and

licence-reproduction

conditions

are

complied

with

every

channel

through

which

Bria

redistributes

the

component.
• Components are used without modiﬁcation. Where Bria does modify a component, any applicable
obligation

state

changes

complied

with.

14
• Each component is identiﬁed in the Public Summary of Training Content for the model concerned,
together

with

any

summary

training

content

published

its

provider.
• Bria makes no representation as to the provenance of the data used by third-party providers to
train

such

components,

and

its

representations

concerning

licensed

training

data

relate

the

model

weights

Bria

trains.
4.3 Data Procurement and Usage In addition, Bria maintains the following processes to ensure that all data licensed and used by it will not
subsequently

create

harmful

implications

for

rightsholders

users:

Data Quality and Diversity: Bria does not train its models on synthetic data generated for that purpose,
whether

through

model

distillation,

alignment

against

another

model's

outputs,

otherwise.

Consistent

with

the

Ofﬁce's

instructions

the

Public

Summary

Training

Content,

this

distinct

from

the

use

models

clean

enrich

data.

Bria

does

use

captions

generated

with

third-party

models

enrich

its

licensed

image

data,

disclosed

the

Public

Summary

Training

Content

for

each

model.

Bria

does

not

acquire

data

that

will

not

lead

model

quality

improvement

and

would

only

dilute

the

attribution

existing

partners.

Dedicated Storage and Access: Training data for generally available models is stored in dedicated
training

accounts.

Proprietary

data

provided

customers

for

training

private

models

segregated

from

general

models'

training

data

accounts.

Access

proprietary

data

provided

customers

for

training

private

models,

and

models

informed

customers'

proprietary

data,

restricted

such

customers

and

authorised

Bria

support

for

maintenance

functions

only,

under

rigorous

permissions

system.

Access

models

subject

periodic

penetration

tests.

Asset Cataloguing: An automated pipeline records each asset in a dedicated catalogue, maintaining a
clear

link

its

origin.

Output

Protection

and

Compliance

Bria implements robust technical and policy measures to prevent copyright-infringing outputs from our
general-purpose

models:

Technical Safeguards We employ state-of-the-art technical measures to prevent our models from reproducing training
content

infringing

manner,

including:

• Advanced content ﬁltering systems • Real-time output monitoring • Regular testing protocols to verify safeguard effectiveness • Detection of potential copyright violations post-generation
Acceptable Use Requirements Our terms and conditions explicitly prohibit:
• Using models to generate copyright-infringing content • Circumventing technical safeguards
15
• Using our models to create any unlawful content
Documentation Requirements For all model deployments, we provide clear copyright and acceptable-use compliance guidelines,
explicit

prohibition

infringing

uses,

and

technical

documentation

protective

measures.

Cross-Platform Application These protective measures apply consistently across direct model implementations, on-premises
model

integrations,

API

access,

and

all

other

deployment

scenarios.

Monitoring and Enforcement We maintain ongoing oversight through regular audits of model outputs, compliance veriﬁcation
procedures,

incident

response

protocols,

and

continuous

improvement

safeguards.

Rightsholder

Contact

and

Complaint

System

We maintain accessible communication channels and procedures for rightsholders, through an email
message

legal@bria.ai,

submit

precise

and

substantiated

complaints

regarding

non-compliance

with

commitments,

potential

infringement

concerns,

rights

reservation

violations,

and

any

other

matters.

This reporting mechanism complements but does not replace or limit any legal remedies available
under

Union

and

national

law.

Other

Protective

Measures

Beyond our copyright compliance framework, Bria maintains additional protective measures for the
beneﬁt

both

our

data

licensors

and

end

users.

Privacy and Data Protection • We place strong emphasis on privacy protection for individuals. We use human likenesses only with
explicit

releases

for

commercial

use.

strictly

prohibit

the

use

public

ﬁgures'

data

our

models.
• We implement robust data protection measures to safeguard personal information and comply with
relevant

privacy

and

data

protection

regulations.

Identity Protection and Training Architecture No Identity Mapping: Bria's training datasets are anonymised through the removal of naming and
identifying

metadata.

Our

models

are

designed

prevent

any

mechanism

for

linking

visual

likenesses

real

person

identities.

Balanced, Multi-Source Training Catalogue: We maintain a wide-ranging, globally sourced dataset that
deliberately

balances

input

across

creators,

contexts,

content

types,

and

personalities.

This

diversity

reduces

the

prominence

any

single

source

and

minimises

the

potential

for

overrepresentation

unintended

memorisation.

Architecture Design: Bria's models are architected to learn general visual patterns rather than memorise
or

reproduce

speciﬁc

individuals.

Our

technical

architecture

and

sampling

methodology

discourage

identity

persistence

overﬁtting

any

single

visual

subject.

16
Commitment to Diversity, Inclusion, and Bias Mitigation • We actively seek and incorporate diverse datasets representing a wide range of cultures,
ethnicities,

ages,

genders,

and

geographical

locations.
• Our development process includes testing and reﬁnement to identify and mitigate potential biases
in

our

models.
• We implement content moderation ﬁlters to prevent the generation of harmful or biased content. • We consult with diverse stakeholder groups, including underrepresented communities, and
incorporate

their

feedback

into

our

development

and

deployment

processes.
• Our AI models are designed with cultural sensitivity in mind. • We publish reports on our diversity and inclusion initiatives, including the composition of our
datasets

and

the

results

our

bias

mitigation

efforts.

Safety, Provenance, and Mitigation of Harmful Content Bria applies C2PA Content Credentials to content generated by its models and holds a root on the C2PA
trust

list.

Content

Credentials

are

cryptographically

signed

provenance

metadata

bound

the

generated

ﬁle,

allowing

recipients

verify

that

the

content

was

produced

Bria

model.

Provenance

metadata

this

kind

can

removed

lost

where

ﬁle

subsequently

re-encoded,

converted,

stripped

metadata

downstream

tools

that

not

preserve

it.

Bria

developing

additional

marking

techniques

intended

improve

the

durability

provenance

signals

across

such

transformations.
• We enforce strict guidelines against generating harmful or offensive content. • We implement measures to prevent misuse of our platform. We do not accept data that features
public

ﬁgures

harmful

content,

which

makes

impractical

use

our

models

generate

these

concepts.

Conclusion

This Copyright Policy establishes Bria's framework for responsible AI development through ﬁve core
pillars:

1. Exclusive use of commercially licensed training data, with veriﬁed legal authorisation, for the
model

weights

Bria

trains,

and

disclosure

any

third-party

model

components

incorporated

into

Bria

model.
2. Proprietary attribution technology that provides ongoing compensation to data licensors based
on

measured

contribution

impact.
3. Robust technical safeguards preventing copyright-infringing outputs. 4. Rigorous due diligence procedures for third-party data sources, and deﬁned controls for
third-party

model

components.
5. Accessible rightsholder communication channels for compliance concerns. Our approach encompasses additional protective measures including privacy protection, diversity and
inclusion

initiatives,

bias

mitigation

protocols,

and

safety

measures

against

harmful

content

generation.

Through our data partner empowerment and collaboration framework, we maintain long-term
partnerships

that

create

mutual

value

while

advancing

responsible

innovation.

Many

our

data

licensors

leverage

our

models

offer

their

creator

communities

and

customers

advanced

generative

capabilities

under

preferred

commercial

arrangements.

view

our

data

licensors

integral

our

sustainable

development

approach.

Contact

Information

Bria Artiﬁcial Intelligence Ltd.
Email: legal@bria.ai
Website:

https://bria.ai/security-and-compliance