Changes in elements and items

• Issue #196 on logging of inferences/predictions: as a new element 5.6
• Issue #197 on enhancing element 6.1 with aggegated carbon footprint of AI activities, and/or a SCAP
• Broaden language of element 6.1 to energy consumption measurement (issue #197)

Changes in resources

• Add https://arxiv.org/ftp/arxiv/papers/2104/2104.10350.pdf to element 6.1 (issue #205)
• Add https://arxiv.org/abs/2111.00364 to element 6.1 or to section 6 directly? (issue #206)
• Add Quantmetry's blog article on frugality (issue #202)
• Issue #203 to introduce a first reference of compliance of foundational models / LLMs with AI Act

Changes in elements and items

• Fix a typo in element 2.1 wording (issue #189)

Changes in resources

• Add Giskard AI as reference tool for model testing and monitoring (issue #192): added to resources of elements 2.4, 3.5 and 5.3

Changelog of the 2022 H1 release

Changes in elements and items

• Finetune wording of elements 2.3 and 2.4 to widen discrimination to population bias in general
• Finetune wording of item 2.3.b to widen to knowledgeable and/or trained
• Remove item 3.1.d on multiple testsets as it didn't prove operationnally relevant
• Add a answer item 3.1.d on documenting the train-test split technical choices (#175)

Changes in resources

• Indicate in 1.1 that the CNIL MOOC on GDPR is currently being upgraded (#180)
• Add new CNIL technical resources on AI compliance with GDPR in elements 1.1 and 1.2
• Add a new academic paper on reconstructing training samples from a model, in elements 1.7 (thank you @celinejacques)
• Add paper on System Cards in elements 4.1 resources (#182)

Misc. changes

• Add new misc. articles in the references section of the repository (e.g. FTC 'algorithm destruction' capability, Covid19 AI models attempts)

Changelog of the 2021 H2 release

New evaluation elements and answer items

• Add a new element, numbered 2.1, on the gathering of data and the preparation of datasets for model training and evaluation (#173)
• Add a new element, numbered 2.3, on the evaluation of the risk of discrimination in data science projects (#166)
• Add an intra-element condition for non-concerned org. to elements 2.4 (ex-2.3) and 2.5 (ex-2.4) (#166)
• Add new answer item 6.1.f on transparency of CO2 impact (#164) by @gmartinonQM
• Add new answer item 6.3.b for work in progress on an ethical policy (#165 and #169)

Changes within evaluation elements and answer items

• Add a slight precision to answer item 1.9.b on communicating to stakeholders (#166)
• Rearrange answer items of element 1.9 in progressive order to make it a single-answer element (#169)
• Add a slight precision to answer item 3.7.a on needing or not to communicate on performance metrics (#165)
• Add a slight precision to answer item 4.2.a on relying on the practices of collaborators involved (#169)
• Split answer item 6.1.c into 6.1.c and 6.1.d to facilitate unambiguous answers (#164)

Misc. changes

• Add examples to element 1.4 on certifications related to personal data (#166)
• Add examples to element 2.2 (ex-2.1) on sensors/capture bias, and attention to data labels/annotations (#173)
• Replace references to "predictive models" by "AI models" to enable a more generic perspective (#166)
• Replace wording "model genealogy" by "model lifecycle documentation" for clarity (#170)
• Add numerous references on environmental impact of AI (#164) by @gmartinonQM
• Add Numeum's guide and the LNE's certification framework

Resources/references to add

• Fixes Add OpenDP as a ref and resource on diff priv #97 : OpenDP on differential privacy
• Fixes Add HRX's article into the "various controversies" section #99 : HRX article on AI use cases controversies (via @meuce)
• Fixes Add interesting references #103 : Misc. references
• Fixes Add interesting reference: factsheet by IBM #106 : IBM factsheet
• Fixes Add Shapash and FACET in resources #112 : Shapash & FACET on explainabillity
• Fixes Add DataforGood ressources #114 : Resources elaborated during the Dataforgood season 8 project (fairness, genealogy, robustness)
• Fixes Références : The Global Landscape of AI Guidelines & Code Carbon #116 : Code Carbon (via @SaboniAmine)
• Fixes Reference: Principled Artificial Intelligence | Berkman Klein Center (harvard.edu) #120 : Berkman meta study on AI ethics principles
• Fixes ML Exploit : Remote code execution from pickle files #134 : ML Exploit with pickle files
• Fixes Add Counterfit to resources on ML security #144 : Counterfit to test ML models vulnerabilities

Changes in the evaluation elements and answer items

• Fixes Add an answer to Q4.3 #108 : New answer item to Q4.3 on sharing/publishing AI incidents (via D. Bartolo)
• Fixes AI registers #123 : Public AI Registers (new reference + new answer item to 5.5)
• Fixes Moving beyond "algorithmic bias is a data problem" #136 : New evaluation element on modeling- and learning-related biases
• Fixes [User feedback on the assessment] Missing answer item #140 : New answer item to 1.4 on compliance to personal data regulations (user feedback)
• Fixes A bunch of suggestions #133 : New answer item to Q1.5 on alternatives to minimisation principle (via @JustineBoulant)

Misc. fixes

• Fixes Set an explanation to each evaluation element #109 : Add missing explanations/hints in some elements
• Fixes Resource - dead link #118 : Broken link
• Fixes Missing blank lines between section titles and keywords #119 : Missing newlines
• Fixes A bunch of suggestions #133 : Clarify README and References.md, fix typos, enhance explanations/context infos on ML vulnerabilities in Q1.7-1.8)

• Merge of former sections 4 and 5 into a new Section 4 titled "Assurer la reproductibilité des modèles et en établir la chaîne de responsabilité"
• Updated and finetuned wording for evaluation elements 1.7, 1.8, 2.1, 2.3, 3.3
• New technical resources and real-life illustrations of risks

Complement the assessment referential with new evaluation elements, updated answer items, and a lot of new and reorganised resources.