We aim to act responsibly, create and promote an AI that is lawful, ethical, inclusive and robust. It is important to ensure the protection of fundamental (human) rights and user safety.

To that end, this document gathers a set of concrete guidelines, that are structured around 7 principles. Those principles about AI trustworthiness were primarily identified by the European Union (EU). The reason why we tackle AI trustworthiness by complying with these principles is that we believe they provide a relevant overview of all components involved in the development of an AI system.

The primary audience of this document is AI practitioners, especially because it conveys a set of technical guidelines. However, concepts are always introduced on a high level scale, before being tackled on a lower level - and technical - scale. As a result, we strongly encourage non-AI practitioners to also read it.

Human oversight

As stated by the EU, “human oversight ensures that an AI system does not undermine human autonomy or cause other adverse effects”.

Guidelines depending on the AI system life cycle: AI contributors should always assess “why” they set up an AI system prior to actually doing it ; and users of an AI system must be able to control it.

Risk & criticality: AI contributors must evaluate the risk and criticality of an AI system before starting to implement it, namely during ideation phases, among following risks: Unacceptable Risk, High Risk, Limited Risk, Minimal Risk.

Practical guidelines: https://datacraft-paris.github.io/trustworthyai/human.html 

Privacy and data governance

Trustworthy AI must ensure that privacy, as a fundamental right, is respected for all parties – users, individuals targeted or employees. Quality and integrity of the data must also be protected, and access to data must be regulated.

Respect for privacy and data protection: Any system gathering data must ensure that the consent of the user is respected, and his privacy protected.

Quality and integrity of data: To avoid any harmful consequence of an AI system, the quality and integrity of the data must be ensured at any time

Access to data: As a general principle, only people with valid reasons should access each dataset

Practical guidelines: https://datacraft-paris.github.io/trustworthyai/privacydatagov.html 

Technical robustness and security

Development robustness: AI contributors must comply with development standards to ease reliability, readability of procedures and handovers.

Data project lifecycle safety:

• AI contributors must account for potential biases and not introduce additional ones While training AI models, AI contributors must carefully understand model’s underlying mechanisms, carefully calibrate them, and carefully analyze results
• While industrialising AI models, AI contributors must ensure AI system’s sustainability

Security: Artificial Intelligence is no exception when it comes to security, AI systems are exposed to a wide range of attacks. Consequently, AI contributors must be sensitised to Machine Learning security and enforce security audits, especially when humans are impacted by an AI system decision.

Practical guidelines: https://datacraft-paris.github.io/trustworthyai/robustnesssafety.html 

Transparency and explicability

Transparency and explicability are two fundamental and distinct notions when it comes to AI trustworthiness. While explicability aims at understanding the logic behind an algorithm, transparency aims at making “public” (totally or partially - within the scope of an organisation e.g.) the AI system and anything that lives with it (data, …).

Communication: AI contributors should always clearly and timely communicate results to stakeholders. 

AI contributor’s management: AI contributors should always be put in a favourable condition by the management to give the alert in case of fraudulent, illegal, illegitimate, discriminatory or unethical results.

AI explicability: AI explicability aims at creating numerical methods to extract the logics behind an AI system’s decision. AI contributors must be sensitised to understand results output by an AI system and study both global interpretation (at model level) and local interpretation (at individual level). This balance is the only way to achieve an exhaustive understanding of the AI system’s behavior, which in turns unlocks trust in the AI system.

AI transparency: While creating an AI system, AI contributors should enforce the system’s transparency to challenge the AI system’s weaknesses, challenge potential pitfalls of the training data, which in turn helps to check legal & ethical compliance, and to protect individuals.

Practical guidelines: https://datacraft-paris.github.io/trustworthyai/transparencyexplicability.html 

Diversity, non discrimination and fairness

AI contributors must foster the creation of bias-free AI systems and give themselves the means to reasonably arbitrage between performance and fairness, through the use of proper frameworks. We will discuss an AI project lifecycle through the eyes of fairness, and explain how one can detect discrimination through statistical methods.

AI Project lifecycle

• • At the start of the ideation & design phase of an AI project, AI contributors must be aware of potential discrimination risks/breaches in an AI project and avoid initiating a project that discriminates on purpose. Stakeholders should always be involved throughout the whole project and must have complete trust & understanding of the system and its risks.
  • AI contributors must anticipate bias and question the risk of discrimination in the data collection and preprocessing of data (origin, gender, age, characteristics related to the brand image, …)
  • While training (programming) AI models, AI contributors must carefully understand the system, its mechanisms, analyze results on fairness and correct bias where needed.
  • While industrializing AI models, AI contributors must ensure that the AI’s system is fair and non-discriminative. Feedback loops should be put in place when the model is deployed and proper actions should be taken when there is a signal of discrimination.

Identification metrics and ways for correcting discrimination bias and unfairness

• Quantify the risk of breach of fairness on identified sensitive populations, through the use of metrics
• After identification of bias, one might correct the bias through the use of algorithms
• Some open sources packages are created that might help in detection & correction of unfairness in an AI system

Practical guidelines: https://datacraft-paris.github.io/trustworthyai/diversitynondiscrimination.html 

Environmental and societal well-being

The EU Ethical AI Guidelines notes the following on this topic:

“AI systems should benefit all human beings, including future generations. It must hence be ensured that they are sustainable and environmentally friendly. Moreover, they should take into account the environment, including other living beings, and their social and societal impact should be carefully considered.”

This guideline, although a bit philosophical, means that we should be raising awareness of the impact of AI on the environment and societal well-being. The following table has the purpose to raise this awareness and it mentions some practical steps that AI contributors may take in this regard.

• AI contributors must design AI systems so that the resources used to implement them minimize the environment footprint, whether it is during the model development phase, model deployment phase, or during the inference phase. They should report environment-related metrics, and account for them as much as they account for pure machine learning or $-related performance metrics.
• AI systems may alter our conception of social agency or impact our social relationships. While AI systems may enhance social skills, it might equally contribute to social deterioration, affecting people’s physical and mental wellbeing. The effects an AI system might have on this should be carefully tracked and monitored.
• The effects of an AI system on institutions, democracy and society at large should be taken into account. In particular, situations related to the democratic process, including not only political decision making but also in electoral context.

Practical guidelines: https://datacraft-paris.github.io/trustworthyai/environmentalsocietal.html 

Accountability

Mechanisms must be put in place to ensure responsibility and accountability for AI systems and their outcomes, both before and after their deployment.

• AI contributors are responsible
• Evaluation by internal and external auditors
• Identification of negative impacts
• Ensure redress is possible

Practical guidelines: https://datacraft-paris.github.io/trustworthyai/accountability.html