Using AI to make predictions or decisions in e.g. quantitative science, healthcare, biology, economy and finance has been extensively, and possibly excessively done over the past several years. While the application of AI to these domains remains an active area of research, we believe that the biggest challenge for the future of AI lies ahead. Rather than just predicting or making decisions, AI solutions should be developed to conduct exploratory analyses, i.e., to find new, interesting patterns in complex systems or facilitate scientific discovery (Raghu and Schmidt, 2020). Specific cases where this direction has already been explored include e.g., drug discovery (Vamathevan et al., 2019), the discovery of new material (Butler et al., 2018), symbolic math (Lample and Charton, 2019; Stanley et al., 2019) or the discovery of new physical laws (Both et al., 2019; Iten et al., 2020; Udrescu and Tegmark, 2020). Will AI succeed in assisting humans in the discovery of new scientific knowledge? If so, in which domain will it happen first? How do we speed up the development of new AI methods that could reach such goals? These are some questions that should inspire and drive the applications of AI in other fields.

Another possible approach consists of using AI models as experimental “guinea pigs” for hypothesis testing. In the domain of neuroscience, one standard methodology consists of analyzing which AI model is best at predicting behavioral data (from animals or humans) in order to support or inform hypotheses on the structure and on the function of biological cognitive systems (Gauthier and Levy, 2019). In that case, the process of training the AI-agent is an experiment in itself since the intrinsic interest does not lie in the performance of the underlying algorithm per se but instead in its ability to explain cognitive functions. Can we create an AI algorithm that will replace all stages of scientific process, from coming up with questions, generating the data, to analysis and interpretation of results? Such automated discovery is considered as the ultimate goal by some experts, but so far remains out of reach (Bohannon, 2017).

Whereas AI approaches are readily impacting many scientific fields, those approaches also continue to benefit from insights from fields such as neuroscience (Hassabis et al., 2017; Samek et al., 2019; Ullman, 2019; Parde et al., 2020), for example the similarities between machine and human-like facial recognition (Grossman et al., 2019) and the use of the face space concept in deep convolutional neural networks (O’Toole et al., 2018; Parde et al., 2020). Other fields impacting AI research include evolutionary biology (Gobeyn et al., 2019) and even quantum mechanics (Biamonte et al., 2017). One of the biggest successes of integrating insights from other fields in modern day AI, the perceptron, became the prelude to the modern neural networks of today (Rosenblatt, 1958). Perceptrons and neural networks can be considered analogous to a highly reduced model of cortical neural circuitry. Other examples are algorithms such as reinforcement learning which drew inspiration from principles of developmental psychology from the 50s (Skinner, 2019) and have been influencing the field of developmental robotics (Cangelosi and Schlesinger, 2015) since the 2010s. Further illustration of this cross-fertilization can be seen in bio-inspired approaches, where principles from natural systems are used to design better AI, e.g., neuroevolution algorithms that evolve neural networks through evolutionary algorithms (Floreano et al., 2008). Finally, the rise of quantum computers and quantum-like algorithms could further expand the hardware and algorithmic toolbox for AI (Biamonte et al., 2017). Despite these important advances in the last decade, AI systems are still far from being comparable to human intelligence (and to some extent to animal intelligence), and several questions remain open. For instance, how can an AI system learn and generalize while being exposed to only a small amount of data? How to bridge the gap between low-level neural mechanisms and higher-level symbolic reasoning?

While AI algorithms are still mostly focused on the modeling of purely cognitive processes (e.g., learning, abstraction, planning…), a complementary approach could consider intelligence as an emergent property of cognitive systems through their coupling with environmental, morphological, sensorimotor, developmental, social, cultural and evolutionary processes. In this case, the highly complex dynamic of the ecological environment is driving the cognitive agents to continuously improve in an ever-changing world, in order to survive and to reproduce (Pfeifer and Bongard, 2006; Kaplan and Oudeyer, 2009). This approach draws inspiration from multiple scientific fields such as evolutionary biology, developmental science, anthropology or behavioral ecology. Recent advances in reinforcement learning have made a few steps in this direction. Agents capable of autonomously splitting a complex task into simpler ones (auto-curriculum) can evolve more complex behaviors through coadaptation in mixed cooperative-competitive environments (Lowe et al., 2017). In parallel, progress has also been made in curiosity-driven multi-goal reinforcement learning algorithms, enabling agents to autonomously discover and learn multiple tasks of increasing complexity (Doncieux et al., 2018). Finally, recent work has proposed to jointly generate increasingly complex and diverse learning environments and their solutions as a way to achieve open-ended learning (Doncieux et al., 2018). One related research direction are studies of systems that sequentially and continually learn (Lesort et al., 2020) in a lifelong setting, i.e., continual learning without experiencing the well known phenomenon of catastrophic forgetting (Traoré et al., 2019). When combined, this research puts forward the following questions: How can we leverage recent advances that situate AI agents within realistic ecological systems? How does the dynamic of such systems drive the acquisition of increasingly complex skills?

To ensure that AI solutions perform as we intended, it is important to clearly formulate the problem and to state the underlying hypothesis of the model. By matching formal problem expression/definitions to laws (intentions), functional and technical specifications, we ensure that the project has a well established scope and a path towards achieving this goal. These specifications have been set forward by the GDPR (General Data Protection Regulation) that published a self assessment template guiding scientists and practitioners to prepare their AI projects for society (Bieker et al., 2016). In short, products and services resulting from AI decision making must clearly define their applicability and limitations. Note that this differs from problem definition since it involves explicitly stating how the algorithm will address part or all of the original problem. The developers have to explicitly detail how they handle extreme cases and show that security of the user is ensured. It should be mandatory for the owner and user of the data to clearly and transparently state the known biases expressed by the dataset (similar to the way the secondary effects of medicines are clearly stated on the medication guide). While some of these are already addressed by the GDPR in the EU, similar regulation and standards are needed globally. An alternative, complementary approach would be to rely on the classical scientific method practices developed over the centuries. Relying on observation, hypothesis formulation, experimentation (Rawal et al., 2020) and evaluation allows us to understand causal relationships and promotes rigorous practices. AI would certainly benefit from explicitly integrating these practices into its research ecosystem (Forde and Paganini, 2019).

To control the functioning of AI algorithms and their potential inherent biases, clear, transparent and interpretable methodologies and best practices are required. Trustworthiness of AI-driven projects can be ensured by, for example, using open protocols of the algorithms functionality, introducing traceability (logs, model versioning, data used and transformations done on data) or the pre-definition of insurance datasets. In transversal domains such as software development, tools have been devised to prevent mistakes and model deterioration over time (such as automated unit tests). Establishing similar standards for AI would force data scientists to design ways to detect and eliminate biases, ultimately making sure that the algorithm is behaving as intended. If ethical standards can be encoded in the algorithm, then regulation can be imposed on the optimized objectives of AI models (Jobin et al., 2019).

The goal of AI should be to improve human condition and not further aggravate either existing inequalities (Gebru, 2019) or environmental issues in our societies. The AI service and product developers are likely to be at the center of this challenge - they are the ones that can directly prevent errors and biases in input data or future applications. They present a priori knowledge that can lead to or prevent misuse (conscious or unconscious). It is tempting to extensively employ libraries and “ready-to-use” code samples, as these make the production process faster and easier. However, especially when used by non-experts, the key features of AI models, e.g., data recasting, could easily be implemented incorrectly. The secondary users of AI tools must be able to measure the biases of their input data and obtained results, which can be done only if they are both aware of potential problems and if they have the necessary tools readily available.

As with any software, failures and mistakes will inevitably arise and a system has to be in place to assess how AI tools and services behave not only during development but also “in production.” The combination of decision logs and model versioning can allow us to verify and ensure the product outcomes are the ones intended. Here the question of independent authorities comes in order to regularly audit the AI products around us. Companies and AI product developers must be capable of “opening the black box” and clearly exposing the monitoring they perform over an algorithm. Opening the black box has already been set as an important goal in AI research (Castelvecchi, 2016), even if not all experts agree that this is necessary (Holm, 2019). It includes not only making the currently used model transparent, but more importantly being able to explain how it was designed, and examining its past states and decisions. For example, developers must track data drifting and deploy policies preventing an algorithm to produce unintended outcomes. So far, this has been left to good practices of individual developers, but we can envision construction of an authority in charge of auditing AI products regularly. One proposed approach has been to impose Adversarial Fairness during training or on the output (Adel et al., 2019). Independently of a particular way to ensure auditability and interpretability, the process should be co-designed not only by AI practitioners but all stakeholders, including the general public, following open science principles (Greshake Tzovaras et al., 2019). Auditability and interoperability considerations complement and extend the more obvious and direct requirements of robustness, security and data privacy in AI.

Finally, as for any technology, the usefulness of AI will have to be assessed against its environmental impact. In particular, life cycle assessment of AI solutions should be systematic. Here also, auditing by independent authorities could be a way to enforce environmental neutrality (Schwartz et al., 2019).

Besides impacting research and industry directly, AI is transforming the job market at a rapid pace. It is expected that approximately 80% of the population will be affected by these technological advancements in the near future (HolonIQ). Highly complex jobs (e.g., the medical, juridical or educational domains) will be redefined, some simpler, repetitive tasks will be replaced or significantly assisted by AI and new jobs will appear in the coming decades. For instance, budget readjustment and reeducation of people who lose their jobs, towards a clean energy shift, with only about 30% coming from governments (which amounts to less than 10% of the funds committed to coronavirus economic relief), could positively shift climate change (Florini, 2011). However, workers of these different fields received little to no formal education on AI, and more initiatives on sustainable AI (such as EarthDNA ambassadors or TeachSDG Ambassadors) are needed. Therefore, the AI transformation should come along with a transformation in education where educational and training programs will have to be adapted to these different existing professions.

The transformation in education can be implemented on four different levels: academic institutions, companies and governments. Academic institutions should not only prepare AI experts by providing in-depth training to move forward AI research but also focus on interdisciplinarity and attract diversity in AI. Three main axes for AI education should be: 1) high level AI experts who can train future generations 2) AI practitioners who can raise public awareness in their research and (3), broader public that can be informed directly, leading to decrease in a priori distrust.

The end users and beneficiaries of AI services and products, as the most numerous part of the population, must play a central role in their development. It is they who should have the final say on what global use of AI technologies should be pursued. However, to do so, they must have a chance to learn the fundamental principles of AI. This is not fundamentally different from educating the general public about any scientific topic with a global societal impact, may it be medical (e.g., antibiotic resistance, vaccination) or environmental (e.g., climate change). Providing the information and training at scale is not a trivial task, due to at least two major issues: 1) the motivation of the general public and 2) the existence of appropriate educational tools. Various online resources are available targeting the general public, such as Elements of AI in Finland or Objectif’IA in France. Interestingly, in the case of AI, the problem itself could also be a part of a possible solution - we can envisage AI playing a central role in creating adaptive learning paths, individual-based learning programs addressing the needs and interests of each person affected by AI technology. Educational tools designed with AI can motivate each individual by providing relevant, personalized examples and do it at the necessary scale. Interactions between AI and education is yet another example of interdisciplinarity in AI (Oudeyer et al., 2016), which can directly benefit not only the two fields, education and AI, but society and productivity as a whole.