Europe loses the AI race – and that’s a good thing!

by Benjamin Paaßen (Berlin), Joachim Wündisch (Düsseldorf) & Torbjørn Cunis (Stuttgart), Fellows of the Young ZiF

It appears that a race for the most powerful artificial intelligence (AI) is in progress. Major corporations in the US and China compete for ever larger data sets, more refined algorithms, and faster computing clusters1. The European Union is left behind: While European research groups are present at major AI conferences, no European institution is even close to having the resources of Alphabet, Facebook, Amazon, or Tencent. Most breakthroughs in recent years, such as solving image recognition challenges, large language models, or games, have been achieved by US-based companies2. In these kinds of tasks, EU research may never catch up. However, the question is: Should we even want to?

Artificial intelligence via human recklessness

Developments in the last five years have shown time and again that a reckless application of AI methods can cause serious ethical problems. In 2016, ProPublica revealed that a system employed in the US criminal justice system judged Black defendants to be more dangerous than white defendants, even if this turned out to be wrong3. In 2018, Amazon scrapped a planned system for pre-screening of job applications because it predicted that men would make better software engineers than women4. In the same year, Joy Buolamwini and Timnit Gebru showed that commercially available face recognition technology was much less accurate for darker-skinned people compared to light-skinned people, and especially inaccurate for darker-skinned women5. Even seemingly innocent language models appear to learn stereotypes embedded in human language, e. g., when reasoning that ‘Man is to Computer Programmer as Woman is to Homemaker’6. It appears that inaccurate model assumptions, biased training data, or questionable embedding into the broader social context can make AI decisions at least as problematic as human decisions.

Bigger is not always better

More broadly, there is a tension between current cutting-edge AI and EU policy goals. The lion’s share of resources in the AI race has been spent on a tiny subset of AI, namely deep neural networks. Deep neural networks are large-scale computers with billions of free parameters that are automatically optimized over weeks or months on terrabytes of data in giant computing clusters. Such systems have, undoubtedly, achieved unprecedented success in image processing tasks – such as face recognition – or language processing tasks – such as automated translation7. However, their sheer size and complexity makes their decisions almost impossible to explain, which is problematic as soon as decisions concern humans89. Further, the amount of energy consumed will likely complicate the fight against climate change10. The insatiable need for more training data has oftentimes led to questionable methods of data acquisition, such as collecting photos of millions of people without their consent [10] or scraping large-scale text data from the web, including harmful stereotypes and hate speech11. Finally, it appears that deep neural networks are particularly hard to protect against hacking attacks, making their application a potential security risk1213.

A vision for better AI in the EU

In summary, it appears that the current AI race focuses almost single-mindedly on achieving higher performance measures without much regard for environmental, legal, or ethical concerns. This kind of AI is not suitable to the EU context. Rather, EU research should focus on different goals: Using as little energy, time, and personal data as possible, making models explainable, and achieving AI that respects the autonomy and dignity of human beings14151617. Ironically, an AI subject to these constraints may even be more practically useful because savings in energy, time, and data, in combination with explainability and ethical acceptability are likely to outweigh a few percentage points of benchmark performance1819. Further, the EU is uniquely positioned to advance this vision: It arguably has the strictest data and privacy protections, including a rule requiring automatic decision-making systems to explain themselves to the user20. Further, it has released guidelines for trustworthy AI which point the way towards a better AI21. And it has a strong, publicly funded research community, which can drive research into better AI even without being led astray by narrow economic incentives. There is a long way to go, though, and success is not guaranteed. More efforts must be made to translate research prototypes into open-source software and datasets that can fuel applications. Beyond innovation in AI itself, a strong programme from the social sciences and the humanities is needed to translate an abstract vision for better AI into actionable guidelines for every single project which attempts to apply AI to human lives. Engineers should not be left alone with ethical considerations.

Overall, AI has the potential to propel humanity towards both dystopian and utopian futures. How it affects us depends on how it is built and how it is embedded into society. Without justification, the performance of AI systems – as measured by narrow metrics – is treated as synonymous with its social utility. Europe should not fall into this trap. Instead, it should move off the race-track and start running orthogonally, towards better AI.

References

  1. Savage (2020). The race to the top among the world’s leaders in artificial intelligence. nature 588, pp. 102-104. doi:10.1038/d41586-020-03409-8 []
  2. LeCun, Bengio, and Hinton (2015). Deep learning. nature 521, pp. 436-444. doi:10.1038/nature14539 []
  3. Angwin et al. (2016). Machine Bias. ProPublica. []
  4. Dastin (2018). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters. []
  5. Buolamwini and Gebru (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of the 1st Conference on Fairness, Accountability, and Transparency, p. 77-91. []
  6. Bolukbasi et al. (2016). Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings. Proceedings of the 30th International Conference on Neural Information Processing Systems, p. 4356–4364. https://arxiv.org/abs/1607.06520 []
  7. LeCun, Bengio, and Hinton (2015). Deep learning. nature 521, pp. 436-444. doi:10.1038/nature14539 []
  8. Guidotti et al. (2018). A Survey of Methods for Explaining Black Box Models. ACM Computing Surveys, 2018(93). doi:10.1145/3236009 []
  9. Brkan (2019). Do algorithms rule the world? Algorithmic decision-making and data protection in the framework of the GDPR and beyond. International Journal of Law and Information Technology, 27, p. 91-121. []
  10. Bender et al. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜 Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, p. 610-623. doi:10.1145/3442188.3445922 []
  11. Bender et al. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜 Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, p. 610-623. doi:10.1145/3442188.3445922 []
  12. Akhtar and Mian (2018). Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey. IEEE Access (6), pp. 14410 – 14430. doi:10.1109/ACCESS.2018.2807385 []
  13. Hamon, Junklewitz, and Sanchez (2020). Robustness and Explainability of Artificial Intelligence. JRC Technical Report. []
  14. Hamon, Junklewitz, and Sanchez (2020). Robustness and Explainability of Artificial Intelligence. JRC Technical Report. []
  15. Craglia et al. (2018). Artificial Intelligence: A European Perspective. Joint Research Centre of the European Commission. doi:10.2760/11251 []
  16. Edwards (2020). European researchers look beyond deep learning. Engineering & Technology. []
  17. Ala-Pietilä et al. (2018). Ethics Guidelines for Trustworthy AI. High-Level Expert Group on Artificial Intelligence. https://ec.europa.eu/futurium/en/ai-alliance-consultation []
  18. Craglia et al. (2018). Artificial Intelligence: A European Perspective. Joint Research Centre of the European Commission. doi:10.2760/11251 []
  19. Edwards (2020). European researchers look beyond deep learning. Engineering & Technology. []
  20. Brkan (2019). Do algorithms rule the world? Algorithmic decision-making and data protection in the framework of the GDPR and beyond. International Journal of Law and Information Technology, 27, p. 91-121. []
  21. Ala-Pietilä et al. (2018). Ethics Guidelines for Trustworthy AI. High-Level Expert Group on Artificial Intelligence. https://ec.europa.eu/futurium/en/ai-alliance-consultation []