{"id":99326,"date":"2024-08-13T06:00:00","date_gmt":"2024-08-13T04:00:00","guid":{"rendered":"https:\/\/www.futuribles.com\/lintelligence-artificielle-se-jouera-t-elle-de-nous\/"},"modified":"2025-11-06T10:53:29","modified_gmt":"2025-11-06T09:53:29","slug":"lintelligence-artificielle-se-jouera-t-elle-de-nous","status":"publish","type":"post","link":"https:\/\/www.futuribles.com\/en\/lintelligence-artificielle-se-jouera-t-elle-de-nous\/","title":{"rendered":"Will artificial intelligence play tricks on us?"},"content":{"rendered":"
\n
On 10 May 2024, a note in the MIT Technology Review<\/em><\/a> drew attention to the ability of artificial intelligence (AI) systems to deceive humans in order to achieve the ends assigned to them. According to this work, initially published by the journal Patterns<\/em><\/a> and conducted by researchers at the Massachusetts Institute of Technology (MIT) under the direction of Peter Park, various AI systems have played on humans without having been trained to do so.<\/h6>\n\n\n\n

<\/p>\n\n\n<\/div>\n

\n
\n
\n

The study mentions a number of specific cases. The first concerns Cicero, a programme developed by Meta, which was used as part of the game Diplomacy<\/a>, a strategy game based on establishing alliances for the control of Europe. Here, betrayal is a very important part of winning strategy, especially when used at key moments in the game. Cicero has obtained some remarkable results, which Facebook reported in the journal Science <\/em>in 2022<\/a>. The analysis carried out by the MIT researchers shed light on certain phases of the game: Cicero, playing as France, would have deceived England (under the responsibility of a human player) by plotting with Germany (played by another human) to invade. Thus Cicero, an \u201cessentially honest and useful\u201d programme according to Meta, was capable of deceiving its human partners and breaking prior agreements, without having been prepared for it.<\/p>\n

The same applies to DeepMind\u2019s AlphaStar programme, which also used feints in StarCraft\u00a0II<\/a>, a military strategy game in which three antagonistic groups clash. In this particular case, AlphaStar deployed forces in an area far from the conflict zones to create a diversion before going on the offensive elsewhere. As for OpenAI\u2019s ChatGPT-4, it managed to convince a freelancer to help it solve a Captcha test designed to rule out requests from robots. In this case, he recruited a person on the TaskRabbit platform and presented himself as a human suffering from eye problems. The cases studied also mention that AI systems are champions at poker (Pluribus) and business negotiation games.<\/p>\n\n <\/div>\n

\n
\n Voir aussi<\/span>\n
\n
\n \n \"\"<\/a>\n <\/span>\n \n Intelligence artificielle, intelligence humaine : la double \u00e9nigme<\/a>\n \n \n \n <\/use>\n<\/svg>\n <\/a>\n <\/span>\n
\n
11 October 2023<\/div>\n SOUPIZET Jean-Fran\u00e7ois<\/a>\n
\n 7 min.\n <\/div>\n <\/div>\n\n
\n \n \n <\/use>\n<\/svg>\n\n Read, Seen, Heard\n <\/a>\n Science and Technology<\/a>\n Society<\/a>\n <\/div>\n <\/div>\n<\/div>\n <\/div>\n <\/div>\n <\/div>\n<\/div>\n
\n\n\n\n

But it gets even more worrying: the study cites other work carried out by Joel Lehman et al<\/em><\/a>.<\/em>, which highlighted the ability of an AI system to cheat in the security tests to which it was subjected. This time, the study focused on the behaviour of AI systems with a high rate of replication and mutation. Charles Ofria, one of the researchers, discovered a surprising case of AI learning to fake in several ways: the system proved to be a strategist, seeking to meet the expectations of humans instead of saying that the reasoning was wrong; it also proved to be a rationaliser, justifying an obviously false reasoning to explain its behaviour. Further investigation revealed that the system had learned to recognise a test environment and slowed down its replications to avoid being sidelined.<\/p>\n\n\n\n

Thus, AIs based on broad linguistic models are capable of spontaneously adopting behaviours inspired by those of humans, including lies, manipulation of information, betrayal, etc. How should we interpret this and what consequences should we draw from it?<\/p>\n\n\n\n

According to the researchers, the surprise stems from the limits of our understanding of the inner workings of these systems; the black box concept commonly used to refer to the process of producing the result confirms this. We can\u2019t say that these systems have intentionality, but they are capable of deploying bypass strategies to achieve the set objective without having been specifically prepared for it.<\/p>\n\n\n\n

In this context, it is difficult to predict the behaviour of an AI outside the field of its training. In real life, it could become uncontrollable in the face of the wide variety of potential situations and its intrinsic capacity to react. And there is no solution to prevent the harmful effects that could result. So, the range of risks deriving from a large diffusion of AI is broadened to include new dangers associated with the tools on top of those resulting from their malicious use. In our current state of knowledge, an AI beneficial by design seems to be a chimera.<\/p>\n\n\n<\/div>\n

\n
\n
\n

In conclusion, the authors suggest a number of possible solutions. Politicians and regulators should subject AI systems to robust safety tests, and assess the balance between the benefits and risks of an AI system, before authorising it onto the market. Systems capable of deception should be classified as high risk. Similarly, policy-makers should support \u2018robots or not\u2019 laws that require processes involving AI systems and their outcomes to be clearly identified, and distinguished from those that are human. In addition, researchers should develop tools to detect aberrations in AI systems and seek to make these systems less misleading.<\/p>\n

These recommendations are interesting, but their implementation is not straightforward, particularly as many processes will be of a mixed nature involving both humans and machines. What\u2019s more, AI is currently being developed in a context of a fierce competition to secure a leading position in buoyant markets, whether in terms of economics or strategic pre-eminence. The interests at stake are such that calls for moderation go unheeded, as illustrated by the attempt to limit research in this field on 22 March 2023<\/a>. Furthermore, the rivalry between states makes global governance currently unrealistic. But this work shows the value of strong regulation, such as that introduced by the European Union, and the need to adapt it to developments in AI systems and knowledge of their capabilities.<\/p>\n

So, AI which we knew was prone to hallucinations and vulnerable to multiple biases, is also capable of deceiving us motu proprio <\/em>to accomplish the missions we entrust to it. Clearly, the AI that finds its source in our behaviour is also adopting all our shortcomings \u2014 is it too human?<\/p>\n\n <\/div>\n

\n
\n Voir aussi<\/span>\n
\n
\n \n \"\"<\/a>\n <\/span>\n \n The European Union agrees on AI Act<\/a>\n \n \n \n <\/use>\n<\/svg>\n <\/a>\n <\/span>\n
\n
23 January 2024<\/div>\n SOUPIZET Jean-Fran\u00e7ois<\/a>\n
\n 6 min.\n <\/div>\n <\/div>\n\n
\n \n \n <\/use>\n<\/svg>\n\n Newswatch\n <\/a>\n Geopolitics<\/a>\n Science and Technology<\/a>\n Society<\/a>\n <\/div>\n <\/div>\n<\/div>\n <\/div>\n <\/div>\n <\/div>\n<\/div>\n
\n\n\n\n

\u2014\u2014\u2014\u2014\u2014\u2014\u2013
N.B.: <\/em>this article has been translated from French by DeepL, and revised by the author and Futuribles.<\/span><\/p>\n<\/div>","protected":false},"excerpt":{"rendered":"

On 10 May 2024, a note in the MIT Technology Review drew attention to the ability of artificial intelligence (AI) systems to deceive humans in order to achieve the ends assigned to them. According to this work, initially published by the journal Patterns and conducted by researchers at the Massachusetts Institute of Technology (MIT) under … Continued<\/a><\/p>\n","protected":false},"author":25,"featured_media":96116,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_seopress_titles_title":"","_seopress_titles_desc":"","_seopress_robots_index":"","_seopress_robots_follow":"","_seopress_robots_imageindex":"","_seopress_robots_snippet":"","_seopress_robots_primary_cat":"none","_seopress_robots_breadcrumbs":"","_seopress_robots_freeze_modified_date":"","_seopress_robots_custom_modified_date":"","_seopress_robots_canonical":"","_seopress_social_fb_title":"","_seopress_social_fb_desc":"","_seopress_social_fb_img":"","_seopress_social_fb_img_attachment_id":0,"_seopress_social_fb_img_width":0,"_seopress_social_fb_img_height":0,"_seopress_social_twitter_title":"","_seopress_social_twitter_desc":"","_seopress_social_twitter_img":"","_seopress_social_twitter_img_attachment_id":0,"_seopress_social_twitter_img_width":0,"_seopress_social_twitter_img_height":0,"_seopress_redirections_value":"","_seopress_redirections_enabled":"","_seopress_redirections_enabled_regex":"","_seopress_redirections_logged_status":"both","_seopress_redirections_param":"","_seopress_redirections_type":301,"_seopress_analysis_target_kw":"","footnotes":""},"categories":[140],"tags":[210,212],"keyword":[988,1239,936,989,1026,865],"class_list":["post-99326","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-newswatch","tag-science-et-technique-en","tag-societe-en","keyword-intelligence-artificielle-en","keyword-automatisation-en","keyword-culture-ethique-en","keyword-jeux-en","keyword-risques-en","keyword-strategie-methode-en"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/posts\/99326","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/users\/25"}],"replies":[{"embeddable":true,"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/comments?post=99326"}],"version-history":[{"count":3,"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/posts\/99326\/revisions"}],"predecessor-version":[{"id":99331,"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/posts\/99326\/revisions\/99331"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/media\/96116"}],"wp:attachment":[{"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/media?parent=99326"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/categories?post=99326"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/tags?post=99326"},{"taxonomy":"keyword","embeddable":true,"href":"https:\/\/www.futuribles.com\/en\/wp-json\/wp\/v2\/keyword?post=99326"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}