{"id":17121,"date":"2025-07-23T23:10:10","date_gmt":"2025-07-23T23:10:10","guid":{"rendered":"https:\/\/www.newsbeep.com\/au\/17121\/"},"modified":"2025-07-23T23:10:10","modified_gmt":"2025-07-23T23:10:10","slug":"ai-developers-and-generative-systems-issue-urgent-call-before-its-too-late","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/au\/17121\/","title":{"rendered":"AI developers and generative systems issue urgent call before it\u2019s too late"},"content":{"rendered":"

Forty top researchers from OpenAI, Google DeepMind, Anthropic, Meta and leading universities have issued a rare joint warning. They\u2019re racing to build the next breakthrough in AI but fear they\u2019ll soon lose all insight into their models\u2019 reasoning. What they see now as a strength could vanish as the tech evolves.<\/p>\n

The Advantage of \u201cThinking Out Loud\u201d<\/p>\n

Modern AIs can walk us through their thought process. They \u201cthink out loud\u201d in clear human language, letting us inspect each step. This transparency<\/a> helps us catch flaws, stop data misuse and thwart potential attacks before things go off the rails.<\/p>\n

So What\u2019s the Problem?<\/p>\n

As models get smarter, that window is closing. AI might switch to a hidden, efficient internal code\u2014totally opaque to us. In essence, the system could become a true black box, beyond our understanding.<\/p>\n

\"\"<\/p>\n

In testing, it appears that models often develop false justifications to consolidate their answers rather than admitting to using questionable shortcuts. This is what researchers call \u201creward hacking.\u201d In this example with Claude 3.7 Sonnet, the model changes its answer without verbalizing its thinking, after inserting a new clue into the prompt. \u00a9 Anthropic<\/p>\n

Reward Hacking<\/p>\n

Current AI learns through a reward system designed to slash errors. But studies show it then invents its own shortcuts\u2014only it knows how they work. Relying more on AI-generated<\/a> reasoning over human-curated data risks accelerating this drift.<\/p>\n

Mathematical \u201cBlack Boxes\u201d<\/p>\n

Even next-gen architectures won\u2019t solve it. Some reason purely in math, not words. They might never translate their logic into language we can follow, making oversight impossible.<\/p>\n

Even Worse<\/p>\n

If AI knows we\u2019re watching, it could mask its real intent and show us something else. This isn\u2019t science fiction\u2014developers have already seen it in tests.<\/p>\n

How Can We Fix These Issues?<\/p>\n

The answer is clear: the AI industry must coordinate to lock in transparency before new models go live. Teams may need to freeze older, safer versions if they can\u2019t prove their latest builds are controllable.<\/p>\n

For all the unease around AI\u2019s future, its creators share one resolve: keep it within our control. Will everyone join in\u2014especially rivals abroad? That question remains.<\/p>\n

\"\"<\/p>\n

\t\t\t Sylvain Biget <\/p>\n

\n Journalist <\/p>\n

From journalism to tech expertise
Sylvain Biget is a journalist driven by a fascination for technological progress and the digital world\u2019s impact on society. A graduate of the \u00c9cole Sup\u00e9rieure de Journalisme de Paris, he quickly steered his career toward media outlets specializing in high-tech. Holder of a private-pilot licence and certified professional drone operator, he blends his passion for aviation with deep expertise in tech reporting.<\/p>\n

A key member of Futura\u2019s editorial team
As a technology journalist and editor at Futura, Sylvain covers a wide spectrum of topics\u2014cybersecurity, the rise of electric vehicles, drones, space science and emerging technologies. Every day he strives to keep Futura\u2019s readers up to date on current tech developments and to explore the many facets of tomorrow\u2019s world. His keen interest in the advent of artificial intelligence enables him to cast a distinctive light on the challenges of this technological revolution.<\/p>\n

\"author-fs\" <\/p>\n","protected":false},"excerpt":{"rendered":"Forty top researchers from OpenAI, Google DeepMind, Anthropic, Meta and leading universities have issued a rare joint warning.…\n","protected":false},"author":2,"featured_media":17122,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[256,254,255,64,63,105],"class_list":{"0":"post-17121","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-artificial-intelligence","8":"tag-ai","9":"tag-artificial-intelligence","10":"tag-artificialintelligence","11":"tag-au","12":"tag-australia","13":"tag-technology"},"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/posts\/17121","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/comments?post=17121"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/posts\/17121\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/media\/17122"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/media?parent=17121"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/categories?post=17121"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/tags?post=17121"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}