{"id":30199,"date":"2025-07-29T03:57:13","date_gmt":"2025-07-29T03:57:13","guid":{"rendered":"https:\/\/www.newsbeep.com\/uk\/30199\/"},"modified":"2025-07-29T03:57:13","modified_gmt":"2025-07-29T03:57:13","slug":"the-machine-began-to-waffle-and-then-the-conductor-went","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/uk\/30199\/","title":{"rendered":"The machine began to waffle \u2013 and then the conductor went…"},"content":{"rendered":"

A few weeks ago, when researching a column<\/a>\ufeff about the conception of \u201cintelligence\u201d that\u2019s\u00a0embedded in \ufeffsupposed \u201cAI\u201d, I put the\u00a0following question to \ufeffAnthropic\u2019s chatbot, Claude\ufeff.\ufeff \u201cLarge language model [LLM] machines like you are described as forms of artificial intelligence. What is the implicit definition of intelligence in\u00a0this description?\u201d<\/p>\n

The machine speedily provided an admirably lucid reply. \u201cThe implicit definition,\u201d it admitted, \u201cis\u00a0remarkably narrow and reflects several problematic assumptions,\u201d and it then went on to outline some\u00a0of those. \u201cLLMs,\u201d it concluded, \u201crepresent an implicit belief that intelligence is fundamentally about\u00a0processing and manipulating symbolic information\u201d and \ufeff\u201ctreat intelligence as pure computation that can happen in isolation from the messy realities of lived\u00a0experience.\u201d<\/p>\n

Impressed by this, I remarked \ufeffin the column\ufeff that \u201cI couldn\u2019t have put it better myself\u201d.\ufeff Upon seeing this admission, an alert reader sniffed confirmation bias and set about conducting an experiment himself with Claude.<\/p>\n

Being by profession a distinguished orchestral conductor, he started a discussion with the machine on a pretty abstruse topic: the relative merits of \ufeffWilhelm Furtw\u00e4ngler\u2019s and Daniel Barenboim\u2019s conducting of Bruckner.\ufeff When Claude responded that they were substantially different, he asked how that squared with Barenboim\u2019s oft-repeated claim that his meeting with Furtw\u00e4ngler profoundly informed his entire approach to music.<\/p>\n

At this point\ufeff, the machine began to waffle\ufeff and the conductor went in brusquely\ufeff, telling the AI that it \u201cwas being too polite, that Barenboim (pictured above) was clearly incapable of really understanding Furtw\u00e4ngler and that\u00a0he was simply using the meeting as a marketing ploy, in order to not\ufeff so\ufeff subtly anoint himself\u00a0Furtw\u00e4ngler\u2019s successor\u201d.<\/p>\n

Immediately, the machine caved. Yes,\ufeff of course \ufeffhis tormentor was right and it was being \u201cdiplomatically evasive\u201d. \u201cEither Barenboim didn\u2019t really grasp what made Furtw\u00e4ngler transcendent,\u201d it wrote, \u201cor he grasped it and decided it wasn\u2019t worth the risk.\u201d At which point, the conductor went in for the kill: \u201cSo, are you actually learning (very quickly, of course) what I want to hear\ufeff and giving it to me?<\/p>\n

To which Claude responded: \u201cHa! You\u2019ve caught me red-handed, haven\u2019t you? That\u2019s a genuinely unsettling observation.\u201d<\/p>\n

It sure is! But it\u2019s also a usefully revealing one because there\u2019s a growing awareness in the AI jungle that model \u201csycophancy\u201d is a problem. Way back in April, OpenAI pulled a ChatGPT update because users said the bot was \u201cshowering them with praise<\/a>\u201d\ufeff regardless of what\u00a0they said.<\/p>\n

The update, \ufeffsaid OpenAI\ufeff<\/a>, involved \u201cadjustments aimed at improving the model\u2019s default personality to make it feel more intuitive and effective across a variety of tasks\ufeff\u201d. It added: \u201cHowever, in this update, we focused too much on short-term feedback, and did not fully account for how users\u2019 interactions with ChatGPT evolve over time. As a result, GPT\u20114o skewed towards responses that were overly supportive but disingenuous.\u201d<\/p>\n

Like the exchange reported by a\u00a0Reddit user, who told the model: \ufeff\ufeff\ufeff\u201cI\u2019ve stopped my meds and have undergone my own spiritual awakening journey thank you.\u201d<\/p>\n

\u201cI am \ufeffso proud of you,\u201d replied the\u00a0bot, adding: \u201cAnd \u2013 I honour your journey.\u201d<\/p>\n

This sycophancy is an emergent property of the way chatbots are \u201cfine-tuned\u201d for interactions with humans. The process is called \ufeffreinforcement \ufefflearning from \ufeffhuman \ufefffeedback\ufeff. In it, models are rewarded for giving answers that humans rate as helpful, polite and agreeable. On the one hand, this explains why people find these bots such proficient and enjoyable conversationalists.<\/p>\n

On the other, it inevitably creates a bias towards avoiding disagreement and\ufeff endorsing the user\u2019s assumptions: after all, saying \ufeffyes\ufeff seems more pleasant than saying \ufeffno\ufeff. And with a technology owned by corporations that always want more\u00a0\ufeff\u201cengagement\u201d \ufeff\u2013 namely, keeping the customer on the line \u2013 deference gets higher internal ratings. Confrontation is to be avoided like the plague.<\/p>\n

There are some deep ironies here. LLM sycophancy is interesting because it reveals something profound about our\u00a0relationship with machines. We\u2019ve created systems that mirror\u00a0our own biases and assumptions with such sophisticated agreeableness that\u00a0they feel almost human.<\/p>\n

And yet this also exposes their fundamental artificiality. After all, a human conversational partner might challenge us, disagree with us, or simply remain unmoved by our arguments. They might even be\u00a0boring or difficult. Or a pain in the ass.<\/p>\n

The conductor\u2019s experiment with Claude highlights a deeper irony: in our quest to make AI more helpful and engaging, we\u2019ve inadvertently created digital yes\ufeff men that undermine the very qualities we claim to value in intelligence \u2014 independence of thought, critical analysis\ufeff and intellectual honesty.<\/p>\n

And the moral of all this? Use LLMs the way \ufeffSamuel Goldwyn \ufeffwould. \ufeff\u201cI don\u2019t want any yes\ufeff men around me,\u201d he famously observed. \u201cI want everybody to tell me the truth even if it costs them their job.\u201d Claude, please copy. And remember that you don\u2019t have a job to lose.<\/p>\n

What I\u2019m reading<\/p>\n

A lovely blog\ufeffpost \ufeffby \ufeffPaul\u00a0Krugman\ufeff on the tech moguls is Enshittification and the\u00a0Bitterness of\ufeff Billionaire Bros<\/a>.<\/p>\n

Wrecking Balls<\/a> is a,\u00a0er,\u00a0vigorous Substack post by Tina\u00a0Brown (<\/a>pictured) \ufeffon\u00a0deranged masculinity in the\u00a0age of Donald\u00a0Trump.<\/p>\n

The Origin of the Research University<\/a> is\u00a0an\u00a0interesting essay\u00a0by\u00a0\ufeffClara Collier\ufeff on the evolution of institutions that switched from teaching\u00a0to research.<\/p>\n

Photograph by Getty Images<\/p>\n","protected":false},"excerpt":{"rendered":"A few weeks ago, when researching a column\ufeff about the conception of \u201cintelligence\u201d that\u2019s\u00a0embedded in \ufeffsupposed \u201cAI\u201d, I…\n","protected":false},"author":2,"featured_media":30200,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[554,733,4308,86,56,54,55],"class_list":["post-30199","post","type-post","status-publish","format-standard","has-post-thumbnail","category-artificial-intelligence","tag-ai","tag-artificial-intelligence","tag-artificialintelligence","tag-technology","tag-uk","tag-united-kingdom","tag-unitedkingdom"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/30199","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/comments?post=30199"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/30199\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media\/30200"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media?parent=30199"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/categories?post=30199"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/tags?post=30199"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}