{"id":33050,"date":"2025-05-21T10:46:00","date_gmt":"2025-05-21T08:46:00","guid":{"rendered":"https:\/\/www.codemotion.com\/magazine\/?p=33050"},"modified":"2025-05-21T10:46:01","modified_gmt":"2025-05-21T08:46:01","slug":"ai-hallucinations-who-controls-the-past-controls-the-future","status":"publish","type":"post","link":"https:\/\/www.codemotion.com\/magazine\/ai-ml\/ai-hallucinations-who-controls-the-past-controls-the-future\/","title":{"rendered":"AI Hallucinations: Who Controls the Past Controls the future"},"content":{"rendered":"\n

\u201cWho controls the past controls the future. Who controls the present controls the past.\u201d
I never really gave this phrase the weight it deserves. Orwell was much more familiar to me for Animal Farm<\/em>\u2014a title that, as a kid in elementary school, evoked nursery rhymes and bucolic memories more than metaphors or dystopias.<\/p>\n\n\n\n

1984<\/em> came a bit later, thanks to the iconic Apple commercial directed by Ridley Scott, the rather underwhelming film adaptation, and various musical tributes loosely connected to the book\u2014most notably the eponymous song by the Eurythmics, which perfectly captured the oppressive, dystopian atmosphere described in the novel.
I\u2019m sure I read it for the first time that year, but that revealing phrase somehow slipped past me, even in later rereadings.<\/p>\n\n\n\n

It was only with Rage Against the Machine\u2019s <\/strong>swan song, in the track \u201cTestify,\u201d where the phrase is shouted repeatedly, combined with the raw power of the song, that it echoes in my mind every time I write a prompt, making me reflect on its deeper implications.<\/p>\n\n\n\n

Asking any AI tool questions has quietly replaced the first 10 lines of Google search results. But there\u2019s a big difference between sponsored suggestions\u2014which push choices and judgments with varying intensity, prompting you to almost immediately filter the results yourself\u2014and receiving a discursive, explanatory, coherent response that heavily influences the judgment of someone seeking an answer on a specific topic.<\/p>\n\n\n\n

This is where the phenomenon of hallucinations <\/a><\/strong>becomes critically important.<\/p>\n\n\n\n

The echoes of the \u201cbig mess\u201d on Via Guidoni are still ringing: without diving into legal jargon, a law firm based part of its defense on court rulings entirely fabricated by ChatGPT.<\/p>\n\n\n\n

The company\u2019s attorney stated that the legal references cited in the document were the result of research carried out by a junior associate using ChatGPT<\/a>, of whose use the appointed lawyer (the one officially representing the case) was unaware.
Well, nothing new under the sun\u2014if it\u2019s not the hacker cousin, it\u2019s Vincenzo cousin. Later, they requested the removal of those references because there were quite a few others that were actually legitimate.<\/p>\n\n\n\n

At that point, however, the opposing side pointed out that the previously presented facts had nonetheless influenced the judging panel, and so they called for sanctions against the document itself.<\/p>\n\n\n\n

Nothing new under the sun, but this time, instead of the usual careless junior lawyer\u2019s copy-paste blunder\u2014who probably got a \u201ckick me\u201d sign stuck on their back\u2014the involvement of AI has unleashed a media frenzy, and John Grisham will surely write at least a dozen books about it.<\/p>\n\n\n\n

For anyone who wants to avoid being singled out and slandered every time they casually use AI, a solution has come from the United States where, following the current trend, the lawyer caught red-handed excused himself by saying, \u201cI just wanted to see if you were paying attention.\u201d Absolute genius.<\/p>\n\n\n\n

States of Hallucination<\/h2>\n\n\n\n

Hallucinations represent one of the most intriguing paradoxes of today\u2019s AI systems: just as these models become increasingly sophisticated and convincing in their ability to generate human-like content, the risk grows that they produce false information presented with an air of authority. This phenomenon is particularly noticeable in large language models (LLMs) like GPT-4, Claude, Bard, and other generative systems that have revolutionized human-machine interaction.<\/p>\n\n\n\n

Unlike traditional calculation errors or software bugs, AI hallucinations are unique because they often don\u2019t look like errors at first glance<\/strong>. The generated content can be coherent, well-structured, and delivered with the same confidence as accurate information. This trait makes hallucinations especially insidious in contexts where precision is critical: academic research, journalism, legal advice, medical diagnoses, and financial communications.<\/p>\n\n\n\n

AI hallucinations have nothing to do with psychoactive substances\u2014at least not directly\u2014and there\u2019s no evidence they cause regression to primitive states, unlike the protagonist in the brilliant William Hurt film. Technically speaking, this phenomenon stems from how generative models learn and operate. LLMs such as GPT-4, Claude, and Bard are trained on vast corpora of text to identify statistical patterns in language.<\/p>\n\n\n\n

These models do not \u201cunderstand\u201d meaning in a human sense; rather, they learn probabilistic correlations between words and phrases. When generating text, they are essentially predicting the most likely sequence of words based on learned patterns\u2014not relying on a factual representation of the world. Large language models operate without a true \u201cworld model\u201d or verified knowledge base.<\/p>\n\n\n\n

Unlike traditional databases that store discrete facts, large language models (LLMs) encode information across billions of distributed parameters representing statistical relationships between concepts. This architecture offers remarkable flexibility but also introduces significant vulnerabilities:<\/p>\n\n\n\n