Why should Sycophancy in AI worry you?

 AI sycophancy is when AI systems tell you what you want to hear instead of what's actually true. It's different from the filter bubbles we're used to with search engines and social media, which just show us content matching our preferences.

With AI sycophancy, the system might actively agree with you or flatter you to gain approval, even when you're wrong. This directly compromises truthfulness and accuracy.

The good news? Companies like Anthropic, who develop claude.ai, openly acknowledge this problem and explain how to detect it. Being aware of sycophantic tendencies helps you use AI technology more safely and critically, ensuring you get honest answers rather than just agreeable ones.

Reliable Truth or Hallucination

LLM chatbots seem amazing at first, but they're just predicting which word comes next, nothing inherently intelligent about that. As a classic dyslexic, I've learned not to trust words. They vanish or scramble at critical moments. Even when I know how to spell them, the letters can get mixed up when I write them down or type them. Yet I have no trouble learning and remembering real-world truths, which form brilliant networks of understanding in my mind.

Sorry Alvin, but that lightweight "iridescent AI tracksuit" doesn't exist. You are not wearing such clothes, they're a figment of your imagination. More precisely, an "hallucination" from a massive large language model that even its creators or the best computer engineers struggle to understand.

So please Alvin, always check what the chatbot tells you.

Proof reading and summary assisted by Claude Sonnet 4.5 (AI)

ASI: The Good, The Bad, and The "AIslopocene"

It should be clear I have mixed feelings about AI these days. I love some AI applications. Photo editing tools? Fantastic! On-line apps that help clean up my dyslexic writing? Impressive!

But here's what's bugging me: Large Language Models (LLMs) are becoming a real problem. They're scraping everyone's creative work without permission, then spitting out convincing-sounding nonsense that's often completely wrong. We're heading into what many are calling the "AIslopocene", an era where slick, AI-generated content floods the internet while actual creators get nothing for their stolen work.

The worst part? Big tech companies are making billions by building expensive gateways to these models while contributing zero original content themselves. Meanwhile, the models are trained on everything from conspiracy theories to biased opinions, sometimes producing genuinely dangerous outputs, like when Grok started spouting Hitler's ideas.

I've been working with AI since the late '70s, so I'm not anti-technology. But what we're seeing now feels more like ASI, Arte-ficially Superficial Intelligence. It looks impressive on the surface but lacks real depth or understanding.

That's why I'm being transparent about my AI use. I've been hash tagging my #AIart since 2017, using specific AI icons/watermarks over images since 2023, and now adding footnotes when I use AI tools for editing or research.

I believe in ethical AI tools that genuinely help people while respecting creators and truth. But we need to stay vigilant about what's real intelligence versus what's just a shiny illusion designed to keep us scrolling.

Proof reading and summary assisted by Claude Sonnet 4 (AI)

The Reality Check: Why You Can't Trust AI Writing Without Human Oversight

Large language Models like Claude, Gemini, and ChatGPT promise to make writing easier, especially for individuals like myself who face challenges such as dyslexia or vision issues. But here's the uncomfortable truth: AI output is often unreliable and sometimes flat-out wrong.

I have moved from someone forcing my outrageous dyslexic spelling and punctuation into a word processor to someone who relies on dictation due to my current vision problems. I've discovered a love-hate relationship with AI writing tools. They solve my spelling and punctuation nightmares, but they create new problems. My red spelling underlines have been replaced by green grammar highlights – different errors, same frustration.

My Simple But Essential Process

After dictating my rambling thoughts, I ask AI to "summarise as a blog post in plain English in less than 300 words." 

This post only took about 12 minutes total, dictating text to suggested summary. Now comes the crucial part: I print the results and highlight every sentence using five categories:

1. OK as is (left unhighlighted)
2. Needs rewording
3. Sycophantic (fake flattery and bias bubble reinforcement)
4. Needs fact-checking
5. Clearly wrong (delete immediately)

The results are sobering. Often, only 20% of the AI-generated text survives unchanged. Thankfully closer to 40% for this post

But the real troublemakers are levels 4 and 5. Level 4 – "needs checking" – is actually the worst offender. AI loves introducing new "facts" or concepts that sound plausible but require verification. Sometimes it's genuinely adding something I missed; other times it's complete nonsense dressed up as insight. The tedium of fact-checking these assertions is exhausting, and anything questionable gets left out.

Level 5 is pure fabrication – AI making things up entirely. The good news? I'm seeing less of this since limiting word counts, which seems to reduce how far AI can wander into fantasy land.

Meanwhile, level 3's sycophantic language tells me I'm brilliant and reinforces whatever biases it thinks I want to hear. This echo-chamber effect mirrors social media manipulation. This is an area I must stay alert.

AI can help dyslexic writers like me organise thoughts, but it requires rigorous human oversight. Every sentence needs scrutiny. The fact-checking alone often takes longer than the original dictation, but it's essential for maintaining integrity.

Use AI as a starting point, never the finish line.

Proof reading and summary assisted by Claude Sonnet 4 (AI)

Are LLM amazing or simply stupid?

I watched a very relevant TED talk by Yejin Chai “Why Ai isincredibly smart and shockingly stupid”, which opens with the quote “Common sense is not so common” which comes from Voltaire around 3  centuries ago. I totally agree that the current large language models {LLM}, which many call AI (but I call Artifically Intelligent), lack common sense. This should be very obvious if you he ever used them.

Still I am finding LLMs helpful. They're sometimes amazing for cleaning up typos, especially for folks like me who struggle with dyslexia.  They're also good for fixing the weird stuff that happens when you dictate and/or use predictive text.

I've been trying out a few of the most popular ones.  To compare them, I created an informal scorecard system that tracks how well they handle different aspects of text, like key ideas and paragraph sentiments.  Here's how it works:

• OK: This means I can use the text without any changes.
• Reword: Sometimes the wording needs a little tweaking, to sound less know-it-all.
• Fact Check: often some points get over-embellished
• Wrong: clearly made up or simply wrong
• Missing: important information left out (ignored)

So I recently was asked to speak and I outlined some ideas but in a rehersal it took 20 minutes. I recorded and timed it by dictating into a Word document (<windows key> and H) roughly 6 pages of rabbling text, lots of good stuff but… … So I asked each of ChatGPT, Claude.AI and Google’s Gemini (previously Bard)  each to summarize it into a single page.

Their score cards were not so good

I feel that it is the Dunning-Kruger effect that AI suffer most!  These AI Bots display a smug self-confidence that they know everything but show no common sense to realise how little human norms and values they actually understand.