Not all AI thinks the same way. In fact, none of them really “think” at all. But they do simulate different styles of reasoning, deduction, induction, abduction, and that affects how they respond, what they’re good at, and when they fall apart.
So if you’ve ever wondered: “Which model should I be using for this task?” this episode is for you.
Hello and welcome back to AI SIX. In this episode, we’re going straight into the science of reasoning. Not human reasoning. AI reasoning. Because if you’re using AI tools, whether it’s ChatGPT, Claude, Gemini, or something else, and you’re not getting the results you want, it might not be the prompt. It might be the model.
WHAT IS AI REASONING?
Before we get into the models, let’s define the terms.
There are three main types of reasoning in logic and decision-making:
- Deductive reasoning — start with a general rule, apply it to a specific case. Like: All EU countries follow GDPR. Ireland is in the EU. Therefore Ireland follows GDPR.
- Inductive reasoning — look at a bunch of examples, spot a pattern, make a generalisation. Think: Every time I launch a campaign on LinkedIn, I get better engagement than on Twitter. So maybe my audience prefers LinkedIn.
- Abductive reasoning — this one’s about best guesses. You’ve got incomplete info, so you make the most likely inference. It’s what doctors do when they diagnose symptoms. Or what an LLM does when you ask it to explain why your sales dropped last quarter.
The best models can do all three, but not equally well, and not always reliably.
HOW THE BIG MODELS STACK UP
Let’s talk Large Language Models. I’ve put together a comparison table, and I’ll link it in the show notes, but I want to walk you through the highlights
🎯 ChatGPT (GPT-4o)
This is your all-rounder. Strong logic. Good at following structured prompts. Can generalise. And probably the best at abductive reasoning — those “what’s the most likely next thing?” moments.
Suggested prompt: “If all EU countries follow GDPR, and Ireland is in the EU, what does that mean for a Dublin-based startup?”
🎯 Claude 3 by Anthropic
This one’s the ethical advisor of the bunch. It’s brilliant at nuanced thinking, pulling together complex information, and avoiding overconfident nonsense.
Prompt to try: “You’re advising a charity using AI to analyse donor data. What ethical concerns should they consider?”
🎯 Gemini 1.5 from Google
Strong on multimedia. So if you’re working with video, audio, or images, this is worth testing. Reasoning is decent, but it leans heavily on what Google does best — fast answers with a factual tone.
Test it with: “Here’s a photo of a flooded street. What are three possible causes, and how would you verify them?”
🎯 LLaMA 3 by Meta
Open-source. Flexible. Popular with developers. Logic is okay, but use cases vary because it depends how it’s fine-tuned.
Try: “Here’s our quarterly sales data. Spot a trend and tell me what caused the dip in Q3.”
🎯 Grok by xAI
Built by Elon’s team. Let’s just say it’s… punchy. Not where I’d go for enterprise logic. But if you want a hot take for social, it’s fast and sharp.
Prompt: “Why are Gen Z ditching traditional news apps? Give me one line.”
🎯 Perplexity AI
This one’s different. It doesn’t rely on its internal training data alone — it pulls in real-time info from the web. So its reasoning is grounded in what it finds.
Prompt: “Summarise current EU AI rules from three reputable sources, and tell me what matters most for a university comms team.”
🎯 Microsoft Copilot
If you live in Excel, Outlook, or Word, this is your co-worker. Its reasoning is tuned for business logic, workflows, emails, and slide decks.
Prompt idea: “Read these five emails and draft a follow-up that anticipates the client’s next question.”
WHAT’S COMING WITH GPT-5?
Now let’s talk future.
GPT-5 is rumoured to drop soon. What’s expected? A few things:
- Better multi-step reasoning — it should follow your logic more clearly across long interactions.
- Stronger memory — meaning it can carry context over time, remember your style, your tone, your goals.
- Fewer hallucinations — more trust, less fact-fudging.
- Tool use as part of its core — not a bolt-on, but embedded. Imagine search, code, and document lookup all flowing into the answer seamlessly.
- And likely better RAG — Retrieval-Augmented Generation — pulling in your own data, not just what it was trained on.
In plain terms: GPT-5 will be less like a chatbot and more like a strategic assistant that can actually think with you, not just for you.
So where does this leave you?
If you’re building workflows, writing strategy, briefing leadership, or even teaching AI, don’t just think “What prompt should I use?” Think:
“Which model reasons the way I need it to?”
Pick the one that fits your context. Test your prompt. And never assume the answer is true just because it sounds smart.
That’s it for this week on AI SIX. If you want the full table with prompts and model breakdowns, check the show notes. And if you’re using different models in your day-to-day work, I want to hear about it.
About the Podcast Host
Joanne Sweeney is a specialist in public-sector marketing and digital communications. She’s the author of Public Sector Marketing Pro and the creator of the Professional Diploma in AI for Public Sector Marketing. With years of experience in training, consulting, and advocating for digital transformation, Joanne provides insight and guidance for government agencies navigating the digital landscape. She is also the host of the award-winning Public Sector Marketing Show podcast.
