Contents
  1. Chatbot or Google — what's the difference?
  2. How do I write a good prompt?
  3. Why does AI sometimes make things up?
  4. Who's actually behind it all?
  5. One subscription for all — bundling platforms like Mammouth
01

Chatbot or Google — what's the difference?

The most common mix-up: an AI chatbot is not a search engine. Google searches the web and returns links to pages where you find the answer yourself. A chatbot like ChatGPT, Claude or Gemini generates an answer — based on texts it was trained on. It doesn't search the web in real time (unless it has a dedicated search tool enabled).

That sounds like a technical detail, but it has major practical consequences:

🔍 Google Search
  • Finds existing pages on the web
  • Shows sources and links
  • Always current — knows today's events
  • Best for: News, prices, opening hours
  • Output: A list of links
🤖 AI Chatbot
  • Generates answers from training data
  • Often gives no sources
  • Has a knowledge cutoff date
  • Best for: Explanations, writing, summaries
  • Output: A formulated answer in dialogue

Stiftung Warentest puts it well: a search engine becomes an answer machine. (Source: Stiftung Warentest, October 2025)

💡
When to use Google, when to use a chatbot? For current events, prices or local information: Google. For explanations, writing tasks, ideas or complex questions that need back-and-forth: chatbot.

Worth noting: Google itself launched its own AI mode in Germany in October 2025 — a chatbot built directly into search that delivers answers instead of links. The line between search engine and chatbot is increasingly blurred. (Source: Stiftung Warentest, October 2025)

02

How do I write a good prompt?

A prompt is what you type into the chatbot — the question, task or instruction it responds to. Many people write something vague and wonder why the results are poor. The good news: better prompting is not rocket science. Four simple principles make the biggest difference.

1
Be specific — not vague
The less specific your request, the more generic the answer. Tell the AI exactly what you want, what it's for, and what format you'd like. (Source: FFHS, March 2025)
2
Give context — the AI doesn't know you
The chatbot knows nothing about you, your situation or your level of knowledge. The more relevant background you provide, the more useful the answer.
3
Specify the format
Do you want a list, a short text, a table, step-by-step instructions? Say so explicitly — otherwise the chatbot decides on its own.
4
Follow up — it's a dialogue
If the first answer isn't right, don't start over — refine within the same conversation. "That's too long, cut it to 3 sentences" is a completely normal next step. (Source: OMR, November 2025)

Before / After — what this looks like in practice

❌ Weak prompt
Write something about healthy eating.
✅ Good prompt
Give me 5 simple tips for eating more healthily day-to-day — for someone with little time who doesn't want to cook. Short sentences, no jargon.
❌ Weak prompt
Explain AI to me.
✅ Good prompt
Explain in 3 sentences what a language model is — as if you were talking to someone who has never heard of AI before.
💡
Tip: You can also ask the AI to help you formulate your request. Just write: "Help me turn my idea into a better prompt" — then paste in your rough idea.
03

Why does AI sometimes make things up?

AI chatbots "hallucinate" — meaning they invent information and present it as fact with full confidence. This can be wrong dates, fabricated sources, non-existent court rulings or simply false statements. The term comes from research and describes the fact that the model isn't lying — it has no consciousness — but produces statistically plausible text that happens to be wrong.

A real example: in 2023, a US lawyer filed court briefs containing citations from non-existent rulings that ChatGPT had made up. The court imposed a sanction. (Source: New York Times, May 2023)

⚠
Hallucinations are not a fixable bug. They are a structural feature of today's language models. All models hallucinate — some less than others. Always check AI output critically, especially facts, numbers, names and sources.

When is the risk particularly high?

A recent study found that about one third of AI chatbot responses on controversial current topics were incorrect — nearly double the rate from a year earlier. (Source: Stiftung Warentest, September 2025)

📌
Rule of thumb: Anything you produce with a chatbot and intend to use — check the facts. Especially for health, legal, financial or public-facing content. AI is a working tool, not a substitute for research.

More on this — with figures and a model comparison — in the AI Assistants Comparison article.

04

Who's actually behind it all?

Chatbots appear everywhere today — on bank websites, in shops, in apps, in customer service. What many people don't realise: most of them are not independent AI systems, but wrappers around a handful of large models. The chatbot on your mobile provider's website almost certainly runs on GPT from OpenAI, Claude from Anthropic or Gemini from Google — just with a different look and restricted to specific topics.

This is called API integration: companies buy access to a ready-made model and build their own interface around it. The underlying model stays invisible. The four most important providers globally are:

What this means for you: from a privacy standpoint, it matters which model is running in the background — not just which interface you see. If you share something personal with a chatbot on a website, that data goes to the actual model provider. More in the AI Assistants Comparison.

05

One subscription for all — bundling platforms like Mammouth

Anyone wanting to use multiple AI models faces a cost problem: ChatGPT Plus costs around €23 per month, Claude Pro around €18–20, Gemini Advanced around €22. Subscribe to all three and you're quickly paying over €60 a month.

An alternative is bundling platforms — services that don't develop their own AI model, but instead offer access to several major models through a single interface. The best-known example is Mammouth AI from France: from €10 per month, access to GPT, Claude, Gemini, Mistral, DeepSeek and more. (Source: mammouth.ai)

Mammouth AI
Example bundling platform

French platform, founded 2024. One subscription, many models. From €10 / month. (Pricing: mammouth.ai/pricing)

✓ Pros
  • Cheaper than individual subscriptions
  • All major models in one interface
  • Switch and compare models with one click
  • GDPR-compliant, European provider
  • No training on your data
✗ Cons
  • No native extras (no Artifacts, no custom GPTs)
  • Only 50 premium messages per 3 hours
  • No free trial available
  • Trustpilot: 2.6 / 5 (as of March 2026)
⚠
Who benefits from a bundling platform? Anyone curious to try or compare different models. If you work deeply with one model and need its specific features, a direct subscription is the better choice. (Source: blog-ia.com, January 2026)
📌 Editorial note — thx4data.de

AI chatbots are useful tools — but not oracles. Understanding how they work lets you use them sensibly: for drafts, explanations, ideas, summaries. Treating them like a search engine and accepting answers unchecked is a recipe for misinformation.

Particularly important: hallucinations are not a bug that can be optimised away — they are a property of the system. That doesn't make AI useless. It means critical thinking remains essential.

On privacy: what you tell a chatbot ends up with the provider of the model running in the background — not just the interface you see. For personal, health-related or legal matters: exercise caution.

Go deeper

Which model is best for what? Privacy compared? All key terms explained?

→ AI Assistants Compared