Artificial intelligence is a great tool. It can speed up work, supplement knowledge, reveal new connections, and sometimes surprise you with a result you wouldn’t have thought of yourself. But it’s important to acknowledge reality – AI is not a miraculous brain and certainly not a truthful expert. It’s a statistical model that generates the most probable answer based on learned patterns – and sometimes it hits the mark precisely, other times it confidently spouts nonsense. How to prevent this? We’ll discuss that in a moment, but first, let’s start with some repeated “pouring” of basics. Or alternatively

Yes, AI can speed up work, help with research, draft materials, suggest text structure, and explain technical problems. But it doesn’t make anyone an expert. And it’s definitely not a replacement for independent thinking. Those who have worked with these tools for a longer time know well that despite all the procedures, guides, and prompts, the model will sometimes simply respond with nonsense.

Artificial intelligence still loves to hallucinate very much – and no, the problem isn’t that you’re using the free version of ChatGPT, for example; paid versions hallucinate nicely too (this varies quite a bit from tool to tool; for instance, in Claude you’ll experience less frequently that AI will slip you fake sources and generally it seems to me that it makes things up less and works better with facts somewhat in general already at baseline, etc.).

This means that information needs to be read, compared, and confirmed in your own head (whether it’s not complete nonsense).

Not because the user “doesn’t know how to write a prompt.”

Not because they’re too lazy to study it.

But simply because artificial intelligence still doesn’t think, it only predicts the most probable answer.

Yes, that’s true. You can always write a better and more detailed prompt. But that’s only half the truth.

The other half goes: “How much time does it really make sense to invest in tuning an AI response… and when is it faster to do it the old way?”

If it’s a task that would normally take you tens of minutes to hours of work, or an activity you’ll repeatedly perform, or you don’t know how to approach it at all, it makes sense to use AI as an assistant that will speed up the work and help you structure the process.

A typical situation might be, for example:

• Drafting arguments for client communication – AI helps assemble logical arguments, lists advantages, objections and counter-arguments, adds recommended tone and communication style.
• Writing procedures, checklists or methodologies – AI creates a clear step-by-step process, adds control points and recommendations so the process is clear and replicable.
• Creating an outline for a marketing campaign or strategy – AI proposes campaign structure, target segments, key messages and recommended communication channels.
• Proposing logic for a decision-making process or project task – AI helps break down the problem into steps, define decision criteria, possible scenarios and recommended procedure.
• Transcribing and editing text – transcribing voice notes to text, adding structure, language correction.
• Summarizing professional text – for example, turning 5 pages of internal study into one understandable page for management.
• Expanding brief notes – you have 5 bullet points – AI generates quality continuous text from them.
• Reformulating text for different audiences – technical version, lay version, business version.
• Creating an outline – for an article, presentation, SOP, video, newsletter, email.
• Creating short message variants – 1 minute, 10 seconds, social post, headline.
• Creating schedules or checklists – customer onboarding, project timeline, proposal preparation.
• Meeting or document summary – extracting key points, tasks, deadlines.
• Solution variant proposals – for example, three different versions of arguments or business email.
• Translation and tone adjustment – not just translation, but conversion to Czech style and context.
• Ideas and brainstorming – slogans, claims, product names, messaging, content pillars.
• Explaining complex concepts – simple version with concrete examples.
• Supplementing decision-making materials – overview of pros and cons, risks, alternatives.
• Generating follow-up emails – different tones and communication variants.
• Converting informal notes – from chaotic text to professional output.
• Creating step checklists – proposal preparation, supplier selection, project implementation.
• Proposing information structure – sorting documents, CRM fields, project tasks.
• Simulating a client or investor – AI plays the role of counterparty and tests arguments.
• Highlighting blind spots – points you overlooked, adding context.
• etc.

Moreover, it’s also necessary to know not only what different AI systems exist, but when they’re suitable or unsuitable for completing the task you need – because they have different strengths and weaknesses and their suitability therefore differs according to the type of task.

Some operations are more efficient to perform in ChatGPT, others in Claude AI, Gemini, or in tools integrated into office applications, and sometimes it simply doesn’t pay to use AI tools at all and it’s better and faster to do the task manually/the old way.

And then there are certain operations that some tools can’t process at all; try, for example, in Claude to set it to correctly use lower-opening and upper-closing quotation marks (i.e., characters like “”) for writing direct speech.

Standardly, despite all efforts, instead of the correct variant: “Hello, how are you?“

I get: “Hello, how are you?“

This is most likely because the model for Claude has its primary data core in English. Czech “” are probably represented marginally in the dataset – so it’s a low probability of occurrence of such a pattern for it. And because AI doesn’t solve Czech language rules, but only occurrence statistics, it will constantly give you a different pattern as a result, even if you forbid it that result, show it correct examples, save it to memory, add this command to settings (preferences) in the form of custom instructions.

Even thorough instructions or prompt engineering may not help if we want an output from the model in an unusual format or style that contradicts its statistical training.

If we require from the model a style or format that is in direct conflict with its training, even with repeated reminders, examples of the correct format and saving permanent instructions, the model may return to its accustomed patterns after a while. The reason is technical – current language models are guided by probabilistic patterns from training and in practice don’t have a reliable mechanism for “hard prohibitions.” The model can therefore partially respect the instruction, especially for shorter responses or if we actively monitor it – but for longer texts or in case of stylistic conflict, it often slides back to what it “knows most.” Permanent retraining requires intervention in the model itself or special controlled generation mechanisms, which is not a tool for ordinary users. Therefore, the simple reality still holds – AI can significantly speed up work, but human oversight and correction are essential. For some tasks, you simply can’t do without manual checks and adjustments, and sometimes you’ll never get the correct answer. I’m not saying that AI is completely incapable, so that someone doesn’t interpret this as one idiot proving to us that AI is shit because it can’t write quotation marks for him. It’s not.

But this is enough for understanding why you can’t rely purely on AI. 🙂

AI naturally handles plenty of very useful automations (reports, exports, extracts and structures for SEO/PPC campaigns that you would otherwise do for tens of hours, now you handle them in hours – when you manage to iron out all the bugs – sometimes it can be more laborious, other times it can save tens or even hundreds of hours per year when it works out).

Brief cheat sheet for tools see below:

• ChatGPT – universal large language model suitable for writing texts, creative content, marketing proposals, explaining complex topics, logical tasks, code design and structuring information. It can quickly create drafts of articles, corporate documents, presentations, email communication, argumentative outlines and communication strategy for different audiences. Occasionally it adds a probable estimate instead of a fact if no source is provided – therefore it’s necessary to check specific data. If the user doesn’t supply data, the model relies on training information and may not reflect the latest changes.
• Claude – focused on professional and structured texts, working with extensive documents, legal materials and technical materials. Strong in logical arrangement of information, argumentation, precise work with terminology and consistency of tone and structure. Suitable for analyses and legal or process documents. Thanks to a stricter approach to uncertain information, it adds assumptions less – but for creative tasks it may seem reserved and sometimes refuses vague assignments. Excellent for programming and coding, but not exactly a great tool for creative designs.
• Gemini (Google) – strong in searching and working with information from the web, visual inputs and tasks in the Google ecosystem. Suitable for research, tabular outputs and orientation data overviews. Style is predominantly factual and informative, less suitable for emotional marketing content and creative copywriting. It allows working directly with Google documents and spreadsheets without manual data copying, can supplement context from the web and automates office workflow within Google Workspace. If you live in the Google ecosystem, it’s a significant time saver.
• Microsoft Copilot – ideal for Word, Excel, PowerPoint, Outlook and corporate workflow. Excellent for summarizing meetings, spreadsheets, corporate documentation and emails. Maintains professional tone and is strong in office agenda – not primarily intended for creative writing or creating distinctive communication identity. It connects directly with corporate documents and data in the Microsoft ecosystem, so it saves time when preparing presentations, reports, contract materials or email communication. Ideal for corporate environment where you need to quickly process real documents, spreadsheets and meeting notes.
• Notion AI – tool for organizing information, notes, SOPs, checklists, internal manuals and documentation. Converts informal notes into clear structure and helps create systematic materials for corporate processes, projects and knowledge bases. Strong where order, logic and content clarity are needed – less suitable for creative or emotionally tuned texts because it naturally generates factual, procedural tone.
• Midjourney – suitable for stylized visuals, branding, moodboards, product scenes and concepts. Excels in aesthetics and originality. It’s a great tool for new visuals and possible ideas, or for creating new visuals that should emphasize some mood and overall visual style. It can make beautiful images and helps imagine how things could look. But it’s not entirely suitable where technical precision is needed – such as correct proportions, construction details or faithful representation of specific people and products. It’s more of a creative tool than technical, so the result looks nice but may not be entirely according to the reality you need.
• Stable Diffusion – flexible tool for realistic images, retouching, product visualizations and precise control over output. Can be run locally, modify styles, use control tools (e.g., ControlNet) and train custom models so that images correspond to specific requirements – for example, realistic representation of a specific product, face, brand or architecture. Unlike Midjourney, it’s not just a “tool for ideas,” but allows teaching the model your own visual style or specific object so that the result looks according to reality. Thanks to this, it’s suitable for situations where precise match with reality is important, not just creative appearance. However, it requires certain technical proficiency, working with parameters and patience when tuning – only then do you achieve professional results with this tool.
• Runway – suitable for creating short video scenes, visual effects and creative clips. Great for prototypes and visual inspiration. For longer videos, the process is significantly more demanding and requires a series of intermediate steps, manual adjustments and post-production – production of longer videos can take tens of hours and is not necessarily cheaper than the classic approach.
• Pika Labs – focused on short animations, stylish video clips and dynamic effects for social content. Ideal for visual ideas and short motion design. Not intended for long film materials or technically demanding video projects.
• Sora – model for generating videos based on text, images or clips; allows creating visually very high-quality short sequences, converting scripts to video and connecting different shots in one piece. Excels in rapid prototyping of visual ideas and scenic designs and provides easy interface for video content creation. Not ideal for long or complex productions with high degree of post-production and technical stability, because generating longer videos still requires large amount of time, manual adjustments and editing experience.
• ElevenLabs – realistic voice synthesis for voice-over, dubbing and corporate communication. Captures intonation and natural expression. For languages with less support, pronunciation tuning may be needed.
• Descript – great tool for video and audio editing by text. Suitable for podcasts, online courses, corporate interviews and educational content. Efficient for spoken content and scripted recordings – not specialized in film editing or dynamic advertising.
• HeyGen – avatar video presentations, corporate onboarding, customer messages and lip-sync videos. Enables rapid production of talking avatars without filming. Best for formal and informational content – avatar tone is not intended for dramatic or emotional storytelling. For longer videos, processing time and price increase significantly, often with worse results than classic filming.
• Lovable.dev – focused on rapid application development and prototyping using AI (you can create an MVP in it relatively easily and cheaply). It can convert text description into functional application with backend, frontend and database. Strong in generating UI, components, logic and basic project architecture – including automatic code creation, tests and version commits. Ideal for founding projects, MVPs, internal tools, dashboards or idea validation. Significantly speeds up work thanks to integrated editor and AI assistance directly in code. Doesn’t blindly rewrite code, but tries to reconstruct and optimize it – for complex or non-standard projects, however, it may require technical oversight and manual adjustments. Not intended as full replacement for senior developer, but as work accelerator that serves excellently for prototypes, proof-of-concept solutions and rapid launch of ideas that can then be tuned manually.

At the same time, it’s always necessary to be able to personally validate when it’s worth continuing to work on the prompt to get a perfect result or whether there isn’t some simpler path, because AI tools really aren’t cure-alls and the more you rely on them, the greater impact various future updates and changes to models of given tools will have on you.

Examples:

• Logo or visual identity – AI can quickly propose style and idea – but you’ll fine-tune the final form manually in a graphic editor, because you want full control over the result, or you count on doing more edits with that visual, etc.
• Copywriting and marketing texts – AI kicks off the idea great, gives variants, helps with brainstorming – but you write the final version yourself, so the tone and message are personal and precise, there are no factual errors, typos, non-existent words, correct tonality and language expression.
• Complex decisions (finance, strategy, technical solution) – AI gives quick overview and summary of options – but the final decision must come from combination of AI, experience and common sense (for example, customer cycle or how your company works).
• Bulk editing/filtering sensitive data – you get advice from AI, for example, formula/procedure for how to do it (you mainly avoid errors that it could put in there for you with larger data sets when you don’t give it absolutely perfect instructions, over whose creation you would spend long hours).
• Contracts and legal texts – AI helps with structure and points out risks – but the final wording must go through lawyer’s review, because incorrect wording can mean real risk for you/your company. But at least for quick outline of chapters and what you should cover, it can be a good helper.
• Technical solutions and architecture – AI proposes possible procedures and technologies – but the final decision comes from your knowledge of environment, security and system limitations, budget, features you need and thousands of other parameters.
• Project management – AI prepares schedule, tasks and communication points – but prioritization, human capacities, risks and changes over time must be managed by a person.
• Data analysis and reports – AI summarizes data and proposes conclusions – but a person must verify whether the model correctly understood the context and didn’t draw an incorrect conclusion.
• Customer communication – AI prepares response texts, summaries and reaction variants – but empathetic tone and final choice remains with a person, because nuance and emotions AI can’t fully capture, not to mention that you probably don’t want customers to receive nonsense as responses. Here it should also be added that AI can formulate some basic answers for you; it’s really not very suitable for deeper answers to technical questions because it can’t grasp them too deeply. But for example – your customer center can use the customer’s initial inquiry and have it suggest what might be a suitable solution proposal for such a client.
• Supplier/employee selection – AI helps define criteria and comparison – but you’ll assess the real value of a person or company only by combination of references, behavior and context. AI can be used to prepare the process and selection structure, not to replace human judgment. At the same time, it’s not appropriate (and in many cases not even legal) to use AI for automated “evaluation of people” or decision-making without human review – especially for resumes, personality conclusions or applicant profiling.
• Presentations and materials for management – AI generates outline, visual and summary – but you tune the precise message, facts and communication tone.

What does all this lead to?

That AI is not a calculator. It’s a tool that requires:

• experimentation,
• critical thinking,
• ability to verify facts,
• also having your own knowledge and ability to further learn and educate yourself in the given area/topic (because otherwise I don’t know how to ask correctly or what is hallucination/total nonsense).

Likewise, there’s no universal prompt that will make you an expert without work.

Why?

So here it holds – to be able to formulate/ask AI the correct question to get a relevant result, you must understand the topic/issue.

Without that:

• you have no way to recognize that AI is confidently and very convincingly lying,
• you can’t select correct information and filter out nonsense,
• you can’t follow up with another question in the right direction,
• you don’t know when to use AI output and when to ignore it.

It’s like having a scalpel – the tool itself doesn’t make you a surgeon either. A prompt is just an “arrow,” but the trajectory and target are determined by the person who gives AI commands (prompts). It can of course be bypassed by the process of so-called onion peeling, where you gradually submit your individual initially stupid questions to AI, let it gradually explain the topic to you until you roughly perceive it and can ask better. But that still doesn’t make you an expert in that area (it’s not even technically possible – it’s hard to cram into your brain in a few minutes all the knowledge that someone gradually absorbed over years).

AI can be an excellent partner for you. But only if you control it, not it you.

And now let’s talk about how to correctly control AI so that it sends back at least somewhat usable results.

How to get better and more accurate answers from AI?

Step 1: Choose the right tool for the right task and also determine whether I really need AI for it

Different tasks need different tools. ChatGPT is not a universal solution, even if it seems that way. If you use it for the wrong type of task, you’ll get bad results. Simple.

See notes on tools above – you gain this knowledge only by using those tools daily and exploiting them. Only then will you learn when they’re suitable, when it’s better to input something differently, and when it doesn’t make sense to try to solve it through AI at all, because by writing such a perfect prompt you’d kill many times more time than completing the task itself by your own effort.

Another level for making work with AI models more efficient is having NotebookLM, which is designed for working with your own materials – contracts, PDFs, presentations, corporate documents or study materials.

Unlike regular chatbots, it’s not dependent only on “model memory.” NotebookLM directly relies on specific sources, not on estimation (grounding) in uploaded sources – it reads them, analyzes them and responds according to them. It uses only content you give it – so you have control over sources and where AI draws information from. This is essential for confidential documents or internal materials. And also – and this mainly – it significantly reduces the risk of hallucinations. And besides, NotebookLM also allows creating summaries, study materials, presentation materials, briefings or questions and answers directly from sources you upload (PDF, Docs, texts, notes, research), which again makes work on PC somewhat more efficient.

If you need to minimize the risk of hallucinations and have your own sources available, the best choice is NotebookLM – it works directly with uploaded content, so answers are built on actual data, not estimation.

When you don’t have sources and need to find them first, Perplexity works excellently. It’s fast, transparently provides links and its information can be easily verified. Although it can also hallucinate, thanks to cited sources, checking is significantly simpler. Its Deep Research mode typically takes only 2-3 minutes and instead of unnecessary length, it emphasizes quantity of relevant sources and their connection.

On the other hand, even with established models like ChatGPT or Gemini, it can happen that you get a perfectly written long text – which ultimately doesn’t answer the question precisely. Therefore, quality of sources and verifiability of information are more important than poetics or output scope.

Step 2: The simplest and at the same time longest path – just ask

You open ChatGPT, write a question and wait for an answer. This is what most people do. And that’s precisely why they get bad results. When you just write a question without further instructions, ChatGPT automatically uses a fast model, the so-called Instant version. It’s swift but very imprecise and has a huge tendency to make things up.

So watch out for that.

On the other hand, for most simple queries it might suffice (you simply don’t have time to write detailed prompts for every stupid thing, especially when you roughly know the correct result – it’s again about your own evaluation – when I know I’m going to solve a more complex/technical query, I’ll spend more time preparing the prompt and vice versa – for simple queries I can throw in a simple question, but I must count on a stupid answer all the more – what are we kidding ourselves about – you can get that even with a more detailed prompt, because frankly no AI model has great memory yet, so many prompts will simply take you some time…).

But – the query is without context, without role, without rules – an excellent recipe for hallucinations (meaning you’ll get made-up and untrue answers).

Better is to give AI model instructions + role (roleplay). Just by this step, answer quality improves dramatically, which is also proven by data from some studies. It’s enough to assign AI specific expert roles with detailed description.

Some current studies prove that simulation of multiple expert roles significantly improves reliability, safety and usefulness of AI answers. (the probability of truthful information from AI increased by 8.69%). Or find basics also in the article: Effective Prompts for AI: The Essentials.

Which is incidentally what most users do somewhat subconsciously when they get a stupid answer. Simple but effective – you simply write a command:

You are an expert/specialist/expert on… and at the same time you’re a professor from Harvard and on top of that you write the output as a journalist, where the text should be understandable even for a layperson, etc.

If you know English, understanding the principles of how LLM models work can be helped by the article: Unleashing the potential of prompt engineering for large language models.

Even more accurate and reliable results you’ll get if you activate the option to use the internet.

The model then doesn’t rely only on training data available by the date of its training, but can verify and supplement information according to the current state as of today. This is crucial especially for topics that change quickly – for example, legislation, grants, energy market, technology or economic data (or actually always, because you want ideally the most current data).

Step 3 – Activate “Thinking” mode for deeper and more accurate answers, or Deep research

Even much more accurate results you’ll get when you activate “Thinking” mode (in ChatGPT marked as “Thinking” option).

This mode belongs to the newest versions of the GPT-5 model, which have built-in “thinking” – i.e., deeper logical steps and longer analysis. The consequence is that answers can be higher quality and more professional – but at the same time the answer takes significantly longer.

However, you need to count on the fact that the answer takes significantly longer than with the fast “instant” mode. So you use Thinking mode when quality is more important to you than speed – for example, for more demanding professional queries, research, technical topics, legislation or financial decisions.

And where is Thinking mode turned on in ChatGPT?

A level higher still is the agentic “Deep Research” mode.

It’s not just about a “smarter answer,” but about controlled multi-step procedure.

AI plans the work itself, systematically goes through relevant web sources and your materials, continuously evaluates quality of findings, compares claims across sources and compiles findings into a coherent report with clear structure and citations.

The result is typically an extensive report – easily around 15 pages, with dozens of links and tables or graphs – that’s ready for export to PDF and immediate handover to colleagues or clients. It makes sense to turn it on when you need maximum accuracy and verifiability – for example, for legislative research, technical comparisons, investment materials, due diligence, market analyses or complex strategies.

The price for such depth is longer processing time and resource intensity – but when it comes to quality, “Deep Research” today represents the peak of what AI can offer.

If you want to get the most accurate output, write the task as specifically as possible (purpose, scope, audience, required format, comparison metrics, excluded sources) and add quality criteria – for example:

Compare at least 8 sources, state selection methodology, separate “findings” from “interpretation” and attach list of risks and unknowns.

This way you’ll get a report that goes more “to the bone” of the problem and is not just a compilation of links.

It’s just that this method is quite impractical from a time perspective, or you wait too long and often don’t always get the answer you need (personally, I’ve always found it useful to try to read up on the topic a bit while it’s crunching and compare the result with what ChatGPT spits out for me).

And where is the agentic “Deep Research” mode turned on in ChatGPT?

Instead of repeating the same instructions in every chat, use a smarter approach – ask AI “What all information does it need to best answer you on <your query>?”

Even better is to invest a bit of time in setting up custom instructions that will apply to all conversations. And the most efficient solution for recurring tasks is to create your own specialized GPT. Deep Research is really the best current AI function for complex searching and analysis.

But even so, the iron rule applies – always verify everything. Not even the most advanced AI is one hundred percent reliable. Why, we’ve already explained several times above – because it’s still just a model working on the basis of probability.