Understanding Artificial Intelligence: What AI Can Really Do – and What It Can't
Machine learning, deep learning, language models and Copilot: a clear guide to the technology, products, opportunities, limits and concerns in everyday work.
AI is not a single program. The term covers different methods and products. If you want to use AI sensibly, you should distinguish three things: How does the technology work? What can the specific product do? What responsibility stays with people?
Last edited: 27.10.2026. The product overview depends on timing. You can find more on SEO terms in the SEO & AI glossary.
What is artificial intelligence?
Artificial intelligence (AI) refers to computer systems that take over tasks for which we often use human abilities such as pattern recognition, language or planning. Not every AI generates text: forecasts, image recognition and recommendation systems also belong here.
The terms build on each other:
- Artificial intelligence: the umbrella term.
- Machine learning: learning from data, a branch of AI.
- Deep learning: learning with multi-layer neural networks, a branch of machine learning.
- Many modern language models (LLMs) are based on deep learning.
Simplified classification: Not all AI systems are machine learning; not all deep learning models are language models.
Machine learning: A system recognizes patterns in example data instead of only applying rules programmed in advance. Example: detecting unusual transactions.
Generative AI: A model creates new content such as text, images, music or code. Whether this content is factually correct must be checked separately.
How does a language model work?
A large language model (LLM) is trained on very large amounts of data. It processes inputs as so-called tokens and calculates which continuation fits the context. Through training and further tuning, it learns to follow instructions and handle a wide range of tasks.
The process in four steps:
- Training: learning patterns in data.
- Prompt: your question and the context.
- Model calculation: tokens are processed.
- Answer: an output that you can check.
Important: Fluent language is no guarantee of factual accuracy. When researching, the model can retrieve additional sources through search tools or RAG. Such a connection makes answers easier to trace, but it does not automatically prevent errors.
Rule of thumb: A convincingly worded sentence is not proof. For important decisions, check the original source.
Strengths and weaknesses
Where AI can help
- Structure, translate and summarize texts
- Create ideas, variants and drafts
- Explain and support program code
- Pre-structure large amounts of information
- Support recurring processes
Where caution is needed
- Invented facts or sources
- Unreliable results on complex tasks
- Bias and incomplete data
- Confidentiality and data protection
- Errors from agents that act on their own
Performance depends on the model, the task, the tools and the context. A system can be strong at programming and still be surprisingly error-prone at simple multi-step planning. The International AI Safety Report 2026 describes these uneven capabilities and reliability problems.
Which AI is in which product?
A product is not the same as a model. The application provides the interface, access rights, tools and data connections. In the background, it can use one or more models. The product (for example Copilot) picks a model (for example GPT or Claude), connects it to authorized data such as documents or search, and delivers a result: text, analysis or action.
| Product | Provider | Models / model families | Classification |
|---|---|---|---|
| ChatGPT | OpenAI | GPT model families; depending on plan and mode | A product is not identical to a single model. |
| Microsoft 365 Copilot | Microsoft | OpenAI GPT; in certain functions also Anthropic Claude | Model choice and access can be controlled by the organization and the license. |
| GitHub Copilot | GitHub / Microsoft | Models from several providers, including OpenAI, Anthropic and Google | The model list depends on the plan and the organization. |
| Claude | Anthropic | Claude model family | Standalone assistant and model family. |
| Gemini Apps | Gemini model family; partly automatic selection | Availability depends on plan and usage. | |
| Perplexity | Perplexity | Different models depending on product and mode | Search and answer product with source features; not a single base model. |
| Apple Intelligence | Apple | Apple's own models; integrations depending on the function | Note the dependence on device, region and function. |
As of: 27.10.2026. The table deliberately lists mostly model families rather than a supposedly fixed model version. Functions, plans and availability change. Details: Microsoft, GitHub, Google. Information on other products is a general classification and should be checked in detail before publication.
What fears do employees have?
Worries about AI are not just a communication problem. They concern work, recognition, control and personal development. Managers should take them seriously instead of offering blanket reassurance.
"Will my job be replaced?"
Individual tasks can be automated. Whether this leads to fewer jobs or different jobs depends on the industry, demand and company decisions. Recommendation: Discuss changes to tasks early and in concrete terms.
"Am I being monitored?"
AI can analyse performance data. This raises legitimate questions about transparency, data protection and co-determination. Recommendation: Set out the purposes of use and the limits in writing.
"Can I still keep up?"
New tools create pressure to learn and adapt. Recommendation: Provide training time, hands-on practice and contact persons.
"Who is liable for mistakes?"
AI outputs can be convincing and still wrong. Recommendation: Keep responsibility for approval and control with people.
The longer-term effects on employment are uncertain. The international safety report describes differing scientific assessments. Source: 2026 report, executive summary.
Why do media reports sound so contradictory?
Headlines often mix up harm happening today, forecasts for the next few years and hypothetical extreme events. In addition, "AI" can mean a simple recommendation system or a highly capable autonomous agent. Treating both under one term leads to misunderstandings.
| Category | Examples | Evidence |
|---|---|---|
| Observable today | Incorrect outputs, deception, data protection problems and misuse of AI-generated content. | Documented risk categories. |
| Developments with uncertainty | Changes in the labour market, greater autonomy of agents and new dependencies. | Some indications; consequences open. |
| Long-term scenarios | Loss of human control over future, highly capable systems. | Disputed scenarios, not a situation that exists today. |
Can AI get out of control?
Today's systems can already take unwanted actions and make errors that are hard to control, especially when agents have access to tools. A far-reaching loss of control in the sense of an existential threat, however, is a hypothetical future scenario. The International AI Safety Report 2026 states that current systems do not have the capabilities this would require; experts differ on how likely it is in the future.
Can AI think or be conscious?
Models can solve tasks, reason and produce language. Whether such abilities mean consciousness or subjective experience is an open scientific and philosophical question. Human-like language does not prove consciousness.
Questions to ask about every AI headline: What was actually observed? Is it about today's systems or future scenarios? Who made the statement? How solid is the evidence?
Source: International AI Safety Report 2026
How do companies use AI responsibly?
Four steps: define the task (benefits and limits), assess the risks (data and consequences), measure a pilot (quality, time, errors) and set the approval rules (human oversight).
- Data: Do not enter confidential or personal content without a verified legal basis and suitable safeguards.
- Quality: Check answers against original sources by spot checks, and fully in critical cases.
- Transparency: Disclose where AI is used and who makes the decisions.
- Law: Check the EU AI Act and other applicable rules for the specific use case.
- People: Involve employees, train them and collect feedback systematically.
Further reading: European Commission on the AI Act and NIST AI Risk Management Framework. This overview is not legal advice.
A sensible start in five steps
- Choose one specific task: for example, summarising internal documents.
- Define success: set processing time, error rate and satisfaction in advance.
- Check the data: Clarify which information may be used.
- Run a small pilot: With specialists, real examples and documented results.
- Decide: Continue, adjust or stop, based on the measured results.
The next decision: Does your company have a clearly defined task where AI could bring a measurable benefit without unacceptable risks?
AI glossary from A to Z
The most important terms for quick reference. Choose a starting letter.
Agent
An AI system that can carry out tasks in several steps using tools.
Algorithm
A fixed procedure for processing inputs.
Alignment
Aligning a system's behaviour with human goals and rules.
Artificial General Intelligence (AGI)
Hypothetical, broadly applicable AI with far-reaching capabilities; there is no universally accepted definition.
Benchmark
A standardised task used to compare systems.
Bias
Systematic distortion of data or results.
Computer vision
Machine processing and interpretation of images.
Deep learning
A branch of machine learning that uses multi-layered neural networks.
Embedding
A numerical representation of text or other content.
Fine-tuning
Additional training of an existing model for specific purposes.
Foundation model
A broadly trained base model for many downstream tasks.
Generative AI
Systems that create new content such as text, images or audio.
Grounding
Linking model answers to specific data or sources.
Hallucination
Model output that sounds plausible but is false or unsupported.
Inference
Applying a trained model to new inputs.
AI agent
Software that combines a model with planning and tool access.
Context window
The amount of information a model can take into account in one processing step.
Large language model (LLM)
A large language model that has learned patterns in language data.
Machine learning
Systems recognise patterns from example data.
Multimodal
Processing different types of data, such as text, images and sound.
Neural network
A computing structure with trainable connections for pattern recognition.
Parameter
Model values that are adjusted during training.
Prompt
An instruction or input given to an AI system.
Prompt injection
An attempt to make an AI system behave in unwanted ways through manipulated content.
RAG
Retrieval-augmented generation: generating answers using information retrieved beforehand.
Reasoning
A term for abilities or methods for working through problems in several steps; not proof of human-like thinking.
Reinforcement learning
Learning from feedback on actions or results.
Supervised learning
Learning from examples with known target values.
Token
A unit of text or data that a model processes.
Training
The process of adjusting a model's parameters using data.
Transformer
A network architecture on which many modern language models are based.
Unsupervised learning
Learning structures without predefined target values.
Validation
Checking quality and behavior with suitable tests.
Overfitting
The model learns the training examples too specifically and generalizes poorly.
Sources and editorial notes
Technical statements and time-sensitive product information should be checked against the manufacturer's documentation before publication and regularly afterwards. A source only supports the functions that it actually describes.
- Microsoft: Copilot and model selection
- Microsoft: Copilot Cowork models
- GitHub: Supported models
- Google: Models in Gemini Apps
- International AI Safety Report 2026
- EU: AI Act
- NIST: AI Risk Management Framework
- IBM: What is machine learning?
Editorial compilation as of: 27.10.2026. The graphics are simplified illustrations for teaching purposes. Model names and product functions can change at short notice.