AI myths explained.
Less hype. More understanding.

A confident answer can still contain invented facts or citations.

AI affects tasks within jobs differently. The ILO finds job transformation more likely than complete replacement; this is not a guarantee for any individual role.

AI also supports image recognition, recommendations, spam filtering and navigation. Generative AI is one part of a wider field.

Fluent language is not evidence of human feelings or self-awareness. Human-like wording can encourage misplaced trust.

Training data and system design can reproduce harmful biases. Automation does not make a decision fair.

Clear instructions help, but do not eliminate errors or missing context.

Oversight should match the stakes and context. A useful draft is not an authorised decision.

Privacy depends on the service, settings, contract and data handling. Do not assume confidentiality.

Synthetic media can look convincing. Appearance alone cannot establish authenticity.

Reliability depends on the task, data and deployment context. A demonstration is not operational validation.

Model knowledge and web retrieval are different. Fresh information requires an enabled search or data connection, and a particular answer may not use it.

References can be invented, misquoted or irrelevant. A real source may not support the sentence attached to it.

Generated explanations can sound logical while supporting an incorrect answer.

Detection tools have limitations. Their result alone cannot establish who created a text, image or video.

Generation settings and model behaviour can produce different outputs. Lower randomness does not universally guarantee identical results.

A model can learn noise or unhelpful patterns. Results on genuinely unseen examples matter more than data volume alone.

An evaluation dataset may fail to represent the situations encountered after deployment.

Retrieval can surface the wrong passage, and generation can misinterpret it. Access to documents is not a correctness guarantee.

Malicious instructions in user input or retrieved content can still target a system. Prompt wording alone is insufficient protection.

Existing foundation models can support many tasks through prompting, retrieval or adaptation. Custom training is one option, not an automatic requirement.

Capabilities vary by model and configuration. A system intended for one input or task may not support another.

Drafting faster may add checking, corrections or integration work. Net benefit depends on the whole workflow.

Recognising unreliable outputs and understanding their consequences still require relevant knowledge and judgment.

Length and polished wording do not establish accuracy or relevance. More text can also contain more unsupported claims.

Data and operating conditions can change, affecting how well a model works. Evaluation is an ongoing responsibility.