AI assistants now sit in the middle of homework, job applications, medical questions and even conversations about personal stress. That ubiquity has shifted the debate among researchers from whether people should use AI to how they should use it. Sasha Luccioni, a Canadian computer scientist who leads work on AI and climate at Hugging Face, argues that AI can extend what people are capable of, provided the person stays in control. In a segment for the BBC World Service, she proposed four questions to ask before turning to an AI tool.
Is this the right tool for the job?
No single model is built for every task. Some systems are tuned for scientific or technical reasoning, others for writing, education or creative work, and general-purpose chatbots are not necessarily the best at any one of these. Luccioni’s point is that choosing a tool because it is the most famous, rather than because it fits the task, is a common and avoidable mistake. The same logic applies to features within a tool: a model that is good at summarizing a document may be a poor choice for arithmetic or current events.
Can I trust the answer?
Language models produce fluent, confident text whether or not the content is correct. The industry term for a plausible but false output is a hallucination, and every major provider acknowledges the problem in its documentation. Luccioni’s advice is to treat every answer as a draft to be checked rather than a verdict, especially when the stakes are high. Practical habits include asking for sources and then confirming them independently, cross-checking numbers, and being more skeptical of answers that are unusually specific on topics where the model has no obvious way to know.
Is my data safe?
What happens to the information typed into a chatbot depends on the provider’s policy, which varies widely and changes often. Some services use conversations to train future models unless the user opts out; others retain chats for a period for safety review; business tiers generally offer stronger guarantees than consumer ones. Luccioni’s recommendation is to keep personal, sensitive or potentially embarrassing information out of AI tools altogether, on the assumption that anything entered could be stored, analyzed or, in the event of a breach or error, exposed.
Do I need AI for this at all?
The fourth question is the one Luccioni treats as most important. Using AI for every task risks eroding the skills it replaces: thinking through a problem, writing a first draft, making a judgment call. Decisions that turn on ethics, personal values or relationships are not well suited to delegation, she argues, however capable the model.
There is also a resource cost. Training and running large models consumes significant electricity and, through data center cooling, water. Luccioni has published research measuring the energy use of AI models and has repeatedly argued that routine tasks which do not need a large model should not be sent to one. The International Energy Agency projected in 2025 that data center electricity demand could roughly double by 2030, with AI the main driver, which makes the question of necessity an environmental one as well as a personal one.
A tool, not a default
The thread connecting the four questions is deliberate use. AI is a powerful instrument when it is matched to the task, checked for accuracy, fed only data the user is comfortable sharing, and reserved for situations where it actually adds something. Reaching for it by reflex, Luccioni suggests, offers convenience at the cost of the capacities that make the work worth doing in the first place.
