INPUT, OUTPUT AND HUMAN RESPONSABILITY IN AI SYSTEMS

Using a generative artificial intelligence system does not simply mean asking a question and receiving an answer. The person defines the problem, builds the input, provides the context, checks the output, and decides what to do with it. AI can execute part of the work, but design, supervision, and responsibility remain human.

Using a generative artificial intelligence system does not simply mean asking a question and receiving an answer. Behind the apparent simplicity of the interaction lies a sequence of activities that remains largely in human hands: defining what to ask, building the input, providing the necessary context, evaluating the output, and deciding what to do with it. For this reason, the user is not merely the recipient of an automatically produced result. They become the subject who designs the task, organizes informational resources, checks the quality of the work performed by AI, and makes the final decision.

A generative system produces an output starting from an input. This input can be a question, an instruction, a document, an image, a series of data, an information base, or a combination of these elements. The first essential point is that the input does not arise autonomously from the system: someone must establish what the problem is, what information is relevant, and what result is to be achieved. In other words, even before AI starts working, there is a phase of human design.

The quality of the input does not depend solely on the grammatical correctness of the prompt. It depends above all on the quality of the question and the amount of context provided. Asking generically to prepare a training plan produces one type of response; indicating the recipients, initial skills, training objectives, duration, budget, teaching methods, and organizational constraints produces a very different problem. The AI's response is thus bounded by what the person decides to include in or exclude from the request.

For this reason, speaking only of a 'prompt' can be reductive. The input also includes the informational environment in which the system is made to operate. If AI has access to documents, databases, updated information, or other relevant sources, its work can be more informed. If instead the system receives little data, incomplete data, or irrelevant information, the quality of the result can also degrade. Designing the input therefore means designing, at least in part, the environment in which AI must perform its task.

Not all questions are equivalent. A question guides the space of possible answers and can implicitly introduce constraints, assumptions, or preferences. If one asks exclusively to prepare an eight-hour course on artificial intelligence, the system will work within that perimeter and will not necessarily evaluate whether duration, level, and contents are appropriate for the recipients. If instead one asks which training plan is most suitable for a group with specific initial skills, specifying objectives, available time, budget, and expected results, the system has a more complete decision-making framework. A poorly formulated question can thus lead to a response perfectly consistent with the request but of little use regarding the real problem.

Knowing how to use AI therefore means knowing how to transform a generic need into an operational request. It is necessary to clarify the purpose, specify constraints, provide information that modifies the response, optionally indicate the desired format, and understand which aspects to leave open to the model's exploration. This capability does not coincide with simply knowing how to write well: it is a form of task design. Whoever formulates the request must understand the problem well enough to know which elements are decisive and which, instead, are secondary.

From this perspective, prompt engineering is not just a technique for obtaining better sentences. It is the way in which the interaction between person and system is built. The more important the task, the more necessary it is to carefully define the problem before delegating it. Effective delegation therefore begins before execution: it begins at the moment it is established what must be done and with what information. 
To address this need, Geeks Academy offers Prompt Master courses.

After receiving the input, AI produces an output. The output can be a text, a table, an image, a forecast, an operational plan, a classification, a summary, or another form of result. But the presence of a formally well-constructed result does not guarantee that it is correct. A system can provide plausible but inaccurate information, use outdated data, misinterpret a constraint, or produce a conclusion that is not sufficiently supported.

For this reason, the second major skill required of those who use AI is the ability to check the response. One must ask whether the result is consistent with the question, whether it respects constraints, whether it correctly uses available data, whether it contains statements that must be verified, and whether, in the case of specialized activities, it is compatible with the rules and knowledge of the domain. AI can execute part of the work, but quality control does not complete itself automatically.

The problem becomes particularly important when the output concerns decisions that have economic, legal, technical, scientific, or personal consequences. In these cases, it is not sufficient to accept the response because it is formulated in a convincing manner. The person must be able to distinguish between a plausible response and a correct response. If they do not directly possess the necessary competence, they must set up a different check: consult an authoritative source, compare multiple sources, ask an expert for verification, or use validation procedures appropriate to the context.

When it is said that AI 'does something for me,' there is a risk of confusing two different levels. The first is execution: the system can write, reorder data, search for relationships, propose alternatives, produce a document, or perform other activities that previously required human time. The second level is the judgment on the result: establishing whether what has been produced is correct, adequate, and usable. This second level is not automatically transferred to AI and requires, especially in specialized tasks, competence in the domain to which the output belongs.

Responsibility thus remains tied to whoever uses the result. Saying that a module, an analysis, or a preliminary decision was prepared by an AI system does not eliminate the need to verify them. To truly evaluate the output, it is not enough to observe its form or plausibility: one must possess, directly or through a competent person, the domain knowledge necessary to understand its content, recognize any errors, and judge whether the conclusions are well-founded. The person using AI must therefore know where the system can fail, which elements must be checked, and what consequences an error can have. The more critical the result, the stronger the control must be.

This profoundly modifies the meaning of delegation. It is not simply about transferring work to a machine. It is about building a process in which the machine executes certain activities and the person retains the supervision capacity. Such supervision also includes the ability to explain the output: it is not enough to declare that an answer is correct or wrong, one must know how to motivate the judgment, clarify what the result means, and illustrate its presuppositions, limits, and consequences. In order to be able to explain an output, one must first have understood it; and to evaluate it reliably, one must understand it in light of domain competencies. The sequence thus becomes: understand the output, evaluate it, know how to explain it, and finally decide whether and how to use it. Effectiveness depends not only on how much work AI manages to perform, but also on how well the control over the delegated activity is designed.

It is precisely in this context that AI Literacy plays a fundamental role: developing the ability to understand how artificial intelligence systems work, recognise their potential and limitations, critically evaluate their outputs, and use them consciously and responsibly. Knowing how to use an AI tool is not enough; it is essential to acquire the skills needed to manage its use, verify its outputs, and retain control over decisions. AI Literacy thus becomes essential to turning automation into a genuine advantage for both professionals and organisations. The AI Literacy course is delivered by our lecturer and Geeks Academy Scientific Committee member, Stefano Ridolfi, who contributed to the writing of this article.

Even when the output has been verified and is correct, a final step remains: deciding what to do with it. An answer may be technically right but not be the best choice at that moment. A training plan may be well structured, contents may be appropriate, and time and costs may have been correctly estimated, but the manager may still decide to modify it, postpone it, or not adopt it. Similarly, a business analysis may be correct but not automatically translate into an operational decision.

This distinction is fundamental. AI can contribute to producing information, analysis, and alternatives, but the transformation of output into action remains a decision-making act. Circumstances may have changed, priorities not present in the input may exist, or new elements may emerge. The final decision therefore requires an evaluation that integrates the AI's result with the real context in which the person or organization is operating.

The role of those who use artificial intelligence can be described less and less as that of a simple operator and more and more as that of a resource manager. A manager does not necessarily perform every activity personally: they establish objectives, assign tasks, provide resources, monitor results, and decide how to use them. AI introduces a similar dynamic. The person must choose what to delegate, what to continue doing directly, and where to concentrate their time.

At the same time, the user assumes the role of a designer. They must design the task, decide what questions to ask, build the context, choose the data to provide, and define the criteria by which to judge the response. The interaction designer does not limit themselves to 'writing a prompt': they build the conditions in which the system must work. In more structured applications, this can mean preparing documents, databases, procedures, information access rules, and sequences of controls.

Finally, the person remains a supervisor. They must observe the system's behavior, identify errors, recognize when the response is insufficient, ask for a new processing when necessary, and establish whether the output can be used. Manager, designer, and supervisor are thus not three separate professions, but three dimensions of the same way of using AI consciously.

Table 1: The three dimensions of the human role in AI use

The interaction can be summarized as a chain: the person defines the problem, builds the input, AI processes and produces an output, the person verifies the result, and finally decides how to use it. Automation intervenes primarily in the processing phase, but does not eliminate the other phases. On the contrary, the more the system's capacity to rapidly produce contents and solutions grows, the more important the quality of the input, the control competence, and the decision-making capacity become.

Problem Definition -> Input Construction -> AI Processing -> Output Verification -> Final Decision

The value of AI therefore does not consist solely in doing faster what a person did before. It also consists in allowing activities to be performed that previously would have been too costly, slow, or complex. But this greater capacity also increases the need for governance of the process. The effective user is not the one who renounces their role by blindly trusting the system; it is the one who uses AI as a resource to be directed, verified, and integrated into their own decisions.

The central competence thus becomes twofold: knowing how to ask and knowing how to check. Knowing how to ask means correctly designing the problem entrusted to AI. Knowing how to check means being able to judge what comes back. Automated execution is placed between these two moments; after them, the decision remains. It is in this structure that the human role changes: less direct execution of certain activities, more design, supervision, responsibility, and choice.