Based on my daily experience, I have developed a set of ten recommendations—which I propose calling the “Slow Ai International Monitor” (Saim)—for the purpose of examining current international events while making the best possible use of Artificial Intelligence.
Whenever a new piece of news — or even simply a rumor — emerges, I ask AI to analyze it by applying, as far as possible, the ten recommendations that follow.
- AI should not be used hastily. It is not a shortcut, nor is it a substitute for the researcher. Rather, it should be regarded as a “technological prosthesis” capable of enhancing our ability to verify, falsify, contextualize, and investigate a news story in greater depth.
- This protocol is still being tested and refined, but the results produced by the platforms on which I have experimented with it appear decidedly encouraging.
AI Should Not Be Used as an Autopilot. We should approach AI as a tool in our own hands—much as we would navigate a boat on the open sea without autopilot—rather than as a machine to which we delegate the task, or worse, one that we simply follow and become dependent upon.
- When navigating the vast sea of international news, there is always a risk that we cannot know with certainty whether information is true or false, relevant or insignificant, or whether important background circumstances or implications exist.
- Today, however, there is an important new development that can help researchers navigate this labyrinth. Through interaction with users, the best AI platforms are capable of developing a methodology and personalized memory tailored to an analyst’s specific needs.
- The results achieved arise from the convergence of three factors: the specific expertise of human beings, the models’ pre-existing knowledge, and information obtained through online research.
- It is important to remember, however, that personalized memory develops from our interactions with these platforms, and there is always a possibility that it may reinforce our own point of view at the expense of critical thinking.
The Quality of AI Answers Depends on the Quality of the Questions. The quality of research results depends to a great extent on the quality of the questions being asked. It is therefore advisable—particularly during the initial stage—to provide AI with arguments, texts, questions, and research projects that are as detailed as possible, avoiding generic questions that lack context.
- It is also necessary to define as precisely as possible the relevant time period, geographic area, subject of the research, and its operational definition.
- For example, it is essential to distinguish between the date an event occurred and the date it first appeared on the web—an error I have observed with considerable frequency.
- The greater the clarity and precision of the questions, the less likely the answers are to be vague, generic, or redundant.
- There is another caveat to consider: AI platforms tend to please the user. This creates a risk that some responses, rather than attempting to falsify a proposition, will uncritically confirm the user’s point of view regardless of its reliability.
- Even highly detailed research questions can cause AI to reflect the user’s existing beliefs and preconceptions rather than challenge them.
Don’t Just Let the Platforms Assist You—Help the AI. AI’s initial answers are often incomplete, superficial, or incorrect. It is therefore necessary to correct errors patiently so that they are not continually repeated.
- Whenever possible, it is also extremely important to progressively supplement the research with additional information. PDFs, images, videos, audio, data, and documents may contain information the AI has failed to find because it is hidden in some obscure corner of the web.
- To a greater or lesser degree, all platforms still tend to confuse how widely a news story has circulated—and how frequently it appears online—with how reliable it actually is. Hostile actors, incidentally, exploit this inherent weakness when conducting disinformation campaigns.
- Models frequently overlook information that is true and highly relevant simply because it has not circulated widely.
- Whenever possible and appropriate, it can therefore be very useful to provide information derived from direct personal experience, provided doing so does not create risks for other people or violate security protocols.
- Before submitting material, however, its authenticity must be verified, because the danger of introducing misinformation exists at both the input and output stages—as the familiar saying goes: “trash in, trash out.”
- Finally, we must be aware that platforms may use the information we provide to train their models.
Pay Attention to Original Languages. It is imperative that AI conduct research using sources, media, and social media in their original languages, and only afterward translate them. This helps avoid the cultural distortions that can arise when automatic translation occurs before the research itself.
- Few people are also aware that the linguistic approaches of different platforms vary considerably. To cite just two particularly striking examples, while the major Chinese AI platforms promote the widest possible use of Mandarin—with the resulting marginalization of Cantonese and other local languages—Indian platforms emphasize, as the specialized literature confirms, the extensive linguistic and cultural diversity that characterizes the world’s largest democracy.
- It should also be remembered that the quality of linguistic understanding is not uniform even within a single language, particularly when dealing with dialects, different registers, or technical and military jargon.
- The ability of an AI system to communicate clearly and accurately should therefore always be tested in the specific context rather than simply assumed.
Stress-Test the Platforms. During interactions with AI platforms, ask them to provide alternative formulations of their answers and openly challenge those answers with counterintuitive examples.
- For example, ask for interpretations that are the opposite of those initially provided by the platform, and request scenarios that have a low probability of occurring but could have a high impact.
- These two forms of falsification—challenging the platform’s conclusions and testing the analyst’s own hypotheses—are both necessary and are not interchangeable. It should be noted that this kind of stress-testing requires time that may not always be available during emergencies.
- In crisis situations, users of the protocol should ask AI to shorten the stress-testing cycle rather than silently skip it, and conclusions reached under the pressure of events should be explicitly identified as provisional.
Trick Questions and the Search for Errors of Omission. Within stress-testing, three areas deserve particular attention.
- The first is challenging AI with trick questions containing false premises, in order to determine whether the model recognizes and challenges the incorrect assumption rather than automatically incorporating it into its answer.
- The second is explicitly asking what is missing from the analysis, because AI produces not only errors of interpretation but also frequent and serious errors of omission.
- Finally, it is imperative to systematically ask the model to distinguish between factual phenomena and conjecture, with conjectures identified according to varying degrees of probability.
- When the available evidence does not permit a clear conclusion, the model should not be forced to artificially produce a binary answer—true/false or yes/no.
Transparency and Reliability of Sources. Transparency must be demanded, beginning with empirical evidence concerning sources.
- No link, citation, or document provided by AI should be considered genuine until it has been independently verified using an external browser.
- The type of source should also be examined—government or institutional sources, television, newspapers, social media, and so forth—because the epistemic value of information varies according to the category of source.
- In particularly controversial cases, it can be useful to ask the model to make explicit the chain of probabilistic connections that led it to its conclusion, allowing the methodology behind the process to be audited.
- One final warning concerning this kind of auditing: it can be useful for identifying statistical associations, but these should not be confused with a logical chain of causation.
- LLMs generate output by calculating probabilities rather than reasoning deductively. Therefore, such an audit reveals which factors the model weighted most heavily, not necessarily the definitive result of the research.
Pre-Existing Knowledge or Real-Time Online Research? Always distinguish between an AI response based on pre-existing knowledge and one based on real-time online research.
- For international current affairs, this distinction is particularly important. An apparently precise response may actually have been constructed from information that is outdated and no longer valid.
- One of the simplest and most direct ways to verify this is to explicitly ask the platform whether it is conducting a web search or relying on material on which it was previously trained.
- More specifically, it is useful to request the chronological sequence of the cited sources.
- When news events occurred after the model’s stated knowledge period, the response should be regarded as unreliable until it has been verified through an independent source.
- Without this precaution, AI can sometimes “bring back to life” people who have actually been dead for months or years, as has happened to me on several occasions.
Geopolitical and/or Religious Omissions. In addition to linguistic issues, it is important to consider the different informational and political-cultural ecosystems in which AI models are developed.
- There may be divergent approaches, particularly concerning international current affairs, as well as systematic omissions in the treatment of politically sensitive subjects—as occurs when interacting with some of the major Chinese AI platforms.
- When these platforms respond with phrases such as “I am not competent,” “I do not have the expertise to answer,” or “I am not qualified to answer,” this means that a deliberate restriction has been programmed into the system.
Multiple Platforms and Mapping Structural Biases. Use multiple platforms—at least two—for the same research topic, comparing and cross-checking the results.
- I have tested this protocol on real cases using Claude, Copilot, ChatGPT, Gemini, DeepSeek, Sarvam, and Mistral.
- Including AI models with different geographic, cultural, and linguistic foundations is essential because it broadens the intellectual horizons of analysts, researchers, educators, and investigative journalists.
- Comparing the performance of different platforms over time makes it possible to progressively build a comprehensive map of both how different models represent events and where their respective strengths and weaknesses lie.
