AI in strategic communication? Yes! But...

This is the best-known risk of AI. At the same time, it remains one of the easiest to underestimate.

Imagine a simple situation. A model is given 40 pages of a transcript from a parliamentary debate. In its report, it cites a minister as saying: “we must radically increase the tax burden on the middle class.”

It sounds credible. It fits the topic. The problem is that this quote was never said. The model did not find this statement in the text. It generated it because it sounded like something that could appear in that kind of dispute.

Without human oversight, a client could repeat the statement publicly — and then have to walk back words that no one ever actually said.

Hallucinations are only the most visible part of the problem. Fabricated footnotes, false sources, incorrect calculations, or answers in which current data is mixed with the model’s outdated knowledge all stem from the same issue.

The most dangerous thing is that such errors often look correct. The model rarely says outright: “I am guessing here.” It usually responds with confidence, even when it is wrong.

That is why the expert’s work begins at the very stage of formulating the task for the model. A good prompt can reduce the risk of confabulation, but the prompt alone is not enough. Verification is needed: quotes must be checked against the source, numbers against the data, and dates against reliable materials.

This requires experience that cannot be replaced by mere familiarity with the tool. The expert must understand the situation the report concerns: who is speaking, from what position, to whom, at what moment, and with what consequences.

It is not enough to check whether the model made a mistake. One must also understand what the answer means for the specific situation, client, and decision that needs to be made.

AI models are trained to meet the user’s expectations.

This trait can also affect the way AI analyzes data unless appropriate safeguards are put in place.

A strategist asks: “will topic X gain momentum?” The model gives three arguments that it will. No doubts.

When we ask: “will topic X fade away?” the model again finds three arguments — this time for why the topic will lose significance. And once again, it sounds convincing.

The problem is not that the model is always wrong. The problem is that it often adapts its answer to the way the question was asked.

A similar thing happens with strategic recommendations. The model readily relies on simple patterns: “this kind of crisis usually ends this way,” “this group usually reacts in this way,” “this narrative usually works best.” Meanwhile, the data from a specific situation may point to a completely different conclusion.

There is one more issue: the confidence of the response. The model can sound equally convincing when it has strong grounds and when it is relying on weak premises. For the person reading the report, that difference may be invisible.

That is why good AI-supported analysis cannot stop at the first answer that sounds reasonable. Opposing explanations also need to be tested. We need to ask what could undermine a given conclusion. We need to look not only at the arguments that fit the hypothesis, but also at those that weaken it. Sound familiar? This is nothing other than the scientific method. When working with a model, you need to be like a demanding teacher who does not fully trust their student.

Without this, the model’s output may look like analysis. In practice, however, it will only be a well-written answer to a poorly framed question.

Large language models are trained mainly on data in which English is represented far more strongly than Polish.

This does not mean that AI does not know Poland. It does — but it often understands it more superficially, more schematically, and with a greater risk of error.

In communication analysis, this has very practical consequences.

Take a simple example. In the context of a tax bill, someone writes: “Wonderful law, I’ll finally earn less, bravo the government.” The model may classify such a statement as positive because, at the level of individual words, it sees praise. A person familiar with Polish public debate will immediately recognize the irony.

Across thousands of mentions, such mistakes can shift the entire picture of public sentiment. The report may then show growing acceptance, while in reality what is growing is anger, mockery, or the mobilization of opponents.

A similar problem concerns sources. The model may treat a text from an influential weekly magazine and a post from a website that merely aggregates news in much the same way, if both have a similar reach. In strategic communication, however, reach alone is not enough. What also matters is the sender’s position, their influence on other media, their role within a given environment, and the history of previous disputes.

Socially sensitive topics are a particular example. Models tend to be cautious around them. Sometimes they smooth out their answers or avoid stronger classifications. In the analysis of public debate, this can be a problem, because these are often precisely the areas that need to be examined most carefully.

This cannot be fixed by automation alone. What is needed is someone who knows the Polish context: the media, the actors, the mental shortcuts, the irony, the allusions, the conflicts, and the real weight of individual voices.

That is why it is not enough to run the data through a model and accept the output as a ready-made diagnosis. Sentiment must be checked on key samples, source weights must be adjusted, and one must watch whether the model is applying too generic a framework to the Polish debate.

Without this layer, the analysis may be fast, but it will be incomplete or misleading.

A strategic report very often starts from similar questions: who said what, to whom, when, and in what context.

This is exactly where models can make mistakes that look minor, but change the meaning of the entire analysis.

Imagine a report about tensions within a governing coalition. The model attributes a sharp statement to Leader A. In reality, it was made by Leader B from the same coalition. Two names appeared in one paragraph, with a similar topic and a similar context. The model merged them into a single whole.

The result? The entire diagnosis of a “split” may be based on a statement that the person in question never made.

Such errors occur with similar surnames, organization names, affiliations, indirect quotations, and long conversations with the model. Context from one thread can leak into another. In long documents, the model often handles the beginning and the end well, but loses fragments from the middle.

That is why an expert does not only check whether something “sounds logical.” They check whether the statement exists. Whether it has been attributed to the right person. Whether the quote has not been taken out of context. Whether that person still holds the position the report assigns to them.

This requires not only fact-checking, but also knowledge of relationships: who is in conflict with whom, who quotes whom, who supports whom, and who would never say something in that way.

Without this layer, a report may look coherent while actually being misleading.

This is the most difficult type of error, because it looks the most professional.

The model creates a table. Categories. Percentages. A ranking. Recommendations. Everything looks like quantitative analysis. The problem is that sometimes there is no sound method behind this order.

Example: a report recommends withdrawing from a topic because “70% of mentions are negative.” The strategist checks the data and sees that this 70% comes from 200 mentions in a niche group of the competitor’s supporters. At the same time, 1,200 mentions from the mainstream show a predominance of positive or neutral sentiment.

The model counted mentions, but did not weight their significance. It treated a voice from a small, closed group the same as a statement from a place that genuinely influences the debate.

On paper, it looks precise: “70%.” In practice, it is a number without context.

The same can happen with narrative frames. One day the model identifies five narratives. The next day, four. A week later, six. Not because the debate has really changed that much, but because the model organized the material differently each time.

There is also a tendency to reproduce a template. If the previous report contained five narratives, the model may again look for five, even if there are actually three or seven. If the instruction asks for specific coding rules, the model may apply them unevenly: sometimes consistently, sometimes only partially.

That is why the expert must control not only the content of the report, but also the method. Are the categories repeatable week after week? Do the percentages add up correctly? Have the mentions been weighted properly? Have time, source, reach, influence, and context been taken into account?

Without this control, the decision-maker receives a report that looks like quantitative analysis, but is in reality a series of impressive-looking props (lub: nothing but impressive vanity metrics).

After these five points, it would be easy to think that AI in communication analytics is mainly a source of risk. That would be the wrong conclusion.

When implemented well and consistently controlled, AI opens up fundamentally new possibilities. It allows us to do things that previously could not be achieved either with a larger budget or with more team hours.

First: scale. We can analyze a much broader field than before. Not just a few selected sources. Not just a sample. Not just what someone has time to read. The picture still depends on the quality of the data, but it is much closer to the real scale of the debate.

Second: speed. The cycle from observation to diagnosis shrinks from days or weeks to hours. Analysis stops being merely a summary after the fact. It becomes a working tool during events as they unfold.

Third: multidimensionality. Sentiment, dynamics, distribution, segmentation, sources, actors, and the evolution of narratives can all be brought into one coherent view. The decision-maker does not receive five separate pieces. They receive a map of the field.

Fourth: detecting weak signals. AI helps identify patterns that a human would not catch across hundreds of thousands of mentions: emerging topics, recurring narrative frames, and similarities between seemingly distant discussions.

Fifth: a new economics of analysis. What used to be a large report prepared once a quarter can now be built into weekly working procedures. The scope of analytics that was once available only to a few is becoming much more broadly accessible.

But all these benefits are real only when AI works together with an expert, not instead of one.

Fully automated analyses and reports lead to the opposite situation. Scale becomes chaos. Multidimensionality becomes a patchwork of misleading indicators.

AI does not replace the strategist. It gives them a wider field of view, a faster rhythm of work, and tools they did not have before.

This text was, of course, created with the use of AI. But it was not “written by AI.”

The model helped us turn a long list of specific errors we had observed in our work into broader categories.

It helped us write in plain language, so that the text would be understandable not only to technical readers. But every paragraph and sentence in this text was not only revised many times through prompting; in many cases, the final wording was ultimately written manually.

While working on this text, the model repeatedly did exactly what we warn against. It invented problems we had never encountered. It stretched real observations into areas that were not confirmed by our practice. It added things that sounded good, but were not true. That is simply how these models are. And yet we still love them, because they allow us to do things that, just a year ago, we could not even have dreamed of doing.

If you want to see what this approach looks like in your specific field, get in touch.

The first 30 minutes are sales-free. We will show you concretely what can be done today, where AI helps, where it needs to be controlled, and how we work with it every day.

→See if this approach is right for you.