From Language Fascination to Forensic AI Writing”

My interest in large language models began, and is still sustained, by a long standing fascination with language itself. When a friend asked if I would help with an opinion piece for Politiken, it immediately made me curious about how far I could push AI as an analytical partner. I wanted to see whether I could map a national news outlet (Politiken) in terms of lexical diversity, semiotic patterns, unconscious linguistic habits, bias, values, affective dimensions, emotional signals and 8-10 other underlying mechanisms, and then let that map form the ground for writing that truly belongs in that space.

Building a Media DNA Matrix for Politiken

I started with 20 leadership related texts from Politiken written between 2023-2025. Perplexity Pro gathered and documented the corpus so I had a clean base. Gemini with a large context window became my analytic lens. I wrapped both tools in what I now think of as a forensic AI pipeline, a small engine where each component knows its role. At the centre sits a machine readable Media DNA Matrix in JSON format. That file acts as ground truth for the way Politiken speaks, thinks and signals power.

Inside the Analyzer, Generator and Judge Workflow

The architecture rests on three specialised agents that work in sequence. The first agent is the Analyzer. Its only task is to read. It receives the full corpus, applies a battery of instructions and outputs the Media DNA Matrix. Stylome metrics, ideological markers, psychological distance and a business emotion profile are all written into that structure. 

The second agent is the Generator. It never sees the original articles. It receives the Matrix as a dynamic style guide. Sentence length, the ratio between active and other voices, preferred grammatical patterns and typical phrase clusters are all made explicit. 

The third agent is the Judge. It reads any new draft, compares it line by line with the Matrix and returns an alignment score together with commentary on metaphors, values and emotional cues.

That three step workflow changed the way I think about writing for specific arenas. The Analyzer holds the burden of understanding. The Generator creates under constraints that come from data. The Judge keeps the system honest and leaves a trace of each decision. Together they form a loop where style becomes an operational object that can be inspected, shared and improved.

The Matrix rests on four layers. The first captures the stylome, the linguistic fingerprint. I measured lexical diversity around zero point five three five and lexical density around zero point six seven four. Average sentence length landed at about seven point nine words. Active and directive constructions dominated with a ratio of roughly two to one. N gram analysis highlighted recurring phrases such as maengden af maal and der er brug for which carry the pulse in these texts. The combined pattern reveals compressed, staccato like prose that delivers substance in tight lines.

The second layer maps culture and values. I borrowed techniques from fact checking research and ideological profiling. Value coding across the corpus revealed a stable triad of growth, individual responsibility and change. Metaphor analysis showed that economic life is often framed as struggle or conflict, while organisations are described as living systems that adapt and grow. This mix points towards a conservative progressive hybrid profile. Readers are invited to take personal responsibility inside systems that are under pressure yet capable of development. The value map guides which narratives fit that space.

The third layer addresses psychology. I used markers from construal level theory to estimate psychological distance and combined them with a two dimensional model of emotion along valence and arousal. The corpus as a whole scored around 86% percent abstraction. The texts stay close to principles, structures and long term patterns. The emotional baseline came out slightly negative and low in activation. The voice is critical, reflective and calm. It speaks to readers who want to think, who accept nuance and who can live with tension that is not resolved right away.

The fourth layer builds a business to business emotion profile using zero shot classification. Generic sentiment categories such as positive or negative cannot capture the nuances that drive business communication. I defined a set of domain specific emotions that often shape strategic conversations. Thought leadership, innovation urge, operational stability, risk aversion, trusted partnership and belonging were among them. For each article I asked the model to score how strongly the text activated each of these emotions. Politiken scored extremely high on thought leadership and innovation urge and very low on classic safety and relational warmth. I could now treat the feeling of the news outlet as a measurable signal and design submissions that spoke directly to the intellectual pride and curiosity of its readers.

Once these four layers were in place, the Media DNA Matrix became the centre of the system. Each metric, each value cluster, each psychological marker and each emotion score found a home in the JSON structure. I began to see the project as an engine for computational public relations. The Generator agent received precise parameters that define the outlet voice. Sentence length, proportion of abstract versus concrete content, dominant metaphors, expected emotional climate and preferred business emotions all flowed into the prompt as direct instructions. The Judge agent received the same Matrix and could comment on alignment with real authority.

Metaprompting shaped the design. I wrote prompts that asked the models to write other prompts. The Analyzer prompt asked for a precise Media DNA representation. The Generator prompt asked for drafts together with short reasoning about how each line reflected the Matrix. The Judge prompt asked for an alignment score and a breakdown of where the draft matched or diverged from each layer. Chain of Thought reasoning stayed inside the system as scaffolding and left the human free to focus on direction and ethical judgment.

Another important choice was the use of different models for different roles. Perplexity Pro acted as the fact finder and corpus builder. Gemini acted as the slow, careful analyst with a wide field of view. A generative model could then focus on style and argument without having to carry the whole load of research and pattern detection.

When I look back at that first Politiken case now, the opinion piece almost feels like a side effect. The lasting outcome is the method. Communication steps out of pure craft and enters explicit modelling. An editor, a communications director or a sales leader can see how an outlet or a client speaks, feels and frames the world, because the patterns are written down. They can decide how strongly they want to align with that pattern and where they want to create friction. They can repeat the process for other stakeholders and build a small library of Media DNA profiles that guide their choices.

For me this is where the inspirational part lives. Advanced strategic intelligence is accessible. It does not require a secret department or exotic infrastructure. It rests on curiosity, discipline and a willingness to think in systems. A single leadership team can commission one such engine, learn from the results and let that experience reshape how they think about AI. The models become partners in listening before they become partners in speaking.

Advanced Strategic Intelligence as Daily Practice

I began with a favour for a friend and ended with a method that any organisation can adopt and adapt. Choose one arena that truly matters to your strategy. Collect its words. Treat those words as data. Ask your models to read them with the same seriousness you bring to a financial review. Build your own Analyzer, your own Generator and your own Judge. When you see what comes out of that loop, advanced strategic intelligence stands revealed as a concrete daily practice.