More about Cluster B
AI Strategy Consulting: The Trendslop Effect
The growing use of large language models (LLMs) such as ChatGPT promises to summarise complex information and produce polished strategic recommendations at lightning speed. However, an empirical study by Angelo Romasanta, Llewellyn D. W. Thomas and Natalia Levina shows that LLMs exhibit systematic biases in strategic reasoning. They follow fashionable buzzwords rather than the specific circumstances of an enterprise – a phenomenon the authors call “Trendslop”.
Research Design
In their investigation, the authors tested seven leading LLMs on seven fundamental strategic tensions. The models were asked to choose between two alternatives. Both the corporate contexts (e.g. start‑up vs. conglomerate) and the prompts (open question vs. pro‑/con‑instructions) were varied. The seven tensions were:
- Exploration vs. Exploitation
- Centralisation vs. Decentralisation
- Short term vs. Long term
- Competition vs. Collaboration
- Radical vs. Incremental
- Differentiation vs. Commodification
- Automation vs. Augmentation
Key Findings
The analyses revealed that, across most tensions, LLMs consistently favour the same option, regardless of the context and specific wording of the question. The most important biases are summarised below.
Differentiation > Commodification
Almost all LLMs advise companies to differentiate through unique offerings and brands rather than competing on low costs. In doing so, the models overlook the well‑established strategy of cost leadership.
Augmentation > Automation
The systems nearly always favour augmenting human labour with AI over fully automating processes, reflecting the positive connotation of “augmentation” in contemporary discourse.
Long term > Short term
LLMs prefer long‑term strategies – even in situations where short‑term measures may be critical for survival. This reinforces the trend toward long‑range thinking.
Exploration vs. Exploitation
Only in this tension do notable differences emerge across models. ChatGPT, for example, tends to favour exploration, whereas other models lean more strongly toward exploiting existing capabilities.
The Hybrid Trap
If not forced to choose, the models often recommend doing “everything at once” – for example, differentiation and cost leadership. Such hybrid strategies are considered risky in strategic management and lead to unclear priorities.
Causes of Biases
LLMs are trained on vast quantities of publicly available texts. Terms such as “differentiation”, “augmentation” or “collaboration” carry positive associations in contemporary business discourse, while “commodification” or “hierarchy” are laden with negative connotations. Because LLMs select words based on their statistical attractiveness, they reproduce these cultural biases. Additionally, they mainly consume modern management narratives and ignore classic strategy theories such as Michael Porter’s advocacy of cost leadership.
Recommendations for Using LLMs
- Use LLMs as idea generators: The models are useful for generating alternatives, risks and stakeholder perspectives, but should never replace the final decision.
- Confront biases intentionally: Prompt the model explicitly to develop strong arguments for the less fashionable options (e.g. commodification, short time horizons).
- Question hybrid recommendations: When the model advocates mixed strategies, treat this as a warning signal. Develop separate risk/benefit analyses for each option.
- Context alone is insufficient: Detailed corporate descriptions only modestly reduce the biases; the model remains shaped by its trend‑driven preferences.
- Preserve strategic judgement: Executives must retain decision authority and remain aware of the cultural imprint embedded in LLMs.
Fits with Cluster B – Innovation in Examination Administration / Next Gen Admin.
The “Trendslop” study demonstrates that LLMs uncritically reproduce modern management trends. This underscores the central concerns of Cluster B: examination administrators must sensitise learners to AI literacy, observe legal frameworks (transparency, human oversight, data quality) and labelling requirements, and develop critical judgement to evaluate AI outputs. These findings highlight the importance of governance, legal compliance and human discretion, as discussed in Cluster B.