Trendslop: Why LLMs Struggle with Strategy

A 2026 Harvard Business Review article, “Researchers Asked LLMs for Strategic Advice. They Got ‘Trendslop’ in Return,” examines how large language models behave when tasked with generating business strategy. The central finding is that regardless of prompt quality, context detail, or model choice, these systems tend to produce advice that is fluent and credible on the surface but largely generic, repetitive, and aligned with prevailing management trends rather than the specifics of a situation.¹

The researchers describe this output as “trendslop”. It consists of recommendations that sound correct—often echoing familiar strategic language such as “invest in AI”, “focus on the customer”, or “differentiate rather than compete on cost”, but fail to engage with the actual constraints, tradeoffs, or unique dynamics of the problem at hand. LLMs are insufficiently specific to be useful for real decision-making.

The Underlying Mechanism

Large language models are not designed to reason about strategy in the way a human expert would. Instead, they are inherently trained to predict the most likely continuation of text given a prompt. This objective function has important consequences.

It drives the model toward high-probability outputs, language patterns that appear frequently and consistently in its training data. In domains like business strategy, the most common patterns are not detailed case-specific analyses but rather generalized frameworks, widely accepted best practices, and stylized management language. Over time, the model internalizes this distribution and learns that the safest and most “correct” response is one that resembles the consensus.

This produces what can be more precisely described as mode-seeking behavior in token space. The model does not explore the full range of possible strategic interpretations; instead, it gravitates toward the center of mass of its training data, the region where probability density is highest. I.e. it reproduces the average way strategy is discussed, rather than generating novel or sharply differentiated positions.

Prompting Doesn’t Fix It

A common assumption is that better prompts or more detailed context could force the model into more specific reasoning. The article shows that this is only marginally effective. Even when given rich, scenario-specific inputs, models tend to snap back to familiar strategic tropes.

Prompting operates within the same probabilistic framework. While it can shift the distribution slightly, nudging the model toward certain topics or constraints, it does not change the underlying objective. The model is still selecting the most likely continuation, and in most cases, the highest-probability continuation remains a generalized, socially reinforced answer.

We learn again that prompting can steer models, but cannot fundamentally alter their tendency toward consensus outputs.

The Role of Training Data

The composition of training data further amplifies this effect. Business and management writing, consulting reports, academic articles, executive summaries, is often heavily skewed toward clarity, coherence, and defensibility. It emphasizes frameworks, best practices, and narratives that are easy to understand and justify.

A lot less visible in the training data are failed strategies, messy decision processes, and the kind of uncertainty and tradeoff discussions that real strategy involves. The model thus learns to associate “good strategy” with legible, well-structured narratives rather than the complex, often contradictory realities of strategic decision-making.

As a result, the model’s internal representation of “good strategy” becomes biased toward legible, defensible narratives rather than messy, uncertain decision processes. It learns how strategy is described, not how it is derived.

There is also a recency and popularity bias. Widely discussed ideas (e.g., AI transformation, platform strategies) appear more frequently and thus carry higher probability weight. This reinforces the tendency to reproduce what might be called the current thing, which means decisions are already late to the party by the time they are generated.

Missing Tradeoffs and Commitment

A deeper limitation is that LLMs do not naturally encode tradeoff logic, which is central to strategy. Real strategic decisions require choosing among alternatives, accepting opportunity costs, and committing to a course of action under uncertainty. This involves elimination as much as generation, the model must not only propose options but also select among them.

Language models, by contrast, tend to accumulate rather than eliminate options. They produce recommendations that are additive (“do X and Y and Z”) rather than selective (“do X instead of Y because…”). This reflects the structure of their training objective: there is no explicit reward for exclusion or commitment, only for plausibility. This leads to outputs that feel comprehensive but avoid the core difficulty of strategy, making irreversible choices with incomplete information.

Alignment and the Reinforcement of Safe Outputs

Post-training alignment processes, such as those described in Training language models to follow instructions with human feedback, further push models toward safe, broadly acceptable responses. Human evaluators tend to prefer answers that are generally reasonable, non-controversial, and easy to agree with. This creates a feedback loop that reinforces the production of “trendslop”-outputs, answers that are unlikely to be wrong, but also unlikely to be decisive.

Position Within Broader Research

This behavior aligns with a broader body of work on language models. “On the Dangers of Stochastic Parrots”² argues that such systems reproduce dominant linguistic patterns without grounded understanding, while “Language Models are Few-Shot Learners”³ demonstrates that their apparent reasoning ability often emerges from pattern completion rather than structured inference.

Taken together, these findings suggest that LLMs are best understood as statistical compressors of human discourse. They capture what is commonly said and recombine it fluently, but they do not inherently evaluate competing hypotheses or construct decisions from first principles.

Implications

The practical implication is however not that LLMs are entirely useless for strategy, but that their role is often misunderstood. They are effective at a narrower set of tasks:

• expanding the space of possible ideas
• articulating known frameworks
• generating plausible starting points

We must acknowledge that they are structurally weak at:

• selecting among alternatives
• reasoning through tradeoffs
• producing context-specific strategic commitments

When used as decision-makers, they tend to default to minimum viable consensus outputs that reflect what strategy typically sounds like, rather than what a particular situation demands.

Conclusion

The phenomenon labeled “trendslop” emerges from the way large language models are trained, to produce the most probable continuation of text within a distribution shaped by human communication. In domains like strategy, where real value lies in differentiation, constraint, and judgment under uncertainty, this leads to a consistent mismatch.

LLMs excel at generating language that resembles strategic thinking. What they lack is the mechanism to reliably produce the non-obvious, context-sensitive decisions that actual strategy requires.