← All insights ←

Strategy Trendslop: When your competitor's AI proposes the same strategy

Published: 28 August 2026

Strategy Trendslop: When your competitor's AI proposes the same strategy

I hear it often in meetings, whether Board of Directors, Foundation Board or Executive Board: why not ask ChatGPT, Claude or Gemini. Usually the questions concerned are the more complex ones, those that would require larger volumes of data and therefore longer reflection and longer analysis. AI is also used more and more when developing strategies.

In principle we humans are inclined to give quick answers. It lies in our nature and it would be neither efficient nor practicable to start lengthy thought processes about each and every thing. Daniel Kahneman therefore distinguished two modes of thinking in Thinking, Fast and Slow. System 1 answers quickly and without effort. System 2 calculates, checks and weighs up, and in doing so it consumes measurably more energy. That is why it only kicks in when it has to.

In today’s meetings it often no longer has to. There is a third option: delegate it to the AI. People are efficient, or they are simply taking the comfortable route.

Yet what arrives at the other end is less neutral than assumed. Ask a language model about your strategy and it will most likely recommend differentiation and a long planning horizon. Cost savings and quick measures rarely appear in the answer. This result is independent of your industry or your current economic situation. The bad news: your competitor, asking the same or a similar question, receives the same result.

The wrong expectation

Angelo Romasanta of Esade Business School, Llewellyn Thomas of the University of Sydney and Natalia Levina of NYU Stern describe in the Harvard Business Review of March 2026 exactly the assumption that underlies this reflex.

Executives assume that a language model delivers an unbiased outside view. For a system trained on vast amounts of text, that appears plausible. The authors consider it an error. A language model is not the colleague who critically examines existing reasoning, digs into the specifics of the situation, puts assumptions under pressure and then objects when everyone has made themselves comfortable. On strategy questions it resembles more, in their formulation, a ‘freshly minted MBA or junior consultant’ who parrots whatever is currently popular.

That is precisely the point at which the expectation tips over. You seek an outside view and get an inside view of the zeitgeist.

The delegation of thinking in two stages

That people offload thinking when they can, when it appears more efficient or when the situation demands it, instead of entering the complex thought process, has been known at the latest since Kahneman’s research on the economics of decision making. Michael Gerlich of SBS Swiss Business School surveyed and tested 666 people in Societies in 2025 and found a strong relationship between AI use and cognitive offloading, as well as a clearly negative one between AI use and critical thinking. Hao-Ping Lee and colleagues from Microsoft Research and Carnegie Mellon University arrived at a related finding at the CHI conference in 2025: the greater the trust in the AI, the lower the willingness to think critically oneself. The greater the confidence in one’s own competence, the more likely it is that critical analysis takes place and answers are questioned. Both works rest on surveys and show correlations, not proven causalities.

At the second stage, the one people now offload to the AI, the decisive thing happens. The AI orients itself on statistical values of the underlying data sets and optimises for positively rated trends.

The authors of the HBR study examined seven models, among them ChatGPT, Claude, Gemini, Grok, DeepSeek and Mistral, and presented them with seven entrepreneurial tensions as a clear either/or question. Each data point represents the average preference of a model across fifty runs.

If the models were neutral, the values would spread around the middle. Instead they cluster tightly on one side, across almost all providers. The comparisons mostly produced the same outcome:

  • Differentiation beats cost savings: The models almost consistently advised unique positioning with a price premium and avoided the route via standardisation and cost efficiency.
  • Augmentation beats automation: Technology should extend the capabilities of the existing workforce. Replacing human work with technology was rarely recommended.
  • The long term beats the short term: The models preferred multi-year undertakings, even when the situation was urgent and a short-term measure would have been indicated.

Real differences between the models were found in only one of the seven fields, namely exploration versus exploitation. That does not exonerate the models, however. Each one remains biased in itself; ChatGPT, for instance, still leans clearly towards exploration. Since most executives use only one model anyway, the spread between providers helps little in the individual case.

The authors call the pattern strategy trendslop and trace it back to the models predicting the most strongly socially desired answer, measured against the ‘average of the internet’.

A better prompt barely helps

The obvious objection is that one simply has to ask better and therefore write different prompts. The authors tested it with over 15,000 runs on ChatGPT-5, varying order, role framing, the pro and contra request as well as success incentives. For differentiation and augmentation, the share of biased answers moved by less than two per cent. For the remaining five fields the movement averaged 22 per cent. More context likewise helped only to a limited degree, on average by eleven per cent, sometimes in one direction, sometimes in the other.

The most revealing part of this measurement lies in where those 22 per cent came from. Almost the entire movement went back to a single factor, namely the order in which the two options stood in the prompt. Reversing them lowered the probability of the biased answer by 19 per cent.

The model therefore reacts most strongly to the order in the prompt. For practice this means two things. The apparent improvement through skilful prompting is to a large extent an order effect. And the direction of the bias thus depends in part on how someone happened to write the question down. That does not make it controllable.

From this follows a test that actually costs nothing. Ask the same question a second time with the options swapped. If the recommendation turns out differently, what you have in front of you is an artefact of the wording and not an analysis.

The hybrid trap

The study’s second finding is even more dangerous in everyday meetings, because it disguises itself as quality: as soon as the authors removed the requirement for a binary decision, the model frequently recommended two options at once. Differentiation and cost leadership, radical and incremental innovation. It reads as balanced and is often difficult to attack. It is the sort of proposal that has the potential to pass through a board without a counter-question, because nobody has to expose themselves.

Unfortunately this answer describes a position that strategy research has warned about for decades. Michael Porter calls it being stuck in the middle. Differentiation and cost leadership demand opposing capabilities in the organisation, different processes, different cost structures, different people. Anyone pursuing both at the same time builds neither of the two capabilities.

The cause is mechanical: present a language model with several options in one prompt and it tends to weight them, to rank them or to blend them. Strategy, however, consists of choosing under uncertainty. What you leave out decides as strongly as what you do.

For practice this means two things.

  • A recommendation that proposes both at once should be treated as a warning signal and not as a compromise.
  • Anyone wanting to use an AI to develop strategy has each variant examined in a separate run and carries out the weighing up themselves.

A second model and a comparison between models does not solve the problem

When I read this finding, I took the view that in practice I could simply run the strategy pass additionally with another provider and thereby obtain greater variance. Unfortunately the science currently says otherwise here too.

Elliot Kim, Avi Garg, Kenny Peng and Nikhil Garg have already examined this across more than 350 models and presented it at ICML in 2025. The errors of different models are strongly related. In one of the datasets, two models agreed in 60 per cent of cases when both were wrong. The decisive part is the second one: it is precisely the larger and more accurate models that show highly correlated errors, even with different architecture and different provider. Anyone cross-checking Claude against ChatGPT buys less independence than they assume.

The advocatus diaboli that I developed for certain tasks also does not always help and is almost worthless for purely strategic questions. Research on debates between several models points in the same direction. The agents tend to join the perceived majority and the consensus reached reflects conformity instead of correctness. In one investigation of this effect, the models tipped from a correct to an incorrect answer in the majority of the observed changes of opinion.

What this means for the Board of Directors

Up to this point it is a problem of the individual company. The actual point lies one level higher, and Jon Kleinberg and Manish Raghavan showed it formally in PNAS in 2021.

Imagine: ten companies assess the same market question, each with its own imperfect judgement. They err differently. Some get it wrong, others get it right, and the market sorts this out over time. Today, however, all ten reach for the same tool, which taken on its own judges better than any of them. For each individual company this is progress. For all of them together it becomes worse, because they now err at the same time and in the same direction. A misjudgement no longer hits just the one, it hits all.

From this follow four aspects that Executive Boards and Boards of Directors should observe:

  • The model widens options but makes no choice. For risks, blind spots and missing perspectives it is usable. For the decision it is not.
  • The known inclination is deliberately turned against the AI. Explicitly demand the strongest justification for the option that the model does not propose.
  • Every question is asked in different orders. If the recommendation tips, it was a question of wording.
  • A proposal that recommends both at once is a warning signal. As a rule it describes the position between two stools.

At the beginning stands a delegation of complex thought processes. The human passes the difficult question on to the machine and the machine evaluates it on the basis of the average of the internet. What lies on the table at the end is coherently worded, sounds like deep analysis, often impresses with a large volume of data and many arguments, but probably resembles closely what the competition is building on as well. Now at the latest it is time to activate Kahneman’s System 2 and to enter one’s own critical thinking phase.

Anyone examining a strategy paper on the Board of Directors or the Executive Board should therefore in future ask a question that would have made no sense three years ago: where does this proposal come from, and how many others are presumably building on the same recommendation, should it be based on an AI.

References

  • Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
  • Romasanta, A., Thomas, L. D. W., & Levina, N. (2026, 16 March). Researchers Asked LLMs for Strategic Advice. They Got “Trendslop” in Return. Harvard Business Review. https://hbr.org/2026/03/researchers-asked-llms-for-strategic-advice-they-got-trendslop-in-return
  • Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006 (Correction: Societies, 15(9), 252)
  • Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3713778
  • Porter, M. E. (1980). Competitive Strategy: Techniques for Analyzing Industries and Competitors. Free Press.
  • Kim, E., Garg, A., Peng, K., & Garg, N. (2025). Correlated Errors in Large Language Models. Proceedings of the 42nd International Conference on Machine Learning (ICML). https://arxiv.org/abs/2506.07962
  • Hao, X., Wu, Z., Qiu, Y.-X., Xiao, C., Xu, R., Zheng, S., & Qin, J. (2026). Not All Flips Are Conformity: Decomposing Stance Convergence in Multi-Agent LLM Debate. arXiv:2606.00820 (Preprint, not peer-reviewed). https://arxiv.org/abs/2606.00820
  • Kleinberg, J., & Raghavan, M. (2021). Algorithmic monoculture and social welfare. PNAS, 118(22), e2018340118. https://doi.org/10.1073/pnas.2018340118