AI’s now universal corporate adoption means that every B2B marketing team is essentially relying on the same set of generative AI tools and prompts to help deliver their insights. In turn, content outputs from many companies look increasingly similar. LinkedIn is the most obvious channel to see this, where it’s all too easy to slowly drown in a sea of reasonably polished and articulate reports and articles, few of which offer much interesting insight. 

When everyone’s using the same AI tools to analyse the same underlying information, then it’s all too easy to fall into the homogenisation trap. Everyone arrives at the same five trends, structured in the same tone, with similar conclusions. It’s helpful if you were pretty terrible at finding and creating insights before; but it drags you down if you were somewhat decent at it. 

A fascinating new HBR article, Research: The innovation problems AI can’t solve, helps shed some light into how and why this happens. For example, when AI is used during the research ideation process, LLMs will typically steer you towards familiar, more conventional ideas. While these tools are generally better at summarising a wide range of inputs and ideas, they’re typically poor at suggesting more interesting ideas that lie further away from the mean.

This is a problem for B2B firms using thought leadership to help sell complex, high-value services. Distinctive, research-based content is supposed to establish a distinct point of view, with some rigorous research to underpin it. The issue is not the use of AI; it is the reliance on AI as a substitute for primary insight. As the article explains, when AI is tasked with coming up with the core hypothesis for any primary research, it naturally selects the most statistically probable narrative. This is “thought followership”, though, not thought leadership. 

In search of the outlier 

As the authors point out, because “a single exceptional idea is worth more than dozens of merely good ones, the task isn’t to produce a large pile of solid options—it’s to surface genuine outliers”. But when people use LLMs to support their brainstorming and ideation process, they get pulled towards the statistical mean, based on what AI regards as the most likely idea. And as researchers start to fixate on that idea, it actually narrows their thinking. (An earlier HBR article made a similar conclusion, albeit through the lens of using AI for strategic advice.)  

A related risk that the HBR authors flag is the bias towards more “polished” ideas from AI. Research ideas originating via AI are typically fluent and well-structured by design, which makes it easier to mistake them for high-quality ideas. In turn, it’s easier for companies to greenlight these ideas for a new campaign, without properly kicking the tyres on it. In these situations, the scrappier, less polished diamond in the rough gets overlooked. 

There are other issues too. The article highlights the risks of relying on simulated or synthetic data within thought leadership research (an approach I’ve been concerned about for some time now). The researchers find that these synthetic respondents will choose a rational choice far more often than actual humans do. The article discusses this in the context of conducting market research (eg, how will customers respond to a change in price), but the issue applies for thought leadership research too. The safer option is to rely on real-world human research.  

Evading the trap

How then, can marketing teams steer clear of these risks? The authors’ recommendations vary, depending on the research stage, and on the type of research being conducted. Within the ideation phase, they encourage users to revisit how they prompt AI. For example, making use of chain-of-thought prompting: asking the model to draft a set of ideas, and then explicitly prompting the AI tool of choice to revise these into bolder and more distinct options. (On a side note, the authors warn that while this works well for AI tools, the same guidance to “think more broadly” is much less effective for people, and the tendency is for people to fixate on the initial ideas suggested by AI. The authors call this a variation of the “don’t think of a white bear” problem: trying to suppress an unwanted thought makes your brain more likely focus on it). 

Once a set of hypotheses or ideas for a new research initiative are mapped out, they recommend presenting these all in an identical form to the decision-making group. This makes it easier to judge the ideas on their inherent quality, as opposed to simply opting for the one that is presented in the most compelling way. 

In other findings, the authors counsel firms to steer clear of synthetic data inputs, or at least to exercise strong caution. I’d go further: an essential input for any thought leadership campaign should be some kind of original, proprietary research. Feeding new data into your AI tools is much more likely to deliver differentiated and interesting insights that help your work stand out; relying on secondary and/or synthetic data will make your work much more likely to blend in. (An important side note here: AI can clearly be a very powerful tool for delivering this new primary research, and a small but interesting range of AI-enabled research studies are worth a look.)     

Key steps for differentiated insight 

Escaping the homogenisation trap requires a clear division of labour between human editorial judgment, primary research, and AI. Extrapolating from these HBR insights, a revised thought leadership research process could look more like this:  

1. Map the consensus: Before starting a research project, use AI to map out the standard industry narrative on the topic. Identify the expected recommendations, and the common assumptions being made. Doing this at scale is very time consuming for humans, but it’s an area where AI usually excels. 
2. Look for the anomalies: Have your experts identify where proposed narratives and ideas fall short. Which recommendations are simply just common sense, or a reversion to the mean? Is there a more counter-intuitive alternative to consider? Either rely on human judgement to get this right, or make use of smarter prompting to get AI to shift away from the mean. 
3. Capture original data: Design a primary research methodology that can test that more interesting hypothesis. This could be done via traditional quant and qual research methods, or by using AI tools to underpin new research approaches. Steer clear of synthetic data and methods.  
4. Accelerate your content creation. Once you have gathered unique, primary data and have your core messaging locked down, then make use of AI tools to accelerate the creation of a wider range of spin-off content. AI accelerates the execution, but the core insight remains distinctly human and backed by original data.