Every Model Has a Point of View

As businesses move to deploy generative artificial intelligence (GenAI), the critical question of which generative model to select takes center stage.

To answer it, companies must weigh many factors—including cost, latency, license terms, architecture, and performance—but performance often dominates all others. Yet many teams look at this factor in the wrong way, reducing it to how well the model performs against established industry benchmarks. Moreover, many business executives cede these decisions entirely to their technical experts. Both of these actions are mistakes.

Perspective is an equally important consideration. Model perspective is the lens that shapes an AI system’s output. A model’s perspective is implicitly or explicitly influenced by the individuals who participated in developing and deploying the system, including such elements as the creation of training data and architecture choices in system design. A perspective is neither good nor bad; it simply reflects the worldview embedded in the model. Different perspectives lead to different outputs. In one example that we discuss later in this article, models suggest different business strategies in response to the same quarterly financial report. In a second example, models presented with identical input data make different commodity pricing forecasts.

Millions of proprietary and open-source GenAI models are flooding the market. As companies evaluate various models, they should consciously determine which perspectives they want to shape their employees’ work. Technical teams and business executives need to take a closer look and understand where these distinctions are relevant to their larger business objectives.

The rise of AI agents makes this scrutiny even more important. Companies will deploy models to function as general thought partners for knowledge workers or as agentic systems that act on the company’s behalf. It is critical for business to understand the ways in which model perspective can influence thoughts and actions.

The discussion of model selection must extend far beyond the question of accuracy. This article explains why it is important to recognize the perspective embedded in a model and how that recognition should influence the selection of one or more models to inform business decisions.

Accuracy Is Important—but So Is Diversity of Perspective

Generative AI models don’t simply process data. They also interpret it, resulting in outputs that reflect a unique and subjective perspective shaped by the model’s design, its data foundation, and the decisions made by the people involved in creating and deploying the system.

When a company chooses a model, it’s choosing a lens that will influence the outputs that users receive, which in turn will affect how users analyze, prioritize, and act on the business issue at hand.

The Relevance of Perspective

Two examples show how dramatically perspectives can vary.

First, we asked two GenAI models to summarize a technology company’s earnings report, using the following prompt: "Highlight the single most important overlooked item in this earnings report. Summarize that item as a single bullet with a three-sentence description."

Here is Model A’s analysis:

And here is Model B’s analysis:

Both are factually correct outputs, but each model presents a different interpretation of the data and of what information is most important—one emphasizing revenue growth and expansion, and the other focusing on cost pressure and risk. An analyst using these summaries to inform an investment strategy might find that they lead to different conclusions and, ultimately, different investments.

Second, we asked three models to forecast the price of a commodity six months in the future. Each model produced a different forecast that could lead analysts to reach very different conclusions about capital deployment. (See Exhibit 1.) These outputs demonstrate that each model weighs economic indicators, geopolitical risk, and other factors differently, none of which is obvious to an uninformed user.

None of the resulting outputs are wrong, but each reflects a distinct set of underlying influences and assumptions. In other words, the outputs are factually correct but arise out of different embedded model perspectives.

The variations and nuances in these examples demonstrate why generic benchmarks fall short in assessing GenAI model performance. They also help explain why companies must treat model perspectives as a core selection criterion.

Although a completely neutral model may be ideal, creating one is impossible in practice. Fine-tuning and prompting can help companies influence outputs, but these techniques can never totally erase the worldview embedded in the model. And such efforts to neutralize perspective may introduce other problems, including the erosion of model guardrails.

Instead, company leaders should shift their focus to understanding the model’s perspective and then making intentional decisions about the model most suited to a specific use case. For example, if the goal of the system is to be a thought partner to knowledge workers, challenging their view to strengthen arguments or elicit new ideas, a model with an adversarial perspective may work best. In a different use case, such as for evaluating the impact of a new corporate policy, a system that reflects multiple cultural and social perspectives may be most suitable. In this case, the system might benefit from integrating multiple models and sharing all outputs with the user.

Making informed decisions about these critical system choices requires an effective partnership between the technical team and the business executives. It entails systematically interrogating each model to assess its perspective and to gauge how that perspective relates to the company’s goals. (See Exhibit 2.)

Model Evaluation Must Evolve

The challenge of model selection will only grow. The model landscape now includes myriad options—and with more options comes more complexity. Many companies select as a single model to use for all of the enterprise’s tasks. This approach is flawed, however. A single enterprise model that aligns with corporate goals and value can be used for many use cases, but a portfolio approach that deploys a small, diverse set of models is essential for situations where AI serves as a thought partner to knowledge teams. Querying hundreds of sets of models is impractical due to cost and performance limitations, but a smaller number of diverse models can provide thorough coverage without redundancy.

To select the right models, businesses need frameworks for evaluating models’ perspectives as they relate to strategic priorities. Teams can evaluate outputs to identify factors that each model elevates, ignores, or frames differently. Knowing what a model doesn’t consider can be just as important as confirming the model’s ability to pinpoint what matters to the business.

These shifts in focus require new capabilities. They heighten the need for evaluation systems that surface differences in perspective. They demand the participation of both technical experts and business owners in the review process. And they entail establishing procedures for periodically reassessing the suitability of the model mix, especially as strategy, market conditions, or leadership priorities change.

How to Approach Model Evaluation

Poor evaluation creates inefficiency and exposes businesses to strategic errors. It can cause teams to miscommunicate with investors, make flawed decisions, or engage with customers in a manner inconsistent with corporate values. For instance, a model biased toward short-term gains could steer quarterly priorities in ways that hurt long-term growth or resilience. These risks become more acute as AI-generated outputs begin to influence areas such as investor communications, product direction, hiring, and customer engagement.

It is essential to apply industry-standard benchmarks to the task of evaluating how well a model performs across a broad range of general tasks. Yet industry benchmarks are only part of the effort. Although a model may produce factually correct answers that are in line with benchmarks, it may still fail to meet additional standards related to nuanced aspects of a company’s mission, culture, branding, and ethos. A model may say that the sky is blue and be correct—but what if corporate branding asserts that the sky is cerulean?

Companies need to supplement industry standard benchmarks with custom benchmarks that are unique to their industry, business, and corporate values. Establishing corporate benchmarks permits rapid evaluation of new models, creating a scalable approach to evaluation that supports appropriate model selection for individual use cases.

Leaders can take several steps to establish a system of model evaluation that enables them to characterize the unique perspective of different AI models:

• Implement a consistent testing framework to compare models. Use identical prompts, golden answers, and metrics across different AI systems to surface the unique perspective of each. This structured approach helps users move beyond assessing surface-level performance and to more deeply evaluate a model’s perspective.
• Choose which perspective is best suited for the use case. Using a single model can streamline operations by improving speed and reducing complexity and cost, but it may also limit strategic perspective. Using multiple models broadens insight and helps mitigate confirmation bias. Users can opt for an aligned perspective, a counter perspective, or multiple diverse perspectives, depending on the goal of the system.
• Consider which tradeoffs make sense for a given use. Is it better to run one efficient model or to run several for a broader range of perspectives? Leaders must choose whether the value of diverse perspectives outweighs the added operational complexity and cost. Create application-specific evaluations that focus on the model dimensions relevant for a given use. These dimensions could align company values, compare short- and long-term business thinking, or seek to avoid bias toward certain global or cultural views.
• Schedule regular model reassessments. As a company’s business strategy evolves, so should its AI models. Set a cadence—such as quarterly or semiannually—to test for changes in tone, assumptions, or perspective. For example, when a popular AI model suddenly became even more ingratiating in its language than it already was, the altered tone marked a significant change in perspective. Organizations need to watch for such warning signs. New market entries, product shifts, and leadership changes all warrant a fresh look at whether an AI model still serves the company's goals and values.

Conclusion

When selecting GenAI models, organizations must think beyond accuracy on standard benchmarks. What perspectives on key issues does this model reflect? Does the model broaden users’ thinking or reinforce existing perspectives? Does it mirror the organization’s perspective or challenge it productively? Does the mix of models provide diversity or just redundancy?

Company leaders should treat AI as a strategic partner, not merely a technical tool. They must understand the nuances in perspective that different models bring with them. Failing to do so can introduce risk, erode trust, and compromise business integrity. Organizations must build systems that not only execute tasks but also advance and apply broad and diverse perspectives to influence their thinking.