AI Isn't a Stakeholder. Here's Why the Label Matters.

An idea steadily gaining traction in corporate reputation circles is that artificial intelligence should be treated as a stakeholder. The argument runs that because companies can now be “seen” by AI — described, evaluated, or recommended by large language models such as Claude or ChatGPT — AI should be measured like any other stakeholder. It’s an interesting idea, but it doesn't hold up. Here are five reasons why.

1. It fails the basic definition

A stakeholder is someone who has a transactional relationship with a company: someone who can be affected by it and, in turn, affect it. Customers buy or don’t buy. Investors put down their money or pull it out. Employees decide whether to stay or leave. In each case, there’s a meaningful ongoing relationship involving real-world behavior with material consequences. AI has none of that. Claude doesn’t wake up with a point of view. ChatGPT doesn’t take the initiative to recommend, complain, or disengage from a brand. Its output about a company comes only in direct response to a human prompt.

2. It doesn’t behave like one

What makes a stakeholder more than just someone with an opinion is that they act on it. We measure stakeholders precisely because of the behaviors their perceptions unlock: the more someone trusts a company, the more likely they are to buy, recommend, support it in a crisis, or consider working for it. AI does none of this on its own behalf. Yes, agentic systems like those emerging from OpenAI’s Operator can undertake some tasks without a human in the loop. But the agent is still executing someone else’s intention. It will never spontaneously recommend a product to a friend or quietly boycott a brand, at least not without prior contextual framing. It has no skin in the game, no agenda of its own.

3. Its “opinions” aren't stable

AI doesn't have opinions in any stable, meaningful sense. Real stakeholder research works because people have reasonably coherent views that can be measured, tracked, and analyzed over time. You can identify what drives their trust, correlate attitudes with behavior, and build a driver model. LLMs are uniquely sycophantic: they’ll reverse their apparent view at the slightest touch. Ask about ethics first and the model locks into an ethics-focused frame. Lead with product quality and you'll get an entirely different set of answers. Tell Claude it’s wrong and it’ll often immediately capitulate. This isn’t the natural variance of human opinion research but something more corrosive: a system that instantly abandons its position depending on what it thinks you want to hear.

4. The better frame is media, not stakeholder

AI behaves more like a media source — a highly personalised one that synthesises existing content on demand. When a user asks an LLM about a company and the model reads ten articles and returns a summary, the output isn’t an “opinion” but a curated digest of whatever the web currently says. That framing makes the problem tractable. The variables that matter are familiar: sentiment (how does the aggregated corpus skew toward this company?) and volume (how often are people consulting AI about this company?). Both are measurable, and both are more analytically honest than asking “what does the AI think?” — a question that anthropomorphises a system that has no thoughts to speak of.

5. Getting the framing wrong has real costs

An LLM’s “views” on ethics or products aren’t coherent facets of a single opinion, but entirely separate outputs that happen to be generated by the same model. Treating them as connected is methodologically dishonest. The “stakeholder” framing also implies a parity that doesn’t exist: if three out of five LLMs say something negative, is AI necessarily “against” you? The question lacks a clean answer because treating AI as a stakeholder is a category error. And ultimately, doing so risks mismeasuring it, misinterpreting it, and misallocating the resources required to influence it.