Why AI Gets People Wrong: The Missing Mental Layer That Changes Everything for Your Business

Why AI Gets People Wrong: The Missing Mental Layer That Changes Everything for Your Business

An AI model watches a cup get moved into a cabinet. It tracks the object. It tracks the motion. It updates the scene. And then it predicts the completely wrong next action because it has no idea what the person in that scene actually believes. That single scenario, drawn from new research published in August 2026, exposes a gap at the heart of how AI systems currently model the world, and it has direct consequences for every business owner using AI tools to understand, serve, or communicate with customers.

The research, authored by a team that released both a framework and a working implementation on GitHub, argues that today's leading world models, including OpenAI's Sora, Google DeepMind's Genie 3, Meta's JEPA, and Marble, only simulate the physical layer of a scene. They track objects, positions, motion, and occlusion. What they never track is what the people in that scene believe, want, feel, intend, or consider socially appropriate. The researchers call this the missing mental layer, and they built a framework called Mental World Modeling, or MWM, to address it. The framework extends classic world models with mental variables including beliefs, attention, goals, intentions, emotions, norms, and social relationships.

To test whether this approach actually works, the team built MENTIS, a modular, training-free pipeline, and benchmarked it against eight major language models, including GPT-5.6-Sol, GPT-4.1, Claude Fable 5, Claude Opus 4.8, and Claude Haiku 4.5. They evaluated all of them on Menti-Bench, a dataset of 448 decision scenes covering text, picture stories, and sound-video clips, each with six response options and a reference solution that documents both the correct action and the underlying mental and physical states driving it. The results were decisive. Direct AI answers averaged an F1 accuracy score of 63.3. Self-consistency, where the model answers the same question six times and selects the most common response, pushed that to 77.9. The full MWM pipeline reached 87.9. People reached 98.5. Critically, the weakest model using MWM, GPT-4.1 at 84.9, outperformed the strongest model using only direct answers with self-consistency, GPT-5.6-Sol at 83.6. Adding more compute power alone could not close the gap that better mental modeling closed.

For small and mid-size business owners, this research reframes a question you may not have thought to ask: does the AI you are using to talk to customers, write copy, or handle service interactions actually understand what your customers believe, not just what they say or do? Most AI marketing and customer service tools today operate closer to the physical-only world model described in this research. They process inputs, they track patterns, they generate outputs. But they are not modeling the belief state of the person on the other side of the conversation. That means they can produce responses that are technically accurate and contextually wrong, the digital equivalent of handing someone a cup they no longer believe is in the cabinet.

The implications for customer experience and marketing strategy are significant. The research found that in interpersonal scenes, the full MWM framework improved prediction accuracy by 26.4 F1 points, compared to only 14.0 points in object-focused scenes. In other words, the gap between good AI and great AI is biggest precisely where relationships, emotions, and social context are involved. That is the territory your marketing lives in every day. A sales email, a chatbot response, a retargeting ad sequence, all of these succeed or fail based on whether the AI behind them understands where the customer is mentally, not just what page they visited or what they last clicked.

The good news embedded in this research is structural. The MWM framework does not require stronger models. It requires better-structured reasoning that includes mental variables. That means businesses using AI tools today can start building this kind of thinking into their prompting, their workflow design, and how they instruct AI to approach customer interactions, without waiting for the next model generation.

This week, take one AI-generated customer-facing asset, a chatbot script, a follow-up email sequence, or an ad campaign brief, and revise the prompt that generated it to explicitly include mental context about your customer: what they currently believe about your product, what they are uncertain or anxious about, what social or professional norms are shaping their decision. Run the output side by side with your original and compare. You are not adding complexity. You are adding the mental layer the research shows current AI is missing, and closing the gap between a response that is technically correct and one that actually lands.

AI in marketing is not just about what gets said. It is about understanding the belief state of the person you are saying it to, before, during, and after every interaction. That is where the next edge in AI-powered growth is being built.

Originally inspired by: World models that ignore human beliefs predict the wrong actions, new research shows (https://the-decoder.com/world-models-that-ignore-human-beliefs-predict-the-wrong-actions-new-research-shows/) See how Leads to Conversion can help your business use AI to understand and convert your customers more effectively. Get your free AI audit