Behind the buzzword: What are world models?
Want to know what AI world models are, but you're sick of wading through all the hype? Here's what you need to know about it with no fluff, only facts.
Oct 6, 2025 • 3 Minute Read
In this edition of Behind the Buzzword, we cover a hot topic in the tech industry right now: world models. Here's what you need to know about it, explained in a three minute read.
How AI currently fails to understand reality
Imagine you're holding a rubber ball in your hand. What would happen if you let go of it? Before you even do it, you know that it would drop down, bounce for a bit, and then settle on the ground.
That's because in your head, you've got an understanding of how the world works---including things like physics and gravity. That's why you don't need to jump in front of a train to know that's a bad idea. This ability helps you make smart and safe decisions.
However, AI has no such understanding of how the world works. That's why when Generative AI tries to make videos of our world, they can often make hilarious mistakes. For example, with OpenAI's Sora, it has been known to create videos of beer pouring into a glass in strange ways or people blowing on candles without them going out.
If an AI can't understand reality, that limits how useful it can be; not just in video production, but everything from applied robotics to climate simulations. World models are a proposed solution to this problem.
What are world models?
World models are generative models that try to represent and simulate the workings of the real world, and predict future states. The idea is this allows AI to gain that understanding of the world they’re lacking, much like we do intuitively before we make a decision.
What are the potential benefits of world models?
1. Better AI decision making
Knowing how the world works is crucial when it comes to making decisions (or providing others with advice) and would make an AI significantly more useful.
2. Faster, cheaper, and safer AI training
Training an AI in the real world and having it learn through trial and error can be time-consuming, dangerous, and expensive. World models would provide the option for it to use simulated experiences instead to predict outcomes.
3. Improved robustness and performance
Right now, if you throw generative AI a curveball---like asking it to plan a path for you through a city, and then blocking some of the streets---it can come up with a really convoluted, winding way to get you to your destination. With a world model, it would know how the world works, and that it could just reroute you around those obstructions.
4. A new tool to empower industries
World models could be exceptionally useful in industries such as game development, robotics, science, and medicine. Think genomics, drug discovery, or climate modelling. For example, you could simulate climate changes over a long period of time.
5. A potential path towards reducing AI hallucinations
Some researchers see world models as a way to reduce AI hallucinations and enable reliable reasoning, which current AI cannot do (Something I cannot emphasise enough in articles like these.) Others see it as a path towards Artificial General Intelligence (AGI).
What are the barriers to world models?
There are a lot of unsolved issues with making world models a reality, such as:
- Figuring out the "how" of actually creating a true world model
- Creating world models with high enough accuracy that they're not creating unrealistic simulations
- Overcoming architectural obstacles around long-term memory and high-dimension data
- Making sure the models successfully generalize the real world effectively
Are world models a buzzword or a real thing?
World models have existed for a while in some form or another, it's just whether or not they're effective enough to be used as expected. For example, in the late 1960s, an AI system called SHRDLU created a "world block" to answer questions about tabletop objects, but didn't scale up. Recently, Google released Genie 3, a general purpose world model, which creates an environment based on a text prompt that the user can navigate in real time.
Note that there is a trend of companies claiming to offer a cutting edge AI feature, when in reality it barely counts, and they've just shifted the goalposts. For example, many companies claim to offer agentic AI features in their products, while behind the scenes, it's just plain old chatbots or task automation (There's even a term for it: agent washing.) The industry isn't short of people who are claiming they've achieved AGI, if you just turn your head and squint at their product a little.
There are a lot of incentives for a company coming out as the first to achieve one of these milestones (brand awareness, diriving investment, standing out from the competition, not being seen as hitting a wall in AI advancement, etc.) and very little downside.
In short? When it comes to big AI claims, taking them with a little bit of salt isn't bad for your diet, and world models are no exception.
Conclusion
Hopefully from this extra-short model, you now have a bit more understanding of how AI works to add to your own world model. That's a wrap on the latest Behind the Buzzword!
Futher learning about Generative AI
Pluralsight offers a range of beginner, intermediate, and expert AI and ML courses, including dedicated courses on generative AI. Since you can sign up for a 10-day free trial with no commitments, it’s a great way to take some professionally authored courses with a set course structure. Why not check them out?
More Behind the Buzzword articles
Advance your tech skills today
Access courses on AI, cloud, data, security, and more—all led by industry experts.
