The grounding problem is a challenge to the idea that a system which only has access to symbols, can refer to the reality beyond these symbols. This problem throws into doubt the ability of AI systems to understand or produce meaningful text.
While current AI systems produce text that we can treat as meaningful, it is questionable whether a system that operates only on text can (a) understand what meaningful text refers to or (b) produce text that can refer.
Can AI systems that only act on symbols and are never exposed to real objects truly understand text and is the text they produce meaningful?
Key Points:
- A grounded AI system establishes appropriate relations between the symbols it uses and reality.
- An AI system producing cogent text is insufficient to establish grounding.
- According to some referential theories of meaning, to understand meaning requires a system to be grounded.
1. Text and meaning
The grounding problem arises for any AI system which uses symbols, but does not connect these symbols to the world. Such a system might have a symbol cat and encode that this symbol relates to the symbol pet, without establishing an appropriate relation to real cats. In particular, an ungrounded AI system has never seen, touched, or interacted with a cat. The system, therefore, lacks the kind of causal relations that human language users typically have to the objects they talk and write about.
According to many referential theories of meaning, AI systems might lack what it takes to establish the reference of words in a text. Such theories hold (1) that the word “cat” has its meaning due to referring to cats in the world and (2) that such reference is established by creating an appropriate link between world and symbols.
The lack of reference might be used to argue that AI systems exhibit one or both of the following two shortcomings:
- Ungrounded AI systems might be unable to fully understand text because they cannot establish reference themselves. The underlying assumption is that a system which cannot establish reference cannot comprehend the reference of text.
- Ungrounded AI systems might fail to generate meaningful text, instead producing text that appears to be meaningful. As the system cannot interact with the world, it cannot establish the reference relations required for the text to be meaningful. The word “cat” in the text would lack any reference.
Both arguments rely on assumptions about the role of reference which are up for debate in the field known as metasemantics. For example, it is not obvious that the AI has to establish reference relations for its output to be meaningful. Instead the text could be meaningful because the ungrounded AI system engages in an exchange with humans who have established conventions about meaning. Thus, the exact deficiencies of an ungrounded system depend on the nature of meaning and its dependence on reference.
2. LLMs and the Grounding Problem
The symbol grounding problem (or vector grounding problem) arguably afflicts large language models (LLMs) that are only trained on text. Text is the paradigmatic symbolic resource. An LLM trained exclusively on text can only encode the relations between the textual symbols, without connecting them to any experience with the referents. Such a system might generate a lexicon entry about cats, but has never seen, touched, or interacted with a cat.
In writing about cats, the system is limited to repeating and inferring from the relationships between symbols it has encountered. Thus, one might argue that, like other ungrounded systems, the LLM fails to understand the text it produces or even that the text lacks meaning because the LLM does not stand in the right relationship to real objects. The question arises how LLMs can be put in touch with the world to ground their symbols.
3. Meaning and Modalities
To address the grounding problem, researchers have trained AI models with information sources other than text. Such other sources can include images, video, and audio data and are known as modalities. A multi-modal AI system might not only have been trained on texts about cats, but also have been exposed to images of cats and even the sound of their meowing.
A system that can both produce the image of a cat in response to text input and can describe the image of a cat, plausibly has encoded some connection between the textual symbols and the visual representation. However, even in such cases one might wonder whether the additional information is sufficient for fully grounding the symbols. For example, real cats are three-dimensional, but a system processing images and text would never have been exposed to three dimensions and therefore might still lack important information.
In addition to lacking information, multi-modal AI systems might also fail to integrate the information correctly. The information has to be used in such a way as to establish the appropriate relation between system and reality. The AI system has to grasp that the description “Toby the cat” and the image of Toby are related due to a unified object — a specific cat — that is causally linked to both the description and the image. What exact information processing is required to achieve such integration of different sensory modalities is an open research question.
4. Purported Benefits of Grounding
Granting that the grounding problem arises for LLMs, what are the consequences of LLMs lacking understanding or being unable to produce truly meaningful text? As long as the output is indistinguishable from human created text, one might feel justified in ignoring worries about how the system relates to meaning in a philosophical sense. That being said, the grounding problem is tied up with at least two practical challenges:
First, a system lacking grounding might struggle to generalise its reasoning to new situations. Having never interacted with a cat, an AI system might fail to predict how a cat behaves in a situation that is not covered in its textual training data. Having interacted with a real cat might make such predictions easier, allowing the model to generalise better to new situations. The grounding problem suggests that information from modalities other than text is required for generalisation.
Second, human use of language and reasoning is grounded and this might guide our communication. We have had similar, though not identical, experiences with cats. This grounding might allow us to be more successful in our communication because it serves as a common basis. Thus, communication with ungrounded AI systems might be less effective and more error prone. In describing a cat, for example, we might leave out those facts we have come to assume due to our interactions with cats. An ungrounded AI system, lacking such encounters, will not have access to these facts and therefore not share our assumptions. Hence, on the odd occasion when such an assumption comes into play, the communication might go awry.
From a philosophical perspective, grounding might not be strictly necessary to achieve generalisation and communication. In practice, however, grounding AI systems probably offers the best or only feasible way to achieve the generalisation and communication abilities we expect from intelligent agents.
5. Other Forms of Grounding
The main focus in the grounding debate has been on sensory modalities, that is sources of information such as images and videos. Depending on one’s theory of meaning, however, other sources of grounding might be required to make LLMs able to understand text or make it truly meaningful. Candidates for such alternative sources of grounding include:
- Access to communicative norms, that is access to how language users should behave in communication in contrast to how users do behave,
- interaction with objects, that is being able to directly manipulate objects in reality, and
- the presence of a body which the system experiences as its own.
Thus, the sensory-modal grounding problem can be distinguished from the normative, interactive, and bodily grounding problem. Similar to the sensory-modal grounding problem, solving these grounding problems might be required for generalisation of reasoning or successful communication with human language users.
The comparative importance of these various grounding problems continues to be a matter of debate due to the unresolved nature of meaning; with the sensory-modal grounding problem receiving the most attention.
With the advance of AI, the challenges of specifying how language use can be grounded will grow more pressing and turn from theoretical to practical. We will increasingly seek to align the reasoning and communication of AI systems with our own grounded cognition and through this process we will learn how language is grounded.