Prepare and test your skills
Prepare and test your skills
Data dependency and the knowledge cutoff explain why large language models (LLMs) (opens in a new tab) only know what was in their training data up to a fixed point in time. Data dependency means output quality depends on the quality, accuracy, and coverage of historical training data and the prompt data supplied at use time. The knowledge cutoff is the point where training stops, so the model holds a static snapshot with no built-in awareness of later events, facts, or changes.
Data dependency in foundation models means response quality depends directly on both learned data and prompt data. Historical training data flows into the model and shapes its patterns, vocabulary, and assumptions, while prompt data flows into the specific response for that request. When training data contains errors, gaps, or biases, the model depends on those flaws and can repeat them in outputs. When prompt data is inaccurate or incorrect, the response built from that prompt can also show poor performance or false results.
The knowledge cutoff is the fixed point where a model's training data ends, leaving the model with a static view of the world as it existed at that time. Because training does not continue during everyday use, events, facts, or changes that occur after the cutoff never flow into the model on their own. The model therefore depends on historical patterns and may respond with overconfidence, misinterpreted context, or outdated facts when asked about new developments. To answer accurately about current or specific real-world conditions, the model requires additional context grounded in real-world information and accurate understanding of the user's specific data.
Exam tip: Training data determines what the model already knows, while grounding context supplied with the prompt determines what current or specific facts it can use accurately.
Training gaps are predictable failures that appear when training data was missing, thin, or skewed for a topic, group, or situation. Because the model depends on its training data, underrepresentation in that data flows directly into weaker or less fair responses. The following effects all trace back to the same dependency between limited historical data and current output quality:
These gaps show why historical data quality cannot be separated from present accuracy. Tuning the model on domain-specific data adds missing expertise by flowing targeted examples into the model, while supplying grounded context at request time connects the static model to current facts. In specialized cases with the potential to materially impact individual rights, meaningful human supervision provides an additional check between model output and real-world action.
Hallucinations and edge case failures are two related limits of foundation models (opens in a new tab). A hallucination is a response that sounds fluent and confident but is not factually grounded, while an edge case failure happens when a model handles a rare or unusually complex scenario incorrectly. Both occur because foundation models are large models trained on vast amounts of data that use statistical modeling to predict likely responses and generate new content.
Foundation models hallucinate because they learn language patterns from training data rather than storing a verified set of facts. When a prompt asks about something the training data covered thinly or not at all, the model still produces its most statistically likely continuation, which can be fluent but wrong. The core distinction is that a factually grounded response is supported by real data, while a plausible but incorrect assertion is generated from pattern alone. A second cause is that built-in knowledge is static, so anything that changed after training or was never in the training data is a gap the model may fill with a confident guess.
Edge cases are the complex or rare scenarios where pattern-based reasoning is most likely to break down. In common, well-represented situations, the patterns the model learned are strong, so responses tend to be reliable. In rare or highly specific situations, the training data provides fewer examples, so the model has weaker patterns to rely on and is more likely to produce a plausible but incorrect answer. A hallucination is a wrong statement presented as fact, while an edge case failure is a wrong outcome on an unusual input, such as missing a rare condition, mishandling a complex combination of requirements, or applying a common pattern where it does not fit.
Retrieval-augmented generation (RAG) (opens in a new tab) is the main technique for reducing hallucinations by grounding a response in real data instead of relying only on internal knowledge. In a RAG system, an embedding, which is a numerical representation of data such as text, images, or videos that captures relationships between inputs, is used to match a user's query with relevant knowledge. The flow works in a clear order:
A four-step process flow showing how RAG grounds a model response: the user's query is converted to an embedding, compared via similarity search against documents in a knowledge base, the most relevant documents are retrieved, and the model generates a grounded response from the retrieved facts.
Because the retrieved documents supply the facts, the response is anchored to real content rather than to a likely-sounding guess. This is why grounding is described as ensuring informed and accurate model responses.
Function calling is a feature that connects LLMs to external tools like APIs and functions to enhance responses. Function calling lets a model reach beyond static knowledge and pull in real-time information and services such as databases, customer relationship management systems, and document repositories. The sequence for function calling is:
Function calling is a critical component in AI agent architecture because it gives the model a structured way to specify which tool to use and how to format the input, which helps ensure precise interactions with external systems.
Grounding and function calling address hallucinations in different ways, and the right choice depends on what the response needs. RAG supplies stored documents, function calling supplies live data and actions, and search-based grounding supplies web information. Gemini models can evaluate whether model knowledge is sufficient or whether grounding with Google Search is required, so the model itself can flag when it needs outside information.
| Approach | What it supplies | Choose when |
|---|---|---|
| RAG with embeddings | Semantically relevant documents from a knowledge base | The answer should be based on your own stored content |
| Function calling | Real-time data and services from external tools | The answer needs live information or an action in another system |
| Grounding with Google Search | Up-to-date information from the internet | The model must decide whether its own knowledge is sufficient |
Grounding can also be done without vector databases by querying an existing database such as Cloud SQL or Firestore and using the query result in the model prompt. Together, these options give a model a way to check answers against real sources, which is the practical defense against hallucinations and edge case failures.
Exam tip: A hallucination is a plausible but incorrect assertion produced from internal patterns, while a grounded response is supported by retrieved data or an external tool — the difference is whether the facts came from outside the model.
Bias, fairness, and social implications describe how prejudice in training data affects people who use foundation models. Bias is the tendency to reproduce or amplify prejudices already present in training data, which can lead to outputs that discriminate against or underrepresent certain groups. Fairness is the goal of giving comparable service quality to all users and groups, while social implications are the real-world effects, such as unequal treatment, reinforced stereotypes, and lost trust, that follow when bias goes unaddressed.
Bias and unfairness in foundation models come from several distinct sources, and each one shows up differently in the output a user sees. The chain that produces unfair output runs from training data to model to user: data that underrepresents or stereotypes a group teaches the model a distorted pattern, the model applies that pattern when generating a response, and the user receives an output that can carry the distortion into a real decision.
| Limitation | What causes it | How it shows up in output |
|---|---|---|
| Bias amplification | Existing biases in the training data | Outputs that reinforce societal prejudices and unequal treatment of certain groups |
| Language quality | Underrepresentation of some dialects, language varieties, and non-English languages | Inconsistent service quality across users and weaker text generation for less-represented languages |
| Fairness benchmarks and subgroups | Fairness analyses cover only gender, race, ethnicity, and religion, and only English data | Potential risks outside those axes and languages are not exhaustively accounted for |
| Edge cases | Unusual, rare, or exceptional situations not well-represented in training data | Model overconfidence, misinterpretation of context, or inappropriate outputs |
Bias amplification is the key mechanism here: rather than smoothing out an existing prejudice, a generative AI model can inadvertently amplify it.
Fairness gaps appear when some users or groups are less represented in training data than others. The majority of benchmarks used to evaluate models, including all fairness evaluations, are in the English language, so text generation may not be as effective for some dialects or language varieties. Performance can be worse for non-English languages or English varieties with less representation, so different users can receive inconsistent service quality even when they ask the same kind of question. Fairness analyses also do not provide an exhaustive account of risk because they focus on biases along gender, race, ethnicity, and religion axes and are performed only on English language data and model outputs.
Exam tip: Fairness evaluations cover gender, race, ethnicity, and religion, but only for English language data and model outputs — a clean fairness result does not prove the model is fair for other languages or for groups outside those axes.
The social implications of bias and unfairness grow when a model is used in a specialized or high-stakes setting. Generative AI models can lack the depth required to give accurate and detailed responses on highly specialized or technical topics, which leads to superficial or incorrect information. For specialized, complex use cases, models should be tuned on domain-specific data to add missing expertise. There must be meaningful human supervision in contexts with the potential to materially impact individual rights, because a model facing a rare situation it never learned well may answer confidently rather than flag its uncertainty.
Practices that reduce bias and fairness risk combine built-in safeguards with testing that is specific to the use case. Vertex AI Studio (opens in a new tab) has built-in content filtering, and the generative AI APIs have safety attribute scoring that lets customers test safety filters and define confidence thresholds that fit their use case and business. Google recommends taking these steps in order:
Beyond these steps, Google Cloud runs abuse monitoring processes that may help detect potential abuse and violations of its terms for Generative AI Services, with more rigorous requirements for models or features designated as Advanced AI. Customers remain responsible for adhering to the Acceptable Use Policy, the Generative AI Prohibited Use Policy, and other Service Specific Terms. Suspected abuse, or generated output containing inappropriate material or inaccurate information, can be reported to Google Cloud through its report form.
The knowledge cutoff in large language models (LLMs) causes outdated answers because it is the fixed point where training data ends, leaving the model with a static snapshot with no awareness of later events. Because training does not continue during everyday use, the model depends on historical patterns and may respond with overconfidence, misinterpreted context, or outdated facts when asked about new developments unless it receives additional context grounded in real-world information.
Foundation models produce hallucinations because they learn language patterns from training data rather than storing verified facts and still generate the most statistically likely continuation when coverage was thin or missing. A hallucination is a plausible but incorrect assertion produced from internal patterns, while a grounded response is supported by retrieved data or an external tool.
You should use Retrieval-augmented generation (RAG) with embeddings when the answer should be based on your own stored content because it supplies semantically relevant documents from a knowledge base. Use function calling when the answer needs live information or an action in another system, and use grounding with Google Search when up-to-date information from the internet is needed and the model must decide whether its own knowledge is sufficient.
Fairness evaluations for foundation models are limited because they cover biases along gender, race, ethnicity, and religion axes and are performed only on English language data and model outputs. Because of this scope, potential risks outside those axes and languages are not exhaustively accounted for, and text generation may be less effective with inconsistent service quality for non-English languages or less-represented dialects.
A global enterprise deploys a foundation large language model (LLM) to assist compliance officers with policy research and regulatory reporting. During initial user testing, the compliance team identifies two critical failure modes:
Which pair of foundation model limitations directly explains these behaviors?
Last updated