Free 810-110 practice test: 30 questions with answers

Large language models are trained to predict the next token in sequences, which lets them generate fluent text. Regression and clustering models do not generate language, and image classifiers label pictures rather than write text.

Hallucination describes fluent output that is false or unsupported by the input or known facts. Overfitting is a training issue, normalization is data preparation, and batch inference is a deployment pattern.

RAG retrieves relevant documents and includes them in the prompt, so answers can reflect current, organization-specific content without retraining. Pretraining, quantization, and temperature changes do not add external knowledge.

Embeddings map text to vectors so that texts with similar meaning are close in the vector space, which enables semantic search. Signatures, fonts, and compressed copies do not capture meaning in this way.

Fine-tuning updates a pretrained model with curated examples so it adopts specific behaviour and style. Firewall, DNS, and power changes do not affect model behaviour.

Measuring outcomes across groups and documenting them makes bias visible and auditable. Removing review, relying on self-reported confidence, and hiding criteria each weaken accountability.

Data loss prevention can detect and block sensitive content leaving the organization, and an approved tool policy defines where data may be entered. Hardware, email shutdown, and seating changes do not address data exposure.

Transparency requires disclosure of AI use and an understandable account of how outputs are produced. Scalability, latency, and throughput are performance characteristics rather than ethical principles.

Removing or pseudonymizing identifiers and restricting access reduces the chance that personal data is exposed through the model or its pipeline. Public or widely copied datasets increase exposure.

Adversarial examples are inputs modified in ways that cause a model to make wrong predictions, often while looking normal to people. The other terms are unrelated to model attacks.

Correlation analysis measures how strongly two numeric variables move together, which helps guide feature selection. The other techniques apply to images, audio, or network traffic.

A validation split supports tuning while a held-back test split provides an unbiased final estimate. Using the test split repeatedly or splitting by alphabet does not give a reliable evaluation.

Clustering finds groups in unlabelled data, which fits segmentation without outcomes. Supervised methods require labels, and reinforcement learning needs an environment with rewards.

Accuracy can look high by predicting the majority class, while recall and F1 for the minority class show whether the rare cases are found. Record counts, training time, and feature counts do not measure predictive quality.

A random split can place future records in the training data, so evaluation overstates real forecasting performance. A chronological split avoids this leakage.

An agent decides on a sequence of steps and invokes tools, such as look-ups or updates, to achieve a goal. A simple chatbot mainly answers messages, and agents can access external systems through approved tools.

Human approval for destructive actions and logging of tool calls keep people in control and make actions auditable. Administrator rights, self-approval, and removing logs each remove those safeguards.

Tools let the agent act, and memory or state lets it track progress across steps. Display logos, printer drivers, and shortcuts are unrelated to completing multi-step tasks.

Least-privilege tool permissions limit what the agent can do, and regular log review detects behaviour that drifts from policy. Letting the agent edit its own policy, giving unrestricted keys, and skipping evaluation each remove those controls.

Routing the missing approval to a human workflow keeps the decision with an accountable person while the agent completes what it can. Guessing or bypassing approval breaks the control, and removing the tool prevents the task entirely.

Explicit instructions, a defined output format, and a few examples narrow what a good answer looks like and make results easier to parse. Length alone, capital letters, and random seeds do not improve consistency.

Chain-of-thought prompting asks for step-by-step reasoning, which often improves results on multi-step problems. The other terms do not describe a reasoning technique.

The system prompt holds persistent instructions such as role and rules that apply across the conversation. A single user message applies only once, and file names and training logs do not reach the model.

Separating untrusted content from instructions and validating outputs before actions reduces the impact of injected instructions. Unrestricted tool access, trusting retrieved content, and disabling logs each increase the attack surface.

Explicit constraints on length and tone shape the output directly and make the result checkable. Leaving requirements unstated or asking the model to guess produces unpredictable results.

AI-generated code can contain bugs or vulnerabilities, so it should go through the same review and testing as other code. Skipping review or tests removes the safeguards that catch these problems.

An AI workflow pairs a model call with a trigger, such as a schedule, and an integration that delivers the result. A manual spreadsheet, a hardware change, or a backup routine does not describe this automation.

Reviewing and testing generated queries on a safe copy prevents data loss, and read-only access limits the damage of a mistake. Direct production runs, broad write access, and skipped backups each raise the risk.

Version control and peer review make each change traceable and reversible before it reaches production. Manual production edits, emailed files, and skipping review remove those controls.

AI assistants speed up drafting of routine code, but the output can contain bugs or insecure patterns, so testing and review remain necessary. The other claims overstate what the tools deliver.