Synthetic Data: Fake With Benefits

Despite the fact that more and more AI-related businesses use the term “Synthetic Data”, it didn’t just show up with the rise of generative AI. Its roots go back decades — to statistical modeling, early speech synthesis in the 1930s, and even the U.S. Census Bureau’s anonymization efforts in the early ’90s. Researchers like Donald Rubin and Paul Rubin were generating privacy-preserving datasets long before “machine learning” was a mainstream term. 

Still, for years, synthetic data remained a tool for niche applications: census masking, simulation environments, academic experiments. 

That’s changed.

Today, synthetic data is front and center in enterprise AI strategy. From Deloitte to Gartner, from JPMorgan to Tesla, it’s being used to solve problems that real data can’t: privacy constraints, limited samples, missing edge cases, or high annotation costs. Gartner even predicts that by 2030, most AI models will be trained on more synthetic data than real.

In this article, we unpack what synthetic data actually is, how it’s generated, and why it’s suddenly everywhere. We’ll walk through real examples, benefits, generation techniques, and how companies like Sphere are helping global teams integrate synthetic data into real-world pipelines. Safely, efficiently, and at scale.

What Is Synthetic Data?

Synthetic data is information that’s artificially manufactured rather than collected from real-world events. In other words, it’s fake data created by algorithms to resemble real data. Despite being “fake,” well-generated synthetic data retains the statistical patterns and structure of actual data. This means synthetic datasets can stand in for real datasets in many scenarios while looking and behaving like the real thing.

To put it simply, what synthetic data means is data that didn’t actually happen but could have – it’s generated by a computer to mimic the real-world data we need. For example, instead of using actual customer records (with all the privacy issues that entails), a bank might generate a synthetic dataset that has the same format and statistical characteristics as real customer data, but with entirely fictional individuals. The synthetic data feels real (accounts, balances, transactions, etc. in similar proportions) without exposing any actual personal information.

Why the buzz now?

Because obtaining large, high-quality real datasets has become a bottleneck. Collecting real-world data can be difficult, expensive, or fraught with privacy and regulatory hurdles. Synthetic data offers a way to quickly generate as much data as needed, tailored to specific conditions, without those hurdles. As Gartner predicted, by 2030 synthetic data will eclipse data collected from the real world for developing AI models. In fact, a widely cited Gartner study anticipated that by 2024 60% of the data used for AI and analytics projects would be synthetically generated. That’s a striking prediction – and as of 2025, we’re already seeing a rapid shift toward synthetic datasets in many AI initiatives.

So, what is synthetic data in AI? It refers to these artificially created datasets that AI practitioners use to train and test models. Synthetic data in AI is generated with the help of algorithms (often leveraging AI itself) to approximate real-world data distributions. The goal is for AI models to train on synthetic examples as effectively as they would on real examples. Notably, synthetic data can be made to order. 

Need more examples of rare events or edge cases? Generate them synthetically. Worried your training data is biased or incomplete? Synthetic data can fill in the gaps or rebalance the dataset. Because of such advantages, synthetic data has been called “a technical solution to a legal problem,” allowing organizations to use data in innovative ways without running afoul of privacy laws.

The Role of Synthetic Data in AI and Machine Learning

Synthetic data plays an increasingly pivotal role in AI and machine learning. In traditional ML projects, developers require large, carefully labeled datasets (sometimes ranging from thousands to millions of examples) to train accurate models. Gathering that much real data can be prohibitively costly or slow. Synthetic data generation comes to the rescue by letting companies create large, diverse training datasets on demand without the same level of cost and effort. 

As an example, Paul Walborsky, co-founder of the synthetic data startup, noted that a single annotated image that might cost $6 via a data-labeling service could be generated synthetically for about $0.06. That’s a 100x cost reduction, illustrating how synthetic data can dramatically lower the barrier to obtaining training data.

Another critical aspect is privacy. Many AI applications (like those in finance or healthcare) rely on sensitive personal data which is protected by regulations. Using real customer or patient data to train models can raise privacy concerns and compliance issues. Synthetic data offers a workaround: because it’s artificially generated and contains no real individuals, it can be used freely without risking personal privacy breaches. Gartner analysts estimate that by 2026, 75% of businesses will be using generative AI techniques to produce synthetic customer data for exactly this reason. In areas like healthcare, synthetic patient data allows researchers and engineers to develop AI solutions (for example, disease prediction models) without exposing any real patient’s identity.

Synthetic data can also help reduce bias and improve AI fairness. Real-world datasets sometimes suffer from imbalance – for instance, a facial recognition dataset might have fewer examples of certain demographic groups, leading to biased model performance. With synthetic generation, we can create additional data for underrepresented cases or remove biases present in the original data. By augmenting data in this way, synthetic data can ensure a model sees a more balanced, diverse set of examples. As a result, training on synthetic data (or a mix of real and synthetic) can actually lead to better model performance and fairness, addressing biases that existed in the real data. Tech analysts note that synthetic training data can enhance model performance by adding fresh variations and even injecting explainability, since one can generate data to test specific “what if” scenarios.

Beyond training models, what is synthetic data in machine learning practice? It’s a versatile tool throughout the ML lifecycle. Data scientists use synthetic data to prototype and experiment when real data is not yet available. It’s used for validating models against rare conditions (e.g. unusual combinations of inputs) that weren’t present in the original dataset. It’s also increasingly important in software testing for AI-driven systems (more on that soon). According to SAP’s AI team, synthetic data generation is expected to surpass the use of real data in AI models by 2030, given the need for better quality and privacy-preserving data sources. Synthetic data in AI essentially broadens what’s possible – enabling AI development even when real data is missing, sensitive, or biased.

Synthetic Data Generation: How Is Synthetic Data Created?

Now that we know what synthetic data is and why it’s valuable, you might wonder how to create synthetic data. Synthetic data generation is a broad term covering various techniques to produce artificial datasets. The approach can range from simple simulations to advanced AI-driven generation. Here are some common methods of how synthetic data is generated:

Simple Statistical Simulation 

One basic way to generate synthetic data is by sampling from known probability distributions. For example, if you know your real data follows a certain statistical distribution, you can randomly draw numbers from that distribution to create a fake dataset. This approach captures high-level statistics (like mean, variance) of real data. It’s fast and easy, but doesn’t always preserve complex relationships in the data. Still, for some cases (like simulating sensor noise or generating random dummy records), this might suffice.

Rule-based or Programmatic Generation 

Before the era of AI-driven synthesis, many practitioners created mock data using rules or random generators. For instance, to generate synthetic names and addresses, one might programmatically combine random first names and last names with real-looking addresses. This yields data with the structure of real data (e.g., a database of customers), but since it’s often random, it may lack realistic patterns. Rule-based synthetic data (a.k.a. mock data) is useful for testing but usually doesn’t capture the rich statistical correlations of real datasets.

Agent-Based Modeling and Simulation 

This technique is like creating a mini synthetic world. Agent-based models simulate the behaviors and interactions of autonomous “agents” (which could be people, vehicles, processes, etc.) following predefined rules. For example, to synthesize traffic data, you could simulate cars (agents) driving through a virtual city, following traffic rules, and interacting. The simulation can generate synthetic telemetry data, accident scenarios, congestion patterns, and so on. Agent-based simulation is great for scenarios where you want to model complex systems or create synthetic data for scenarios that are rare in real life. It’s used in epidemiology (simulating disease spread through interactions), economics, crowd simulations, and more. With modern tools (like Mesa in Python for agent modeling), creating these simulations has become more accessible.

Generative AI Models 

The most exciting developments in synthetic data generation come from AI itself – specifically, deep learning models that learn to mimic real data. Techniques such as generative adversarial networks (GANs), variational autoencoders (VAEs), and even transformer models can produce highly realistic synthetic data by learning from real examples. For instance, GANs have been used to create incredibly realistic synthetic images: a GAN trained on photographs can generate new images that look like real photos (even though the specific image never actually happened). The GAN does this through a clever two-network system – a generator tries to create fake data and a discriminator tries to tell fake from real, and they both improve until the fakes are indistinguishable. VAEs, on the other hand, encode data into a latent representation and then decode it to produce new samples, which can also yield realistic variations of the original data. Transformer-based generative models (like GPT for text) can produce synthetic text or even synthetic tabular data by leveraging patterns learned from large datasets.

Hybrid Approaches 

Sometimes, synthetic data generation might involve mixing real and fake. Partially synthetic data replaces sensitive parts of real data with generated values (useful for privacy where you keep some real context but mask personal identifiers). Hybrid synthetic data might combine real records with synthetic records to boost dataset size while keeping some real examples in the mix. These approaches ensure the structure of data stays identical to the original while sensitive details are swapped out.

Regardless of method, how to generate synthetic data for machine learning typically involves an iterative process: you often start with a real dataset (even a small sample), train a generative model on it to learn its patterns, and then use that model to create new data. The Synthetic Data Vault (SDV), an open-source library, is one example of a tool that automates this process for tabular data using various generative algorithms. Many AI companies, including startups and big cloud vendors, now offer synthetic data generation platforms where you feed in your real data (which stays private), and the tool outputs a synthetic dataset you can use freely.

Which Two Requirements Must AI-Generated Synthetic Data Have?

Speaking of requirements, you might have heard this question: Which two requirements must synthetic data generated from AI have? It’s a great way to boil down the essentials of good synthetic data. The answer is exactly what we hinted above: 

1. Statistically identical to the training data: This means the synthetic data should mimic the statistical properties of the real dataset. For example, if you calculate things like averages, variances, distributions of values, or correlations between features, those should be very close between the synthetic data and the real data. Essentially, the synthetic data means the same thing as the original in aggregate – it’s as if you sampled more data from the same underlying reality. This ensures any model trained on synthetic data learns the correct patterns.
2. Structurally identical to the training data: The synthetic data must have the same structure, format, and logical layout as the real data. If the real data has 10 columns (age, gender, income, etc.), the synthetic set should have the same columns with the same types of values and relationships. If the real data was images of 256×256 pixels, the synthetic images should also be 256×256 and in the same color format, etc. Structural identity ensures that synthetic data can be used in place of real data seamlessly in whatever systems or analyses you have – your code, database, or ML model sees no difference in schema or shape.

These two requirements imply what synthetic data is not: it shouldn’t be just random gibberish, and it shouldn’t be a carbon copy of real data. A synthetic dataset where each synthetic record is actually just a real record (i.e., containing original data points) would violate both conditions – it’d be cheating on privacy and not really synthetic at all. On the other hand, synthetic data that is completely random might preserve format (e.g., numbers in the right range) but would fail the statistical similarity test, making it useless for training models.

In practice, achieving these requirements involves careful validation. Data scientists generate synthetic data and then compare distributions and model performance. A common approach is to train an ML model on real data and another model on the synthetic data and see if they perform similarly on a validation task – if yes, the synthetic data has effectively captured the needed patterns (statistical goodness) without being exactly the same data (since the models are different). The mantra is often that good synthetic data is “as good as real” in analysis, but “as safe as fake” in terms of privacy.

It’s also worth noting that preventing leakage of real data (to meet those requirements) can be challenging. AI generators have to avoid overfitting – if an AI model memorizes the training data too much, it might spit out something very close to a real record, which we don’t want. Top-tier synthetic data generators incorporate techniques to detect and mitigate this, ensuring that while synthetic data is statistically and structurally faithful, it does not inadvertently include or reveal any actual data points from the training set.

Synthetic Test Data: What Is It and How Is It Used?

When discussing synthetic data, you’ll often hear the term synthetic test data. This refers to artificial data specifically created for testing purposes – typically in software development, quality assurance, or system testing. What is synthetic test data? It’s basically fake data that imitates real production data, used to test applications in a realistic way without using actual user or production records.

In many organizations, testing new software features or performing QA requires a dataset that resembles the real data in production. Using a copy of production data can be risky (you might expose private info) and cumbersome (production databases can be huge, and copying them is slow and expensive). Synthetic test data provides a clever solution: you generate a dummy dataset that has the same shape and characteristics as production, and use that in your test environment. Because it’s synthetic, there’s no sensitive info, and you can make it as large or as varied as needed.

According to experts, synthetic test data is easier to create and more flexible than traditional manual or rule-based test data approaches. It offers realism (so your tests are valid), scalability (you can generate millions of records if needed), and safety (no real personal data included). This is crucial for modern data-driven testing and continuous integration pipelines in DevOps.

Another common use of synthetic test data is in stress testing and scenario testing. You can create extreme or rare conditions in synthetic data to see how your system handles them. For instance, an e-commerce site might generate a spike of synthetic orders to test system load, or a cybersecurity team might generate synthetic logs with various attack patterns to test intrusion detection systems. With real data alone, you might never hit these edge cases until it’s too late.

Benefits of Using Synthetic Data

Why are organizations investing in synthetic data generation? Let’s highlight the main benefits of using synthetic datafor businesses and AI practitioners:

Data Privacy and Compliance

Perhaps the biggest driver of synthetic data adoption is privacy. Synthetic data that mirrors customer or patient information contains no real personal identifiable information (PII), so it poses far less risk under laws like GDPR or HIPAA. This allows companies to share and use data more freely. For example, a healthcare researcher can use synthetic patient records to develop a diagnostic AI tool without ever exposing a real patient’s record – avoiding legal hurdles while still gaining insights. In the words of the European Data Protection Supervisor, “Synthetic data is a technical solution to a legal problem,” allowing innovation without compromising privacy. By using synthetic data, organizations can comply with strict data protection regulations (GDPR, CCPA, etc.) while still leveraging rich datasets for analysis.

Unlimited Data Volume and Scalability 

With synthetic data, the quantity of data is no longer a hard constraint. Need more data to train your model? Generate more. Synthetic data allows scaling datasets to practically any size on demand. Traditional data collection often hits limits – you can only collect so many real examples, or it may be expensive to gather more. Synthetic generation can fill the gap, providing thousands or millions of additional records as needed. This scalability empowers ML projects because larger training sets often mean better models. It also means you can simulate growth or expansion scenarios: e.g., an IoT platform can generate what 10x more devices worth of data would look like, to test future capacity.

Improved Data Diversity and Quality 

Synthetic data can be crafted to represent a wider range of scenarios and edge cases than what you currently have. Real-world datasets might be narrow or biased – for instance, if you only have historical sales data from summer, your model might not know what winter looks like. With synthetic generation, you can introduce new variations (simulate winter sales) to make the dataset more diverse and representative. This ultimately improves model accuracy and robustness, as the model learns from a more comprehensive dataset. It also helps in risk assessment and rare event simulation – you can generate that unusual scenario (financial crisis data, extreme customer behavior, etc.) to ensure your systems handle it well. Deloitte has noted that generating synthetic data opens up opportunities to simulate different “what-if” risk scenarios, so organizations can test and refine strategies before those situations ever occur in reality.

Cost and Time Efficiency

Acquiring real data can be expensive – whether it’s paying for data collection, data labeling, or running lengthy experiments. Synthetic data is often cheaper and faster. As noted earlier, synthetically generating and labeling an image can be cents on the dollar compared to manual efforts. Similarly, consider industries like automotive: crash-testing vehicles to collect safety data is hugely costly and destructive, whereas running thousands of simulated crash scenarios (synthetic data) costs a fraction and can be done much faster. Synthetic data also speeds up data provisioning – instead of waiting weeks to get approval and access to a sensitive dataset, a team can generate a synthetic version in hours and start working immediately. Overall, this accelerates AI development cycles and reduces the time to market for new models and products.

Enhanced Data Labeling and Annotation 

In many ML tasks, especially in computer vision, not only gathering data but labeling it is a huge bottleneck. Synthetic data often comes with labels “for free.” For example, if you use a 3D simulator to generate synthetic images (say of cars on a road), the simulator knows exactly where each car is, what each object is, and can output perfect annotations (bounding boxes, segmentation masks, etc.). This means no human annotators are needed – every synthetic image can have complete ground-truth labels automatically. The TechTarget report highlighted this as a major reason why synthetic data can be so inexpensive: it eliminates the need for manual data labeling because the generation process itself provides all the metadata. For companies working on AI that requires labeled data (like image recognition, object detection, NLP tasks), synthetic data can dramatically cut down the effort in labeling, while also avoiding human error in the labels.

Testing and Experimentation without Risk

Synthetic data allows teams to experiment freely. Because it’s artificially made, you can try out bold ideas on the data, transform it in ways you might not with real data, and share it among teams without red tape. Want to test a new analytics tool on customer data but afraid of a breach? Use synthetic customer data and there’s no harm if something leaks. Want to demonstrate your software to a client but not show real user data? Synthetic data to the rescue – it’s great for demos and training environments where realistic data makes the demo meaningful, but you can’t use production data. Additionally, synthetic data can be used to stress-test models – you can introduce controlled noise or outliers to see how robust your model is, all in a safe sandbox.

Facilitating Collaboration and Data Sharing 

Since synthetic data contains no real sensitive information, it’s much easier to share with partners, vendors, or across different departments of an organization. This improves collaboration and innovation. For instance, a bank’s compliance department might normally refuse to share customer data with an outside fintech consultant. But if that data is synthesized (preserving utility but anonymized), they could share it and gain insights from the consultant. Synthetic data thus breaks down data silos – teams can work together on data science projects without lengthy approval processes because the data is de-identified by design. According to a case study with Deloitte, synthetic data’s privacy-preserving nature enabled better knowledge sharing between departments and even allowed using powerful cloud analytics tools that were off-limits for raw sensitive data. In essence, synthetic data can unlock use of modern AI and cloud services for industries that were stuck with data on local, closed systems due to privacy – now they can leverage the latest tech on synthetic datasets without compromising security.

Better Utilization of AI/ML Resources 

Sometimes data scientists spend more time wrangling or waiting for data than building models. Synthetic data can flip that equation by providing clean, ready-to-use data when needed. It also offers full control over the data characteristics. Practitioners can adjust class balance, add noise, or create specific edge cases by instructing the data generator. This level of control is like giving data scientists a sandbox to fine-tune their training data for optimal results, something not possible with fixed real datasets. Gartner has noted that synthetic data generation “accelerates the analytics development cycle, lessens regulatory concerns and lowers the cost of data acquisition.” In other words, it makes AI development more efficient on multiple fronts.

All these benefits explain why synthetic data is often called a game-changer for AI. However, it’s also important to be aware of its limitations. Poorly generated synthetic data can mislead models (garbage in, garbage out). It might fail to capture subtle complexities of real data, or inadvertently introduce its own biases. Also, synthetic data is not a drop-in replacement in all cases – you still need some real data to ensure validity, and completely replacing real data with synthetic in critical systems requires careful validation. Many experts see synthetic data as a supplement rather than a total substitute: it works best when combined with real data to augment and enhance it.

Deloitte’s 2024 insights also emphasize synthetic data as a way to fill data gaps in simulations and modeling, enabling projects that would have stalled due to lack of data. For businesses, this means new opportunities – using AI in areas that previously had too little data or too many privacy restrictions to proceed.

Examples of Synthetic Data in Action

Concrete examples help bring all these concepts to life. Many leading companies and institutions are already using synthetic data. Here are a few notable synthetic data generation examples and use cases across industries:

• Financial Services (Fraud Detection & Sandboxing): Major banks like J.P. Morgan and American Express have turned to synthetic data to improve fraud detection algorithms. Detecting fraud (e.g., credit card fraud or money laundering) is tricky because fraudulent transactions are relatively rare and often hidden in mounds of normal transactions. By generating synthetic fraudulent transactions and scenarios, banks can train AI models to recognize suspicious patterns more effectively. J.P. Morgan even developed a synthetic data “sandbox” – a safe environment with synthetic transactional data – to speed up data-intensive proofs of concept with third-party vendors. This allows them to innovate with fintech partners without exposing real customer data, accelerating development of new services.
• Healthcare (Research & Medical AI): Healthcare providers and researchers use synthetic patient data to advance AI for diagnosis, treatment, and operations. For example, researchers can create a synthetic dataset of electronic health records that statistically mirrors real patient populations. AI models trained on this can help predict disease outbreaks or test treatment strategies. Because actual medical data is highly sensitive, synthetic data is a boon – it lets researchers collaborate and publish findings without risking patient privacy. There are also open synthetic databases that include demographics and health stats designed to represent a population without the biases or privacy issues of real data. This is extremely useful for public health simulations (like planning emergency responses using synthetic population data that reflects real-world diversity).
• Insurance (Customer Analytics): Insurance companies deal with stringent privacy and often limited data for rare events (like large claims, disasters, etc.). A European insurance group, Provinzial, used synthetic data to fully exploit their customer data for predictive analytics. By synthesizing customer information, they could apply advanced AI models to identify customer needs and tailor services, all while staying compliant with data protection laws. Similarly, insurers use synthetic data to simulate scenarios like catastrophe modeling (what if a once-in-a-century flood occurs?) to better prepare and price risk. With synthetic data, they can generate many hypothetical events and see how portfolios would be affected, which is invaluable for risk management.
• Telecommunications (Customer Behavior Modeling): Telecom companies often sit on heaps of usage data but can’t leverage it fully due to privacy. Vodafone, for instance, used synthetic data for training and testing machine learning models in customer value management. By synthesizing telecom data (calls, messages, data usage patterns), they were able to improve models that predict customer churn and identify which customers might want which new service, etc. The synthetic datasets allowed them to do this faster, saving time and costs, and even improved model performance by providing a richer variety of training examples.
• Automotive (Autonomous Driving & Safety): As discussed, the autonomous vehicle industry is one of the heaviest users of synthetic data. Companies like Nvidia provide simulation platforms where car makers generate photorealistic virtual streets to train their AI. Synthetic driving data includes not just camera images, but also Lidar scans, radar signals, and even simulated pedestrians and traffic events. By training on millions of synthetic miles, self-driving car AI can encounter situations that would be too dangerous or rare to rely on with real-world driving alone. Tesla is reported to use simulation for testing its Autopilot system’s responses to unusual scenarios (though Tesla also collects a huge amount of real data from its fleet). The combination of real and synthetic data is helping push autonomous driving forward more quickly and safely.
• Government & Public Sector: Even government agencies find value in synthetic data. For instance, the U.S. Census Bureau has explored synthetic data techniques to publish statistical data without revealing any individual’s information (a form of privacy protection on public data). Smart cities projects also use synthetic data to model urban mobility or energy usage when they lack sensors everywhere. Synthetic data can simulate how traffic flows or how electricity is consumed in a city under various conditions, assisting in urban planning.
• Manufacturing & IoT: In industrial settings, sometimes there’s not enough data on equipment failures or rare production defects. By generating synthetic sensor data or using digital twins (virtual replicas) of machines, companies can simulate faults and failure scenarios to train predictive maintenance models. For example, a power grid operator might not have many examples of a particular component failing (a rare event), so they create a physics-based simulation to generate synthetic images of defective equipment. Deloitte reported a case where a utility company generated over 2,000 synthetic images of power grid defects (like frayed cables, damaged insulators, etc.) to train their computer vision algorithm, leading to a 67% improvement in defect detection accuracy. This is a powerful testament to synthetic data’s ability to fill critical gaps – the algorithm got much better because it was finally trained on the problems it needed to catch, courtesy of synthetic examples.

These examples scratch the surface, but they demonstrate a common theme: synthetic data enables progress where real data is limited. Whether the limitation is privacy, rarity, cost, or time, synthetic data provides a workaround. Gartner’s research has consistently highlighted synthetic data as a key trend in data and analytics strategies, forecasting that organizations effectively using synthetic data will outperform those that don’t in AI development. With generative AI techniques becoming more advanced, the realism of synthetic data is only increasing – we now have synthetic faces that humans cannot distinguish from real, synthetic voice that sounds natural, and synthetic financial data that fools even expert analysts in blind tests.

Sphere’s View

Many companies are now exploring synthetic data as part of their AI strategy — and with good reason. It’s a practical answer to data access issues, privacy regulations, and the need for scale. But in our experience, success with synthetic data depends less on the hype and more on making the right foundational choices.

From our work at Sphere, we’ve learned that the right partner for synthetic data projects should have three core characteristics:

• A deep understanding of real-world data complexity. It’s not just about generating data that looks plausible — it’s about preserving structure, statistical relevance, and context so your models actually learn what matters.
• Hands-on experience across different use cases. Whether it’s generating synthetic test data for QA environments or training datasets for rare-event classification, the nuances vary — and your partner needs to know how to adapt generation techniques accordingly.
• A pragmatic approach to integration. Synthetic data alone won’t solve broken pipelines or poor model evaluation. The real value comes from knowing how to plug synthetic datasets into broader AI workflows — in a way that supports governance, auditability, and performance.

We don’t see synthetic data as a one-size-fits-all solution. But when it fits, it can be a powerful tool — especially when combined with a clear understanding of where it adds value, where it doesn’t, and how to get the most out of it.

So if you’re looking into synthetic data, don’t just ask whether it’s possible. Ask how it will be used, who will create it, and what standards it needs to meet. That’s where the difference lies — not in the generation itself, but in the thinking behind it.

Frequently Asked Questions