Read the full article below (in Italian):
The larger your database, the higher the possibility of data repetition and inaccuracies that compromise the results you pull from the database. Normalization in DBMS exists to counteract those problems by helping you to create more uniform databases in which redundancies are less likely to occur.
Mastering normalization is a key skill in DBMS for the simple fact that an error-strewn database is of no use to an organization. For example, a retailer that has to deal with a database that has multiple entries for phone numbers and email addresses is a retailer that can’t see as effectively as one that has a simple route to the customer. Let’s look at normalization in DBMS and how it helps you to create a more organized database.
The Concept of Normalization
Grab a pack of playing cards and throw them onto the floor. Now, pick up the “Jack of Hearts.” It’s a tough task because the cards are strewn all over the place. Some are facing down and there’s no rhyme, reason, or pattern to how the cards lie, meaning you’re going to have to check every card individually to find the one you want.
That little experiment shows you how critical organization is, even with a small set of “data.” It also highlights the importance of normalization in DBMS. Through normalization, you implement organizational controls using a set of principles designed to achieve the following:
- Eliminate redundancy – Lower (or eliminate) occurrences of data repeating across different tables, or inside individual tables, in your DBMS.
- Minimize data anomalies – Better organization makes it easier to spot datasets that don’t fit the “norm,” meaning fewer anomalies.
- Improve data integrity – More accurate data comes from normalization controls. Database users can feel more confident in their results because they know that the controls ensure integrity.
The Process of Normalization
If normalization in DBMS is all about organization, it stands to reason that they would be a set process to follow when normalizing your tables and database:
- Decompose your tables – Break every table down into its various parts, which may lead to you creating several tables out of one. Through decomposition, you separate different datasets, eliminate inconsistencies, and set the stage for creating relationships and dependencies between tables.
- Identify functional dependencies – An attribute in one table may be dependent on another to exist. For example, a “Customer ID” number in a retailer’s “Customer” table is functionally dependent on the “Customer Name” field because the ID can’t exist without the customer. Identifying these types of dependencies ensures you don’t end up with empty records (such as a record with a “Customer ID” and no customer attached to it).
- Apply normalization rules – Once you’re broken down your table and identified the functional dependencies, you apply relevant normalization rules. You’ll use Normal Forms to do this, with the six highlighted below each having its own rules, structures, and use cases.
Normal Forms in DBMS
There isn’t a “single” way to achieve normalization in DBMS because every database (and the tables it contains) is different. Instead, there are six normal forms you may use, with each having its own rules that you need to understand to figure out which to apply.
First Normal Form (1NF)
If a relation can’t contain multiple values, it’s in 1NF. In other words, each attribute in the table can only contain a single (called “atomic”) value.
Example
If a retailer wants to store the details of its customers, it may have attributes in its table like “Customer Name,” “Phone Number,” and “Email Address.” By applying 1NF to this table, you ensure that the attributes that could contain multiple entries (“Phone Number” and “Email Address”) only contain one, making contacting that customer much simpler.
Second Normal Form (2NF)
A table that’s in 2NF is in 1NF, with the additional condition that none of its non-prime attributes depend on a subset of candidate keys within the table.
Example
Let’s say an employer wants to create a table that contains information about an employee, the skills they have, and their age. An employee may have multiple skills, leading to multiple records for the same employee in the table, with each denoting a skill while the ID number and age of the employee repeat for each record.
In this table, you’ve achieved 1NF because each attribute has an atomic value. However, the employee’s age is dependent on the employee ID number. To achieve 2NF, you’d break this table down into two tables. The first will contain the employee’s ID number and age, with that ID number linking to a second table that lists each of the skills associated with the employee.
Third Normal Form (3NF)
In 3NF, the table you have must already be in 2NF form, with the added rule of removing the transitive functional dependency of the non-prime attribute of any super key. Transitive functional dependency occurs if the dependency is the result of a pair of functional dependencies. For example, the relationship between A and C is a transitive dependency if A depends on B, B depends on C, but B doesn’t depend on A.
Example
Let’s say a school creates a “Students” table with the following attributes:
- Student ID
- Name
- Zip Code
- State
- City
- District
In this case, the “State,” “District,” and “City” attributes all depend on the “Zip Code” attribute. That “Zip” attribute depends on the “Student ID” attribute, making “State,” “District,” and “City” all transitively depending on “Student ID.”
To resolve this problem, you’d create a pair of tables – “Student” and “Student Zip.” The “Student” table contains the “Student ID,” “Name,” and “Zip Code” attributes, with that “Zip Code” attribute being the primary key of a “Student Zip” table that contains the rest of the attributes and links to the “Student” table.
Boyce-Codd Normal Form (BCNF)
Often referred to as 3.5NF, BCNF is a stricter version of 3NF. So, this normalization in DBMS rule occurs if your table is in 3NF, and for every functional dependence between two fields (i.e., A -> B), A is the super key of your table.
Example
Sticking with the school example, every student in a school has multiple classes. The school has a table with the following fields:
- Student ID
- Nationality
- Class
- Class Type
- Number of Students in Class
You have several functional dependencies here:
- Student ID -> Nationality
- Class -> Number of Students in Class, Class Type
As a result, both the “Student ID” and “Class” attributes are candidate keys but can’t serve as keys alone. To achieve BCNF normalization, you’d break the above table into three – “Student Nationality,” “Student Class,” and “Class Mapping,” allowing “Student ID” and “Class” to serve as primary keys in their own tables.
Fourth Normal Form (4NF)
In 4NF, the database must meet the requirements of BCNF, in addition to containing no more than a single multivalued dependency. It’s often used in academic circles, as there’s little use for 4NF elsewhere.
Example
Let’s say a college has a table containing the following fields:
- College Course
- Lecturer
- Recommended Book
Each of these attributes is independent of the others, meaning each can change without affecting the others. For example, the college could change the lecturer of a course without altering the recommended reading or the course’s name. As such, the existence of the course depends on both the “Lecturer” and “Recommended Book” attributes, creating a multivalued dependency. If a DBMS has more than one of these types of dependencies, it’s a candidate for 4NF normalization.
Fifth Normal Form (5NF)
If your table is in 4NF, has no join dependencies, and all joining is lossless, it’s in 5NF. Think of this as the final form when it comes to normalization in DBMS, as you’ve broken your table down so much that you’ve made redundancy impossible.
Example
A college may have a table that tells them which lecturers teach certain subjects during which semesters, creating the following attributes:
- Subject
- Lecturer Name
- Semester
Let’s say one of the lecturers teaches both “Physics” and “Math” for “Semester 1,” but doesn’t teach “Math” for Semester 2. That means you need to combine all of the fields in this table to get an accurate dataset, leading to redundancy. Add a third semester to the mix, especially if that semester has no defined courses or lecturers, and you have to join dependencies.
The 5NF solution is to break this table down into three tables:
- Table 1 – Contains the “Semester” and “Subject” attributes to show which subjects are taught in each semester.
- Table 2 – Contains the “Subject” and “Lecturer Name” attributes to show which lecturers teach a subject.
- Table 3 – Contains the “Semester” and “Lecturer Name” attributes so you can see which lecturers teach during which semesters.
Benefits of Normalization in DBMS
With normalization in DBMS being so much work, you need to know the following benefits to show that it’s worth your effort:
- Improved database efficiency
- Better data consistency
- Easier database maintenance
- Simpler query processing
- Better access controls, resulting in superior security
Limitations and Trade-Offs of Normalization
Normalization in DBMS does have some drawbacks, though these are trade-offs that you accept for the above benefits:
- The larger your database gets, the more demands it places on system performance.
- Breaking tables down leads to complexity.
- You have to find a balance between normalization and denormalization to ensure your tables make sense.
Practical Tips for Mastering Normalization Techniques
Getting normalization in DBMS is hard, especially when you start feeling like you’re dividing tables into so many small tables that you’re losing track of the database. These tips help you apply normalization correctly:
- Understand the database requirements – Your database exists for you to extract data from it, so knowing what you’ll need to extract indicates whether you need to normalize tables or not.
- Document all functional dependencies – Every functional dependence that exists in your database makes the table in which it exists a candidate for normalization. Identify each dependency and document it so you know whether you need to break the table down.
- Use software and tools – You’re not alone when poring through your database. There are plenty of tools available that help you to identify functional dependencies. Many make normalization suggestions, with some even being able to carry out those suggestions for you.
- Review and refine – Every database evolves alongside its users, so continued refining is needed to identify new functional dependencies (and opportunities for normalization).
- Collaborate with other professionals – A different set of eyes on a database may reveal dependencies and normalization opportunities that you don’t see.
Make Normalization Your New Norm
Normalization may seem needlessly complex, but it serves the crucial role of making the data you extract from your database more refined, accurate, and free of repetition. Mastering normalization in DBMS puts you in the perfect position to create the complex databases many organizations need in a Big Data world. Experiment with the different “normal forms” described in this article as each application of the techniques (even for simple tables) helps you get to grips with normalization.
-
From the same author
Read the articleAs one of the world’s fastest-growing industries, with a predicted compound annual growth rate of 16.43% anticipated between 2022 and 2030, data science is the ideal choice for your career. Jobs will be plentiful. Opportunities for career advancement will come thick and fast. And even at the most junior level, you’ll enjoy a salary that comfortably sits in the mid-five figures.
Studying for a career in this field involves learning the basics (and then the complexities) of programming languages including C+, Java, and Python. The latter is particularly important, both due to its popularity among programmers and the versatility that Python brings to the table. Here, we explore the importance of Python for data science and how you’re likely to use it in the real world.
Why Python for Data Science?
We can distill the reasons for learning Python for data science into the following five benefits.
Popularity and Community Support
Statista’s survey of the most widely-used programming languages in 2022 tells us that 48.07% of programmers use Python to some degree. Leftronic digs deeper into those numbers, telling us that there are 8.2 million Python developers in the world. As a prospective developer yourself, these numbers tell you two things – Python is in demand and there’s a huge community of fellow developers who can support you as you build your skills.
Easy to Learn and Use
You can think of Python as a primer for almost any other programming language, as it takes the fundamental concepts of programming and turns them into something practical. Getting to grips with concepts like functions and variables is simpler in Python than in many other languages. Python eventually opens up from its simplistic use cases to demonstrate enough complexity for use in many areas of data science.
Extensive Libraries and Tools
Given that Python was first introduced in 1991, it has over 30 years of support behind it. That, combined with its continued popularity, means that novice programmers can access a huge number of tools and libraries for their work. Libraries are especially important, as they act like repositories of functions and modules that save time by allowing you to benefit from other people’s work.
Integration With Other Programming Languages
The entire script for Python is written in C, meaning support for C is built into the language. While that enables easy integration between these particular languages, solutions exist to link Python with the likes of C++ and Java, with Python often being capable of serving as the “glue” that binds different languages together.
Versatility and Flexibility
If you can think it, you can usually do it in Python. Its clever modular structure, which allows you to define functions, modules, and entire scripts in different files to call as needed, makes Python one of the most flexible programming languages around.
Setting Up Python for Data Science
Installing Python onto your system of choice is simple enough. You can download the language from the Python.org website, with options available for everything from major operating systems (Windows, macOS, and Linux) to more obscure devices.
However, you need an integrated development environment (IDE) installed to start coding in Python. The following are three IDEs that are popular with those who use Python for data science:
- Jupyter Notebook – As a web-based application, Jupyter easily allows you to code, configure your workflows, and even access various libraries that can enhance your Python code. Think of it like a one-stop shop for your Python needs, with extensions being available to extend its functionality. It’s also free, which is never a bad thing.
- PyCharm – Where Jupyter is an open-source IDE for several languages, PyCharm is for Python only. Beyond serving as a coding tool, it offers automated code checking and completion, allowing you to quickly catch errors and write common code.
- Visual Studio Code – Though Visual Studio Code alone isn’t compatible with Python, it has an extension that allows you to edit Python code on any operating system. Its “Linting” feature is great for catching errors in your code, and it comes with an integrated debugger that allows you to test executables without physically running them.
Setting up your Python virtual environment is as simple as downloading and installing Python itself, and then choosing an IDE in which to work. Think of Python as the materials you use to build a house, with your IDE being both the blueprint and the tools you’ll need to patch those materials together.
Essential Python Libraries for Data Science
Just as you’ll go to a real-world library to check out books, you can use Python libraries to “check out” code that you can use in your own programs. It’s actually better than that because you don’t need to return libraries when you’re done with them. You get to keep them, along with all of their built-in modules and functions, to call upon whenever you need them. In Python for data science, the following are some essential libraries:
- NumPy – We spoke about integration earlier, and NumPy is ideal for that. It brings concepts of functionality from Fortran and C into Python. By expanding Python with powerful array and numerical computing tools, it helps transform it into a data science powerhouse.
- pandas – Manipulating and analyzing data lies at the heart of data sciences, and pandas give you a library full of tools to allow both. It offers modules for cleaning data, plotting, finding correlations, and simply reading CSV and JSON files.
- Matplotlib – Some people can look at reams of data and see patterns form within the numbers. Others need visualization tools, which is where Matplotlib excels. It helps you create interactive visual representations of your data for use in presentations or if you simply prefer to “see” your data rather than read it.
- Scikit-learn – The emerging (some would say “exploding) field of machine learning is critical to the AI-driven future we’re seemingly heading toward. Scikit-learn is a library that offers tools for predictive data analysis, built on what’s available in the NumPy and Matplotlib libraries.
- TensorFlow and Keras – Much like Scikit-learn, both TensorFlow and Keras offer rich libraries of tools related to machine learning. They’re essential if your data science projects take you into the realms of neural networks and deep learning.
Data Science Workflow in Python
A Python programmer without a workflow is like a ship’s captain without a compass. You can sail blindly onward, and you may even get lucky and reach your destination, but the odds are you’re going to get lost in the vastness of the programming sea. For those who want to use Python for data science, the following workflow brings structure and direction to your efforts.
Step 1 – Data Collection and Preprocessing
You need to collect, organize, and import your data into Python (as well as clean it) before you can draw any conclusions from it. That’s why the first step in any data science workflow is to prepare the data for use (hint – the pandas library is perfect for this task).
Step 2 – Exploratory Data Analysis (EDA)
Just because you have clean data, that doesn’t mean you’re ready to investigate what that data tells you. It’s like washing ingredients before you make a dish – you need to have a “recipe” that tells you how to put everything together. Data scientists use EDA as this recipe, allowing them to combine data visualization (remember – the Matplotlib library) with descriptive statistics that show them what they’re looking at.
Step 3 – Feature Engineering
This is where you dig into the “whats” and “hows” of your Python program. You’ll select features for the code, which define what it does with the data you import and how it’ll deliver outcomes. Scaling is a key part of this process, with scope creep (i.e., constantly adding features as you get deeper into a project) being the key thing to avoid.
Step 4 – Model Selection and Training
Decision trees, linear regression, logistic regression, neural networks, and support vector machines. These are all models (with their own algorithms) you can use for your data science project. This step is all about selecting the right model for the job (your intended features are important here) and training that model so it produces accurate outputs.
Step 5 – Model Evaluation and Optimization
Like a puppy that hasn’t been house trained, an unevaluated model isn’t ready for release into the real world. Classification metrics, such as a confusion matrix and classification report, help you to evaluate your model’s predictions against real-world results. You also need to tune the hyperparameters built into your model, similar to how a mechanic may tune the nuts and bolts in a car, to get everything working as efficiently as possible.
Step 6 – Deployment and Maintenance
You’ve officially deployed your Python for data science model when you release it into the wild and let it start predicting outcomes. But the work doesn’t end at deployment, as constant monitoring of what your model does, outputs, and predicts is needed to tell you if you need to make tweaks or if the model is going off the rails.
Real-World Data Science Projects in Python
There are many examples of Python for data science in the real world, some of which are simple while others delve into some pretty complex datasets. For instance, you can use a simple Python program to scrap live stock prices from a source like Yahoo! Finance, allowing you to create a virtual ticker of stock price changes for investors.
Alternatively, why not create a chatbot that uses natural language processing to classify and respond to text? For that project, you’ll tokenize sentences, essentially breaking them down into constituent words called “tokens,” and tag those tokens with meanings that you could use to prompt your program toward specific responses.
There are plenty of ideas to play around with, and Python is versatile enough to enable most, so consider what you’d like to do with your program and then go on the hunt for datasets. Great (and free) resources include The Boston House Price Dataset, ImageNet, and IMDB’s movie review database.
Try Python for Data Science Projects
By combining its own versatility with integrations and an ease of use that makes it welcoming to beginners, Python has become one of the world’s most popular programming languages. In this introduction to data science in Python, you’ve discovered some of the libraries that can help you to apply Python for data science. Plus, you have a workflow that lends structure to your efforts, as well as some ideas for projects to try. Experiment, play, and tweak models. Every minute you spend applying Python to data science is a minute spent learning a popular programming language in the context of a rapidly-growing industry.
Read the articleToday’s tech-driven world is governed by data – so much so that nearly 98% of all organizations are increasing investment in data.
However, company owners can’t put their feet up after improving their data capabilities. They also need a database management system (DBMS) – a program specifically designed for storing and organizing information efficiently.
When analyzing a DBMS, you need to be thorough like a detective investigating a crime. One of the elements you want to consider is DBMS architecture. It describes the structure of your database and how individual bits of information are related to each other. The importance of DBMS architecture is enormous, as it helps IT experts design and maintain fully functional databases.
But what exactly does a DBMS architecture involve? You’ll find out in this entry. Coming up is an in-depth discussion of database system concepts and architecture.
Overview of DBMS Architecture
Suppose you’re assembling your PC. You can opt for several configurations, such as those with three RAM slots and dual-fan coolers. The same principle applies to DBMS architectures.
Two of the most common architectures are three-level and two-level architectures.
Three-Level Architecture
Three-level architecture is like teacher-parent communication. More often than not, a teacher communicates with parents through children, asking them to convey certain information. In other words, there are layers between the two that don’t allow direct communication.
The same holds for three-level architecture. But instead of just one layer, there are two layers between the database and user: application client and application server.
And as the name suggests, a three-level DBMS architecture has three levels:
- External level – Also known as the view level, this section concerns the part of your database that’s relevant to the user. Everything else is hidden.
- Conceptual level – Put yourself in the position of a scuba diver exploring the ocean layer by layer. Once you reach the external level, you go one segment lower and find the conceptual level. It describes information conceptually and tells you how data segments interact with one another.
- Internal level – Another name for the internal level is the physical level. But what does it deal with? It mainly focuses on how data is stored in your system (e.g., using folders and files).
Two-Level Architecture
When you insert a USB into your PC, you can see the information on your interface. However, the source of the data is on the USB, meaning they’re separated.
Two-level architecture takes the same approach to separating data interface and data structure. Here are the two levels in this DBMS architecture:
- User level – Any application and interface in your database are stored on the user level in a two-level DBMS architecture.
- System level – The system level (aka server level) performs transaction management and other essential processes.
Comparison of the Two Architectures
Determining which architecture works best for your database is like buying a car. You need to consider how easy it is to use and the level of performance you can expect.
On the one hand, the biggest advantage of two-level architectures is that they’re relatively easy to set up. There’s just one layer between the database and the user, resulting in easier database management.
On the other hand, developing a three-level DBMS architecture may take a while since you need to include two layers between the database and the user. That said, three-level architectures are normally superior to two-level architectures due to higher flexibility and the ability to incorporate information from various sources.
Components of DBMS Architecture
You’ve scratched the surface of database system concepts and architecture, but don’t stop there. It’s time to move on to the basics to the most important elements of a DBMS architecture:
Data Storage
The fact that DBMS architectures have data storage solutions is carved in stone. What exactly are those solutions? The most common ones are as follows:
- Data files – How many files do you have on your PC? If it’s a lot, you’re doing exactly what administrators of DBMS architectures are doing. A large number of them store data in files, and each file is categorized into blocks.
- Indexes – You want your database operations to be like lightning bolts, i.e. super-fast. You can incorporate indexes to accomplish this goal. They point to data columns for quick retrieval.
- Data dictionary – Also known as system logs, data dictionaries contain metadata – information about your data.
Data Manipulation
A large number of companies still utilize manual data management methods. But using this format is like shooting yourself in the foot when there are advanced data manipulation methods are available. These allow you to process and retrieve data within seconds through different techniques:
- Query processor – Query processing refers to extracting data from your DBMS architecture. It operates like any other multi-stage process. It involves parsing, translation, optimization, and evaluation.
- Query optimizer – A DBMS architecture administrator can perform various query optimization tasks to achieve desired results faster.
- Execution engine – Whenever you want your architecture to do something, you send requests. But something needs to process the requests – that something is the execution engine.
Data Control
We’re continuing our journey through an average DBMS architecture. Our next stop is data control, which is comprised of these key elements:
- Transaction management – When carrying out multiple transactions, how does the system prioritize one over another? The answer lies in transaction management, which is also about processing multiple transactions side by side.
- Concurrency control – Database architecture is like an ocean teeming with life. Countless operations take place simultaneously. As a result, the system needs concurrency control to manage these concurrent tasks.
- Recovery management – What if your DBMS architecture fails? Do you give up on your project? No – the system has robust recovery management tools to retrieve your information and reduce downtime.
Database System Concepts
To give you a better understanding of a DBMS architecture, let’s describe the most important concepts regarding this topic.
Data Models
Data models do to information what your folders do to files – organize them. There are four major types of data models:
- Hierarchical model – Top-down and bottom-up storage solutions are known as hierarchical models. They’re characterized by tree-like structures.
- Network model – Hierarchical models are generally used for basic data relationships. If you want to analyze complex relationships, you need to kick things up a notch with network models. They enable you to represent huge quantities of complex information without a hitch.
- Relational model – Relations are merely tables with values. A relational model is a collection of these relations, indicating how data is connected to other data.
- Object-oriented model – Programming languages regularly use objects. An object-oriented model stores information as models and is usually more complex than other models.
Database Schema and Instances
Another concept you should familiarize yourself with is schemas and instances.
- Definition of schema and instance – Schemas are like summaries, providing a basic description of databases. Instances tell you what information is stored in a database.
- Importance of schema in DBMS architecture – Schemas are essential because they help organize data by providing a clear outline.
Data Independence
The ability of other pieces of information to remain unaffected after you change one bit of data is known as data independence. What are the different types of data independence, and what makes them so important?
- Logical data independence – If you can modify logical schemas without altering the rest of the system, your logical data is independent.
- Physical data independence – Physical data is independent if it remains unaffected when changing your hardware, such as SSD disks.
- Significance of data independence in DBMS architecture – Independent data is crucial for saving time in database management because it reduces the amount of information that needs to be processed.
Efficient Database Management Systems
Database management systems have a lot in common with other tech-based systems. For example, you won’t ignore problems that arise on your PC, be they CPU or graphics card issues. You’ll take action to optimize the performance of the device and solve those issues.
That’s exactly what 75% of developers and administrators of database management systems do. They go the extra mile to enhance the performance, scalability, flexibility, security, and integrity of their architecture.
Performance Optimization Techniques
- Indexing – By pointing to certain data in tables, indexes speed up database management.
- Query optimization – This process is about finding the most efficient method of executing queries.
- Caching – Frequently accessed information is cached to accelerate retrieval.
Scalability and Flexibility
- Horizontal scaling – Horizontal scaling involves increasing the number of servers.
- Vertical scaling – An administrator can boost the performance of the server to make the system more scalable.
- Distributed databases – Databases are like smartphones in that they can easily overload. Pressure can be alleviated with distributed databases, which store information in multiple locations.
Security and Integrity
- Access control – Restricting access is key to preventing cyber security attacks.
- Data encryption – Administrators often encrypt their DBMS architecture to protect sensitive information.
- Backup and recovery – A robust backup plan helps IT experts recover from shutdowns and other unforeseen problems.
Preparing for the Future Is Critical
DBMS architecture is the underlying structure of a database management system. It consists of several elements, all of which work together to create a fully functional data infrastructure.
Understanding the basic elements of DBMS architecture is vital for IT professionals who want to be well-prepared for future changes, such as hybrid environments. As the old saying goes – success depends upon preparation.
-
Go to all authors
Related posts
Source:
- Agenda Digitale, published on June 16th, 2025
By Lokesh Vij, Professor of Cloud Computing Infrastructure, Cloud Development, Cloud Computing Automation and Ops and Cloud Data Stacks at OPIT – Open Institute of Technology
NIST identifies five key characteristics of cloud computing: on-demand self-service, network access, resource pooling, elasticity, and metered service. These pillars explain the success of the global cloud market of 912 billion in 2025
You’ve probably seen two of the most recent popular social media trends. The first is creating and posting your personalized action figure version of yourself, complete with personalized accessories, from a yoga mat to your favorite musical instrument. There is also the Studio Ghibli trend, which creates an image of you in the style of a character from one of the animation studio’s popular films.
Both of these are possible thanks to OpenAI’s GPT-4o-powered image generator. But what are you risking when you upload a picture to generate this kind of content? More than you might imagine, according to Tom Vazdar, chair of cybersecurity at the Open Institute of Technology (OPIT), in a recent interview with Wired. Let’s take a closer look at the risks and how this issue ties into the issue of responsible artificial intelligence.
Uploading Your Image
To get a personalized image of yourself back from ChatGPT, you need to upload an actual photo, or potentially multiple images, and tell ChatGPT what you want. But in addition to using your image to generate content for you, OpenAI could also be using your willingly submitted image to help train its AI model. Vazdar, who is also CEO and AI & Cybersecurity Strategist at Riskoria and a board member for the Croatian AI Association, says that this kind of content is “a gold mine for training generative models,” but you have limited power over how that image is integrated into their training strategy.
Plus, you are uploading much more than just an image of yourself. Vazdar reminds us that we are handing over “an entire bundle of metadata.” This includes the EXIF data attached to the image, such as exactly when and where the photo was taken. And your photo may have more content in it than you imagine, with the background – including people, landmarks, and objects – also able to be tied to that time and place.
In addition to this, OpenAI also collects data about the device that you are using to engage with the platform, and, according to Vazdar, “There’s also behavioral data, such as what you typed, what kind of image you asked for, how you interacted with the interface and the frequency of those actions.”
After all that, OpenAI knows a lot about you, and soon, so could their AI model, because it is studying you.
How OpenAI Uses Your Data
OpenAI claims that they did not orchestrate these social media trends simply to get training data for their AI, and that’s almost certainly true. But they also aren’t denying that access to that freely uploaded data is a bonus. As Vazdar points out, “This trend, whether by design or a convenient opportunity, is providing the company with massive volumes of fresh, high-quality facial data from diverse age groups, ethnicities, and geographies.”
OpenAI isn’t the only company using your data to train its AI. Meta recently updated its privacy policy to allow the company to use your personal information on Meta-related services, such as Facebook, Instagram, and WhatsApp, to train its AI. While it is possible to opt-out, Meta isn’t advertising that fact or making it easy, which means that most users are sharing their data by default.
You can also control what happens with your data when using ChatGPT. Again, while not well publicized, you can use ChatGPT’s self-service tools to access, export, and delete your personal information, and opt out of having your content used to improve OpenAI’s model. Nevertheless, even if you choose these options, it is still worth it to strip data like location and time from images before uploading them and to consider the privacy of any images, including people and objects in the background, before sharing.
Are Data Protection Laws Keeping Up?
OpenAI and Meta need to provide these kinds of opt-outs due to data protection laws, such as GDPR in the EU and the UK. GDPR gives you the right to access or delete your data, and the use of biometric data requires your explicit consent. However, your photo only becomes biometric data when it is processed using a specific technical measure that allows for the unique identification of an individual.
But just because ChatGPT is not using this technology, doesn’t mean that ChatGPT can’t learn a lot about you from your images.
AI and Ethics Concerns
But you might wonder, “Isn’t it a good thing that AI is being trained using a diverse range of photos?” After all, there have been widespread reports in the past of AI struggling to recognize black faces because they have been trained mostly on white faces. Similarly, there have been reports of bias within AI due to the information it receives. Doesn’t sharing from a wide range of users help combat that? Yes, but there is so much more that could be done with that data without your knowledge or consent.
One of the biggest risks is that the data can be manipulated for marketing purposes, not just to get you to buy products, but also potentially to manipulate behavior. Take, for instance, the Cambridge Analytica scandal, which saw AI used to manipulate voters and the proliferation of deepfakes sharing false news.
Vazdar believes that AI should be used to promote human freedom and autonomy, not threaten it. It should be something that benefits humanity in the broadest possible sense, and not just those with the power to develop and profit from AI.
Responsible Artificial Intelligence
OPIT’s Master’s in Responsible AI combines technical expertise with a focus on the ethical implications of AI, diving into questions such as this one. Focusing on real-world applications, the course considers sustainable AI, environmental impact, ethical considerations, and social responsibility.
Completed over three or four 13-week terms, it starts with a foundation in technical artificial intelligence and then moves on to advanced AI applications. Students finish with a Capstone project, which sees them apply what they have learned to real-world problems.
Have questions?
Visit our FAQ page or get in touch with us!
WhatsApp
Write us at +39 335 576 0263
Email
Get in touch at hello@opit.com
Book a meeting
Talk to one of our Study Advisors
We are international
We can speak in:
