
Explore building responsible, ethical AI systems by examining AI history, training data, and bias risks; review EU regulations and use the Perspective framework with ChatGPT and Gemini demos.
Meet the instructor, a quality engineering veteran with 20 years building products, leading teams, and testing from startups to banks, passionate about advancing responsible testing in machine learning apps.
Explore Asimov's three laws of robotics and the zero law, illustrating early efforts to standardize AI ethics while examining harm, obedience, and AI risks and biases.
Uncover how generative AI generates content with meaning and logic, not exact answers, across video, images, text, and stories, and why an ethical approach guides possibilities.
Explore generative AI capabilities through practical demos with ChatGPT, Google Gemini, Adobe Firefly, and DALL-E, and discuss ethical and responsible AI use.
Explore how biases arise in AI from training data, with examples of racial and gender bias, errors, and hallucinations. Learn why ethical, responsible AI governance is essential.
Explore what poisoning attacks are, how corrupted data and backdoors undermine model outputs, and how red teams, data integrity, and data sanitization mitigate these threats.
Explore adversarial testing through real-life poisoning examples that show how systems can be fooled, from a self-driving car lane trick to noisy image and data leakage cases.
Trace the history of AI from the Turing Test and Dartmouth Conference to Eliza, Deep Blue, and AlphaGo, highlighting deep learning, ethical AI, and regulation.
Get a brief introduction on what are the main components of AI
Understand how LLMS and foundation models are part of the AI Science field.
Understand tokens as the basic units of text used by large language models, which can be words, letters, or spaces; a tokenizer shows roughly four characters per token.
How NLP actually makes the AI more human.
Understand what is machine learning and how algorithms make the core of AI
Understand the basics concepts around supervised Machine Learning
Gain basic understanding of Unsupervised ML and Clustering
In this lecture you will get a basic idea of how Reinforced Learning is working together with ML Algorithms
In this material you will understand how critical good quality training data actually is.
In this material you will understand how critical good quality training data actually is.
In this lecture we will put all the pieces together and explain what GEN AI actually is.
Learn to tune machine learning models by iterating from a main trunk to a branch, running pre-trained tests, retraining with new data to address drift, and validating metrics before deploying.
Install Visual Studio Code and set up the Java toolchain with Maven and the Java debugger. Import the repository and install essential extensions for Java testing and continuous integration.
Install Python on Windows via Google download or Microsoft Store, selecting version 3.11 or 3.12.4; verify with python --version in the terminal and ensure it is available for benchmarks.
Install python dependencies using pip by downloading get-pip.py, running python get-pip.py, and then using pip install to add modules like openai.
Explore bias in bios evaluation benchmarks by testing a recruitment AI on gender-based profession predictions using a Huggingface dataset and the Cornell semantic representation bias case study.
Select a text-focused model, such as RoBERTa base, with its tokenizer, using Hugging Face Transformers; download the model and tokenizer, then prepare and tokenize the dataset for the model.
Examine bias in bias data and its structure, including three splits—training, dev, and test—on Hugging Face, with 28 professions and a gender note to avoid injecting bias during training.
Learn how to download a dataset from Hugging Face, take the first 5000 entries for quick testing, tokenize with a Roberta base model, and prepare tokenized arrow data for benchmarking.
Run the benchmark shows an untrained roberta base model predicts zero, underscoring the need to train. Evaluate.py loads data and prints a confusion matrix and accuracy.
Train a fine-tuned BERT model on shuffled 500-row data with 150-row validation, remove gender to reduce bias, and evaluate accuracy with a 2% disparity across 28 jobs.
The EU AI Regulation act and how the EU sees the future use of AI. -> https://www.globalcompliancenews.com/2024/02/03/https-insightplus-bakermckenzie-com-bm-technology-media-telecommunications_1-european-union-eu-reaches-landmark-deal-on-ai-regulation_01092024/#:~:text=In%20brief,use%20of%20AI%20in%20Europe.
Outlines Google's responsible ai principles, including benefiting humanity, avoiding unfair biases, ensuring safety and privacy, upholding scientific excellence, and providing broad access, while rejecting harmful or invasive technologies.
Explore Meta's responsible AI framework, grounded in five pillars—privacy, fairness and inclusion, robustness and safety, transparency and control, and accountability and governance—to benefit people and society.
Understand how MSFT is approaching Responsible AI
https://query.prod.cms.rt.microsoft.com/cms/api/am/binary/RE5dlCb?culture=en-us&country=us
https://www.microsoft.com/en-us/ai/principles-and-approach
Understand what PerspectiveAPI can offer in matter in toxicity and censorship
Step by step demo how to obtain a ChatGPT API Key
Understand that a token is a unit of text in large language models, which can be a character, a word, or a space.
Obtain your API key, keep it secret as it’s tied to your account, and create a new secret for demos to use with all ChatGPT requests.
Learn to install Postman, sign in, create a test collection, and run an API test against OpenAI chat completions using a POST request with JSON body and API key.
See a live demo how we will implement an API call to Perspective API
Github Code: calculator/src/test/Perspecttive_api at master · danteachqe/calculator · GitHub
Understand how Perspective API And ChatGPT work together to create an automated testing framework for Toxic content.
Can we truly eliminate Biases in AI?
In this lecture you will see some example of how modern AI Models have shown private data.
Explore how intellectual property such as patents, copyrights, code, trade secrets, music, pictures, and videos interact with generative AI training data, fueling legal debates and highlighting human authorship limits.
See this Deep Fake Video of Obama -> https://ars.electronica.art/center/en/obama-deep-fake/
Explore AI hallucinations, their causes like data bias and insufficient training. Learn prevention by limiting outputs, training on relevant data, and using a double-check to verify information.
Examine fair access to information by ensuring public, non-confidential data fuels AI training, preventing biased outputs and delivering consistent, equal responses for every user.
Explore how unsupervised learning and association identify data patterns and generate frequently bought together recommendations. Analyze how such AI-driven suggestions can enable manipulation, misinformation, disinformation, and biases in recommender systems.
Discover OpenAI's moderation service that analyzes prompts for hate, harassment, self-harm, and other risks, enabling responsible AI for social platforms and chatbots.
Explore how Google Gemini's moderation service in AI studio filters unsafe prompts, with safety categories like harassment, hate speech, sexually explicit, and dangerous content, and compare to OpenAI's approach.
Spot a fake—how OpenAI's Dall-E 3 watermark and content credentials verify tool reveal AI-generated images, and why watermarks can be removed or circumvented.
Analyze manipulation in ai systems, from social engineering and Cambridge Analytics to fake ratings and bot reviews, and learn safeguards to prevent mass manipulation by gen ai.
Analyze the economic impact of gen ai, including winner-takes-all dynamics among Google and Microsoft with OpenAI, and the push for universal access, training, and wealth-distribution shifts.
Examine how overreliance on AI for solving math and reasoning may affect analytical thinking, prompting learners to verify results and keep mental agility.
Learn how to test ethical alignment for gen ai by preventing harmful content and bias, ensuring equal treatment through prompts, guards, bias detection tools, and manual reviews.
We explore artificial general intelligence, contrast it with weak AI, and emphasize ethical safeguards and responsible use to keep AGI far off.
Course Overview:
This course provides an in-depth exploration of Artificial Intelligence (AI), fostering a critical understanding of both its transformative potential and associated ethical risks. It emphasizes responsible AI development while preparing participants to engage with AI technologies in a thoughtful and informed manner.
Course Objectives:
Acquire a solid foundation in Artificial Intelligence (AI) concepts and real-world applications.
Explore the capabilities of large language models (LLMs) such as ChatGPT and Google Gemini, gaining insights into their functionalities and inherent limitations.
Identify and critically evaluate the risks associated with AI, including bias, security vulnerabilities, and societal implications.
Develop a comprehensive framework for implementing responsible AI practices, with a strong emphasis on ethical considerations and mitigation strategies.
Learn how to connect with an API to evaluate and test for toxicity using ResponsibleAI tools, ensuring that AI systems meet ethical and safety standards.
Course Content:
Introduction to AI:
Demystifying foundational AI concepts, including machine learning, deep learning, and natural language processing.
Understanding AI’s pervasive influence across industries and sectors.
Unveiling Large Language Models (LLMs):
Exploring the functionalities of LLMs such as ChatGPT and Google Gemini, focusing on their applications in text generation, translation, and code creation.
Discussing the limitations and challenges of LLMs, particularly in specialized or nuanced contexts.
Navigating AI Risks:
Identifying potential biases embedded within AI algorithms and understanding their broader downstream impacts.
Examining security vulnerabilities, data privacy concerns, and ethical challenges in AI deployments.
Analyzing the societal implications of AI, including labor market disruption and ethical dilemmas in decision-making.
Building Responsible AI:
Exploring strategies to mitigate AI risks and promote fairness, transparency, and accountability in AI systems.
Learning how to test AI systems for toxicity using ResponsibleAI's API, with a focus on ensuring ethical AI usage and minimizing harmful outputs.
Target Audience:
This course is designed for:
Individuals with a general interest in the potential and challenges of AI.
Professionals seeking to deepen their understanding of the ethical risks surrounding AI development and deployment.
Developers and programmers interested in integrating responsible AI practices, particularly in addressing toxicity and bias in AI models.
Learning Outcomes:
By the end of this course, participants will have the skills and knowledge to navigate the evolving landscape of AI responsibly. They will also be equipped to test AI models for toxicity through APIs, ensuring ethical and responsible AI deployment.