On March 13th, we hosted Eugene Chen, Chair of the Canadian Open Data Society and a member of Canada’s Multi-Stakeholder Forum on Open Government, and Katrina Ingram, Founder and CEO of Ethically Aligned AI for an open Q&A session with the public.
Here are some of the questions, answers, and key takeaways:
What is AI? LLM? GPT? Agents?
A simple way of thinking of it is – if a computer is helping you do something, it’s technically artificial intelligence, even a calculator could be viewed as AI.
In the past, much of AI was centred around processing data and analyzing patterns. In the early 2000s, the main fields of AI or ML (Machine Learning) that were getting most of the attention and funding were
- Image recognition (or image classification)
- Natural Language Processing (NLP)
However, since ChatGPT hit the market in November 2022, much of the attention has shifted towards these LLM (Large Language Model) type AIs that are able to realistically mimic conversation. These LLMs are a form of NLP.
GPT (Generative Pre-trained Transformer) is an AI that is trained on data and generates an output/response with an input/prompt.
These conversational models that we see all around us aren’t the only forms of AI. AI can also be used to do things like more accurately predict future weather based on past data and existing sensor data or optimize healthcare outcomes by more quickly and accurately detecting cancer from scan images.
Agents are basically a specialized type of AI tool that is able to process multiple commands in series or parallel. Agentic AI is a level above that.
Here’s a way to think of it, with the same prompt “Find me 5 restaurants in Whitehorse”, the following will produce different results in different ways:
| Chatbot | AI Agent | Agentic AI | |
| How It Responds | Provides a list of restaurants from a static, pre-defined database. | Fetches real-time information via external sources (APIs or web scraping). | Gives real-time data with possible follow-up questions for personalization. |
| Limitations | Limited to the data it has, static information, no personalized recommendations. | Dependent on the accuracy and availability of external data, might struggle with nuanced queries. | Requires more complex infrastructure, may raise privacy concerns, and relies on context to adapt and learn effectively. |
AGI (Artificial General Intelligence) is the next step beyond Agentic AI. Agentic AIs are designed for a specific function or specialization. AGI on the other hand, should be able to answer/react/produce different response in variety of topics/context, much like how humans can do all sorts of things from driving a car, playing a musical instrument, create art, and discuss biology.
How does AI learn/work?
Supervised – models that learn from labeled data.
Like a flashcard with an image of a dog, and the word “dog” written under it. It learns to associate the image of the dog with the word “dog”.
Unsupervised – models learn from unlabeled data, meaning there are no predefined values.
Clustering images with 4 legs and a tail together. It identifies patterns and structures within the data without explicit guidance.
Reinforcement Learning – through trial an error, receiving rewards or penalties for their actions.
The AI is like a student using flashcards but has to guess the answers and then receive feedback based on whether their guess was good or bad. It learns an optimal policy by maximizing the cumulative reward over time.
How will AI disrupt work?
Instead of increasing the gap, AI actually levels the playing field.
If you’re a beginner, it levels you up.
If you’re an expert, it actually levels you down.
For example, if you’re not good at graphic design or video creation, AI levels you up – you can easily do something that’s much better than what your would be able to create yourself as an amateur.
However, if you’re good at writing, AI might not be as good as you and the quality of the produced work is actually worse.
That said, experts are also able to produce better quality AI creations because they’re able to better prompt and iterate with the AI on responses.
Source: https://www.ethicallyalignedai.com/post/below-average-workers-will-benefit-the-most-from-using-ai
How will AI impact search?
AI tools like ChatGPT don’t perform an actual search of real-time information or data from the web when you ask a query. Instead, they generate responses based on patterns and information they’ve already been trained on. ChatGPT, for instance, was trained on a massive corpus of text data from books, websites, and other publicly available sources up until a certain point (like 2021 or 2023, depending on the version). It can produce text, answer questions, or generate creative content based on that training, without actually querying live or real-time data.
These tools predict the most likely and contextually appropriate responses based on the data it has. It creates text by recognizing patterns and choosing the best match for the input query, often mimicking how humans would respond – because of this, it is possible for the AI to “hallucinate” and provide an inaccurate or non-existing response.
Searching with a search engine like Google is a live query of the web. It pulls up information from real, current, and readily available web pages. Google constantly crawls and indexes the web, keeping an updated database of web pages, blog posts, articles, videos, and other online content. When you search, it returns the most relevant and up-to-date results by referencing these pages. Results aren’t perfect answers either and should still be fact-checked.
Related research paper: https://dl.acm.org/doi/10.1145/3498366.3505816
BUT, most search platforms are already implementing AI – to summarize search results.
Or some AI Agents are capable of Deep Research and can query a live database before jumping to providing a response.
What are some implications?
We say that we should be training people to think critically, yet these tools are reducing the need for us to think critically. It makes it too easy to get answers (and these tools are getting more accurate by the day too).
If I provide text for an AI to edit/proofread, will it use my data for training?
Most AI models use user data for training in free tiers, with some exceptions like Claude, Microsoft’s Copilot, Amazon’s Titan, and Cohere’s Command, where data is not used by default.
Research suggests paid tiers generally follow the same policy as free tiers, but some models like Open AI and Google Gemini use data for training by default, while others do not.
Check out https://glaze.cs.uchicago.edu/ – it prevents AI from using your art as training data by implementing a type of digital watermark.
From their website: Glaze is a system designed to protect human artists by disrupting style mimicry. At a high level, Glaze works by understanding the AI models that are training on human art, and using machine learning algorithms, computing a set of minimal changes to artworks, such that it appears unchanged to human eyes, but appears to AI models like a dramatically different art style. For example, human eyes might find a glazed charcoal portrait with a realism style to be unchanged, but an AI model might see the glazed version as a modern abstract style, a la Jackson Pollock. So when someone then prompts the model to generate art mimicking the charcoal artist, they will get something quite different from what they expected.
