<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Iliada Eleftheriou on Medium]]></title>
        <description><![CDATA[Stories by Iliada Eleftheriou on Medium]]></description>
        <link>https://medium.com/@iliada.eleftheriou?source=rss-8abb2658ce33------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*TvCUW14yfhyiXWQfYsF5wg.jpeg</url>
            <title>Stories by Iliada Eleftheriou on Medium</title>
            <link>https://medium.com/@iliada.eleftheriou?source=rss-8abb2658ce33------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 05:41:53 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@iliada.eleftheriou/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AI Unveiled - Part 2 : Debunking generative AI and Large Language Models]]></title>
            <link>https://medium.com/@iliada.eleftheriou/ai-unveiled-part-2-debunking-generative-ai-and-large-language-models-2da1a27d6244?source=rss-8abb2658ce33------2</link>
            <guid isPermaLink="false">https://medium.com/p/2da1a27d6244</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[limitations-of-chatgpt]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Iliada Eleftheriou]]></dc:creator>
            <pubDate>Wed, 05 Feb 2025 13:14:43 GMT</pubDate>
            <atom:updated>2025-02-05T13:14:43.531Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI Unveiled - Part 2 : Debunking generative AI and Large Language Models</h3><p>Part 2 is all about demystifying ChatGPT, exploring key concepts of generative AI and understanding the limitations of current GPT models by exploring how these tools work behind the scenes — <em>but without any sophisticated math formulas or complicated diagrams, I promise</em>.</p><p>Part 1 in this series explores <a href="https://medium.com/@iliada.eleftheriou/ai-unveiled-part-1-what-is-ai-and-what-isnt-fcaa90fd18ef">what is AI and what isn’t</a>.</p><p><strong>Keywords:</strong> <em>generative AI, large language models, ChatGPT, limitations, hallucinations, training datasets, training process, ethical considerations.</em></p><h3>What is ChatGPT and generative AI tools?</h3><p>Advances in Artificial Intelligence have resulted in the formation of <strong>generative AI</strong> tools, like <strong>ChatGPT. </strong>These tools <strong>generate</strong> an output in the form of text, pictures, videos, and even code based on an <strong>input prompt</strong> (a set of instructions usually in the form of text, or voice) from the user, and that’s where they got their name from.</p><p>These tools are based on <strong>large language models (LLMs)</strong> employing machine learning methods, like deep learning and neural networks. In a nutshell, <strong>LLMs use statistical models to analyse vast amounts of data, learn the patterns and connections between words and phrases, and understand the context of the user’s input prompt to predict an output. </strong>In response to instructions (input prompt), they can generate blog posts, essays, poems, create pictures, synthesise voices, produce songs and videos, design websites, write code and much more.</p><ul><li><strong>ChatGPT </strong>is a conversational generative AI tool, based on the GPT models (i.e. GPT3.5, GPT4o).</li><li>GPT stands for Generative Pre-trained Transformer, and is a <strong>LLM</strong> that uses deep learning to analyse an input prompt by the user and predict an output. In the form of a text-based output, the model generates content by <strong>predicting a word at a time</strong>.</li><li>Other generative AI tools include Microsoft’s Copilot, Google’s Gemini (previously known as Bard), Claude, and Dall-e. There are thousands of gen AI applications with many more cropping up on a daily basis. Most of these tools are based on OpenAI’s GPT models.</li></ul><p>Generative AI tools have the potential to disrupt and transform many aspects of our lives, with a myriad of applications in healthcare, business, education, law, policy, policing and more.</p><p>Whilst Generative AI is a <strong>powerful tool</strong> with great potential, there are <strong>limitations</strong> and potential hazards to its use, particularly if used incorrectly. Therefore, it is vital that we are guided in the proper, correct and <strong>ethical use</strong>, and that we are equipped with the skills required to safely utilise and critically evaluate generative AI tools, understanding their benefits, applications, and limitations.</p><blockquote><strong>Over to you:</strong> Can you spot the limitations and ethical challenges?</blockquote><blockquote>While going through the remaining of this section, make a list of limitations of gen AI tools and the impact it can have on our lives and jobs. Some of them will be quite obvious , others might be slightly hidden waiting for you to discover them. Let’s find out how many you can spot.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AhKN_Or28YLor8ySjpkrog.jpeg" /><figcaption>Image created with DALL-E (Jul 2024). Can you spot any mistakes? Why do you think that is?</figcaption></figure><h3><strong>What generative AI isn’t</strong></h3><p><strong>Generative AI tools are not search engines.</strong> ChatGPT is not a replacement and doesn’t work like the Google search engine<strong> </strong>(even though some search engines use gen AI tools to summarise findings from search results)<strong>. A search engine crawls the web to identify the most relevant list of <em>existing</em> sources that match the keywords the user had provided with.</strong></p><p><strong>Generative AI tools analyse the input prompt and predict a sequence of words that best fit the patterns in the data they were trained on, </strong>without understanding the truthfulness of the generated content and without always fact checking the output with trustworthy sources, just yet(1).</p><h3><strong>Predictions and hallucinations</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/833/1*uA9TEgIy3VsavWyX33PDPw.png" /><figcaption>Screenshot of ChatGPT4o hallucinating, confidently and convincingly communicating wrong information.</figcaption></figure><p>Open AI evaluated responses of the different versions of GPT models against 9 predefined categories to assess and compare the accuracy and reliability of the models. In the figure below we can see that the model with highest accuracy was GPT4 (the paid version of ChatGPT) and the highest percentage was around 80% in the areas of history and science with <strong>~20% (1 out of 5) making wrong predictions.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F2HKN_3GsTAU5J2upQ_acQ.png" /><figcaption>Source: Internal factual evaluation of GPT4 compared with other GPT models by <a href="https://openai.com/research/gpt-4">Open AI documentation(opens in a new tab)</a>, March 2023 [1].</figcaption></figure><p><em>“GPT-4 generally lacks knowledge of events that have occurred after the vast majority of its </em><strong><em>data cuts off </em></strong><em>(September 2021), and does not learn from its experience. It can sometimes make </em><strong><em>simple reasoning errors</em></strong><em> which do not seem to comport with competence across so many domains, or be overly gullible in </em><strong><em>accepting obvious false statements</em></strong><em> from a user. GPT-4 can also be confidently wrong in its predictions, not taking care to double-check work when it’s likely to make a mistake” [1]</em></p><blockquote><strong>Over to you:</strong> What can we derive from the chart above? Have you identified any ethical challenges with the use of different versions of GPT models?</blockquote><h3>Exploring behind the scenes</h3><h4>A. Learning from data</h4><p>As humans, we learn from experience. We observe the world, absorb information and try things based on what we observed, we see if they work, and then modify what we’ve done so that it works better next time.</p><p>Like we’ve seen in the previous section, traditional computer systems can only follow their pre-set programming, whose behaviour is constrained by the rules set up in its code. For example, a manufacturing robotic arm will have its movements pre-determined by the programming it has been given.</p><p>In machine learning, we are developing algorithms that can learn the rules they need to operate from the data that is presented to them. Essentially, <strong>the algorithmic model observes a set of data, identifies patterns in the data and derives rules rather than relying on a team of programmers to define them explicitly.</strong></p><p>As we now live in a world that is incredibly rich with data, this approach is becoming more and more applicable. The increase in computing power, coupled with the big data revolution (the volume and variety of data we create on a daily basis is increasing exponentially), has made machine learning a very powerful option for developing computational tools to work in the real world.</p><h4>B. The training dataset</h4><p>At its core, a large language model works by <strong>analysing massive amounts of data</strong>. This data, known as <strong>a training dataset</strong>, consists of various forms of information from articles, literature, news, social media posts and forums. These are usually open source datasets that are freely available online.</p><p>According to Open AI, GPT3 (one of the earlier versions of ChatGPT) has been trained on 45 terabytes of unfiltered data including books, and the <a href="https://commoncrawl.org/overview">Common Crawl </a>(a massive open access dataset that has been created by crawling the web collecting web pages and metadata) [2].</p><p><strong>The Pile</strong></p><p>The Pile is a massive dataset specifically created to train LLMs. The Pile contains a total of 825 GiB (one gibibyte equals 1.074 gigabytes) diverse, open-source datasets including the entire Wikipedia (English), PubMed Central, ArXiv, Github, Stack Exchange, and many more as illustrated in the figure on the left [3].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/384/1*Lfy_mwQi4wl3D13J093SvA.png" /><figcaption>Source: Leo Gao et al, The pile: An 800gb dataset of diverse text for language modeling. Available at: <a href="http://Source%3A%20Leo%20Gao%20et%20al,%20The%20pile%3A%20An%20800gb%20dataset%20of%20diverse%20text%20for%20language%20modeling.%20Available%20at%3A%20https%3A//arxiv.org/pdf/2101.00027.pdf">https://arxiv.org/pdf/2101.00027.pdf</a></figcaption></figure><blockquote><strong>Over to you: </strong>Can you identify any limitations with this dataset? What are the implications of these?</blockquote><ul><li><strong>Trustworthiness: </strong>Some data sources include trustworthy information that has been scientifically peer reviewed, classic literature or represent an organisations views and perceptions on local and international news. Other data sources contain information retrieved from forums, YouTube video subtitles, and employee’s emails.</li><li><strong>Consent: </strong>An ethical challenge with large language models is whether consent is needed for the data to be used for training these models. Some training datasets (like the Common Crawl) currently operate in an ‘<strong>opt-out</strong>’ manner [<a href="https://commoncrawl.org/blog/balancing-discovery-and-privacy-a-look-into-opt-out-protocols">4</a>]. But, who really owns your social media posts and music preferences?</li><li><strong>Language: </strong><em>The majority of data sources used in training LLMs contain data created by the Western world. According to Leo Gao et al, </em><strong><em>“While only 13% of the world’s population speaks English, the vast majority of NLP (Natural language processing) research is done on English”</em></strong><em> [3].</em></li></ul><h4>C. The training process</h4><p>Training a machine learning model employs the analysis of training datasets to create a model. The model uses statistical measures to identify patterns and trends on how the words are used together and find connections between words and phrases. In simple terms, the model creates a map with the most used words, what words usually come after or used together, how certain topics are connected, and many more.</p><p>For example, the model can spot that whenever the word “rain” appears, words like “umbrella,” “wet,” or “storm” often follow. These connections aren’t arbitrary — they’re the result of the model finding recurring patterns in how language is used in the training datasets. Just like how we instinctively know that “sun” is usually associated with “heat”, “light” or “morning”, the model learns to predict which words are likely to come next based on what it’s seen in the training dataset.</p><p>But here’s where things get really interesting. The model doesn’t just stop at simple word associations, newer versions of LLMs can pick up on context cues, understand synonyms and even grasp some subtle differences in tone and sentiment.</p><p>During training, the model learns to predict the next word in a sequence given the preceding context. Given an input prompt by the user, the model first analyses the sequence of words to understand the context, and the task that is needed, then informs and adjusts the map and finally predicts an output.</p><h4><strong>Learning from human feedback</strong></h4><p>In the very first cycles of the training process of an LLM, the AI model (a series of algorithmic computerised instructions) makes a random attempt at predicting what the next word in the output should be. Periodically, two outputs were provided to a human to decide which of the two is closest to fulfilling the user’s input prompt goal. The human’s decision is used to provide a sense of reward to the AI model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/618/1*VNYIKWYjYgokkel4umOMXQ.png" /><figcaption>Source: “Introducing ChatGPT”, OpenAI documentation. Available at: <a href="https://openai.com/blog/chatgpt">https://openai.com/blog/chatgpt</a></figcaption></figure><h4><strong>Reinforcement learning</strong></h4><p>The AI algorithm gradually builds a model of the goal of the task by finding the reward function that best explains the human’s judgments. It then uses a process called Reinforcement Learning to learn how to achieve that goal.</p><p><strong>Reinforcement learning</strong> is an iterative process where the AI model learns the optimal behaviour in an environment to obtain maximum reward. This optimal behaviour is learned through interactions with the environment and observations of how it responds, similar to children exploring the world around them and learning the actions that help them achieve a goal [5].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/533/1*xSPbK5gv7egjlyw_ut9LKg.png" /><figcaption>Source: “Learning from human preferences” OpenAI documentation. Available at: <a href="https://openai.com/research/learning-from-human-preferences">https://openai.com/research/learning-from-human-preferences</a></figcaption></figure><h4><strong>D. Fine-tuning the model</strong></h4><p>Behind the scenes, the model is constantly fine-tuning itself, continually training the model and honing its understanding of language with each new piece of text it encounters, including new input prompts from users around the world. The more data the model is trained on, the better it will be at generating new, more accurate (and perhaps more creative) content. It’s a continuous process of learning, fine-tuning and adaptation.</p><h4>Summary</h4><p>Financial Times created an infographic visualising how large language models can understand context of the words used and predict an output. Click on the link below to see a summary of how LLMs work behind the scenes.</p><p><a href="https://ig.ft.com/generative-ai/"><strong>https://ig.ft.com/generative-ai/</strong></a></p><h3>Summary of current limitations and (some) ethical challenges</h3><p>While significant progress has been made, there remain some limitations and technical constraints. This is not an extensive list, and will soon be outdated. As the field evolves, new challenges will arise. It’s important to be informed of limitations and current and future ethical considerations, and participate to discussions defining policies and guidelines on safe, ethical and proper use of generative AI tools.</p><p><strong>Accuracy and reliability: </strong>AI generative tools may sometimes produce outputs that are inaccurate or unreliable. They rely on patterns and data they have been trained on, and if the training data is incomplete or biased, it can affect the accuracy of the generated content. Additionally, AI models can make mistakes or produce nonsensical or contradictory outputs.</p><p><strong>Fabrication of facts and dissemination: </strong>Generative AI can fabricate facts or generate false information because the model tries to predict the next word in a given sentence based on the context, statistical relationships and patterns learnt from training data. This can pose a problem when the generated information is used as a source for research or presented as factual information, leading to the dissemination of misinformation.</p><p><strong>Digital divide</strong> between those who can afford the most accurate, more powerful models and those who can’t.</p><p><strong>Data cut-off: </strong>AI generative tools have a knowledge cut-off point, meaning they were trained on data up until a specific date. As a result, they may not be aware of recent events or developments that have occurred after their knowledge cut-off point. This can lead to responses that are outdated or incomplete when asked about current events. For example, ChatGPT (GPT3.5) has a cut-off of data on September 2021.</p><p><strong>Convincing tone</strong> that can be misleading.</p><p><strong>High energy consumption:</strong> training large AI models requires vast computational power, relying on energy-intensive data centres with significant carbon emissions, raising concerns about environmental sustainability.</p><p><strong>Language: </strong>Favour towards western perspectives and cultures.</p><p><strong>Decision making: </strong>Can lead to detrimental results if used in decision making processes without the proper validations.</p><p><strong>Input prompt: </strong>The quality of the generated output is prone to the user’s ability to form an input prompt.</p><blockquote><strong>Over to you:</strong> Have you identified any other? Let us know in the comments.</blockquote><h3>Glossary</h3><ul><li><strong>Generative AI:</strong> tools that employ machine learning models to generate outputs based on instructions by the user. The output can take various formats including but not limited to text, voice, visuals, diagrams, code, and many more.</li><li><strong>Input prompt: </strong>A set of instructions (often in the form of text, or voice) provided by the user.</li><li><strong>LLMs: </strong>A set of machine learning models that power generative AI tools. LLMs employ deep learning mechanisms to <strong>understand</strong> natural languages, words, sentences, context, tone (sometimes even idioms, humour and irony) and <strong>predict</strong> the best possible response based on their understanding of the language.</li><li><strong>ChatGPT:</strong> A conversational generative AI tool that produces text and pictures based on a text-based input prompt given by the user. At the time of writing, the free version of ChatGPT is based on the GPT3.5 model and the paid version on the GPT4o model.</li></ul><h3><strong>Notes, footnotes, and references</strong></h3><p><strong>AI and particularly generative AI is a fast-paced area with new tools being released on a daily basis. Some information provided in this section will soon be outdated. Always check with documentation and please don’t hesitate to get in touch if you spotted a mistake. Thanks!</strong></p><p><strong>Footnotes</strong></p><p><strong>(1):</strong> There’s a feature in ChatGPT (GPT4 paid model) called GPT Browse that autmatically detects if an input prompt requires to search the web and uses Microsoft’s Bing search engine to return an output and links to the sources of the information. More info <a href="https://help.openai.com/en/articles/8077698-how-do-i-use-chatgpt-browse-with-bing-to-search-the-web">here(opens in a new tab)</a>. OpenAI is also testing a prototype called <a href="https://openai.com/index/searchgpt-prototype/">SearchGPT(opens in a new tab)</a> that provides search results with sources.</p><p><strong>References:</strong></p><p>[1] Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S. and Avila, R., 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774. Accessible here: <a href="https://arxiv.org/pdf/2303.08774">https://arxiv.org/pdf/2303.08774(opens in a new tab)</a></p><p>[2] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A. and Agarwal, S., 2020. Language models are few-shot learners. Advances in neural information processing systems, 33, pp.1877–1901. Accessible here: <a href="https://commoncrawl.org/blog/balancing-discovery-and-privacy-a-look-into-opt-out-protocols">https://arxiv.org/pdf/2005.14165.pdf(opens in a new tab)</a></p><p>[3] Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N. and Presser, S., 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027.</p><p>[4] Common Crawl, “Balancing Discovery and Privacy: A Look Into Opt–Out Protocols” <a href="https://commoncrawl.org/blog/balancing-discovery-and-privacy-a-look-into-opt-out-protocols">https://commoncrawl.org/blog/balancing-discovery-and-privacy-a-look-into-opt-out-protocols(opens in a new tab)</a>, Accessed on August 2024</p><p>[5] Synopsys, “What is Reinforcement Learning?” <a href="https://www.synopsys.com/glossary/what-is-reinforcement-learning.html#:~:text=It%20is%20about%20learning%20the,help%20them%20achieve%20a%20goal">https://www.synopsys.com/glossary/what-is-reinforcement-learning.html(opens in a new tab)</a> Accessed on March 2024.</p><p>[6] Bianchi, F., Kalluri, P., Durmus, E., Ladhak, F., Cheng, M., Nozza, D., Hashimoto, T., Jurafsky, D., Zou, J. and Caliskan, A., 2023, June. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (pp. 1493–1504).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2da1a27d6244" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Unveiled - Part 1: What is AI and what isn’t?]]></title>
            <link>https://medium.com/ai-robot-overlord-replacement-or-colleague/ai-unveiled-part-1-what-is-ai-and-what-isnt-fcaa90fd18ef?source=rss-8abb2658ce33------2</link>
            <guid isPermaLink="false">https://medium.com/p/fcaa90fd18ef</guid>
            <category><![CDATA[general-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[education-technology]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Iliada Eleftheriou]]></dc:creator>
            <pubDate>Wed, 31 Jan 2024 11:40:59 GMT</pubDate>
            <atom:updated>2024-01-31T11:40:59.024Z</atom:updated>
            <content:encoded><![CDATA[<p>In this series, we will demystify Artificial Intelligence (AI), distinguish between misconceptions and realities, foster a critical approach to this technology and investigate potential applications and ethical considerations.</p><p>Part 1 of the series provides a brief introduction to the umbrella term of AI, and delves into the differences of AI algorithms and conventional programming.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NMiojBtBitqa5vKF.jpg" /></figure><h3>Imagine this</h3><p>Imagine waking up in a world not too far from our present, and as you glance at your smart mirror, it not only reflects your image but also updates you on the day’s weather, your schedule, suggests a personalised breakfast based on your health data, and even provides clothing options based on your daily planned events. The autonomous car outside your door is ready to take you to work, effortlessly navigating through traffic and predicting alternate routes in real-time to ensure you arrive on time. Meanwhile, your AI-powered virtual assistant has already sorted through your emails, responding to routine messages and flagging the important ones for your attention.</p><p>At work, your team is collaborating on a project with the help of an AI-driven project manager. It tracks progress and deadlines but also suggests innovative solutions based on historical data and emerging trends. During a meeting, you present a realistic AI-generated presentation, powered with automated scenarios (like this one) images and videos, where the content is dynamically tailored to the various audience’s preferences, technical backgrounds and engagement levels.</p><p>Artificial Intelligence has the potential to integrate into our daily lives, enhancing efficiency, decision-making, and personalisation.</p><p>But what exactly is Artificial Intelligence, and how does it differ from other technological concepts?</p><h3>What Is Artificial Intelligence (AI)?</h3><p>We often hear about AI innovations in the areas of health, manufacturing, business, law, and many more. But what is AI and most importantly what isn’t?</p><p><strong>Artificial Intelligence is an umbrella of terms and refers to the development of computer systems that can perform tasks that typically require human intelligence. These tasks include learning, reasoning, problem-solving, perception, and language understanding.</strong></p><p>AI is not a single technology but a broad field of science encompassing various subfields and approaches.</p><p>Examples of AI sub-fields:</p><ul><li><strong>Machine Learning (ML)</strong> is a subset of AI that involves training algorithms on large pre-existing datasets to find patterns and make predictions or decisions about future events.</li><li><strong>Natural Language Processing (NLP)</strong> focuses on the interaction between computers and human languages, enabling machines to understand, interpret, and generate human-like text.</li><li><strong>Computer Vision (CV) </strong>involves training algorithms to interpret and understand the visual world (including static images, videos, and real-time camera supervision) enabling tasks such as image recognition and object detection.</li><li><strong>Robotics</strong>: AI-driven robots can perform tasks in the physical world, ranging from simple actions to complex movements.</li></ul><p>In this video, we give a tour of the various robots of every size and format that are being studied and/or developed at the University of Manchester Schools of Engineering, Computer Science and Mechanical, Aerospace and Civil Engineering (more info <a href="https://www.cs.manchester.ac.uk/research/facilities/cognitive-robotics-laboratory/">here</a>). This is a 360 video, to explore the lab and don’t miss the additional information and videos on the left wall.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FgfTvNyN928U%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DgfTvNyN928U&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FgfTvNyN928U%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/af404c253ab6b39818205a199c29eb9c/href">https://medium.com/media/af404c253ab6b39818205a199c29eb9c/href</a></iframe><h3><strong>Three levels of AI</strong></h3><p>As research and development in the area of AI progresses, we move from simple to more complex algorithms, from lower to higher forms of intelligence. Human-like characteristics, such as emotions and thought processes, could potentially emerge and be replicated by machines and algorithms.</p><p>Currently, we categorise AI in three levels:</p><ol><li><strong>Artificial Narrow Intelligence</strong>, also known as weak AI, is the currently predominant form of AI as we know it. In narrow AI, a learning algorithm is designed to perform a <strong>single task</strong>, such as image recognition, language translation, or playing chess. Any knowledge gained from performing that task will not automatically be applied to other areas. Examples of narrow AI include virtual assistants like Siri or Alexa, recommendation algorithms on streaming platforms, and facial recognition systems.</li><li><strong>General AI </strong>represents a higher level of artificial intelligence that can understand, learn, and apply knowledge across a wide range of tasks and areas showcasing a form of generalising knowledge. General AI is not confined to specific tasks and can adapt to new, unfamiliar situations. It demonstrates a level of cognitive flexibility and autonomy resembling human-like intelligence.</li><li><strong>Super AI </strong>represents a theoretical level of AI where machines surpass human intelligence across all aspects. It implies a machine possessing cognitive abilities emotional intelligence, creativity and adaptability, exceeding the intellectual capabilities of the brightest human minds. Super AI is categorised by autonomous decision making and self-improvement, in which algorithms can independently make decisions, innovate, and solve complex problems without any human intervention.</li></ol><p>As of now, we have achieved narrow AI but progress is ongoing. General AI and Super AI remain theoretical and are subjects of research and often speculation within the AI community. Some organisations use the two terms, general and super AI, interchangeably.</p><h3><strong>What AI isn’t</strong></h3><p>Despite the exciting progress in AI, we are far from creating machines with general intelligence surpassing human capabilities, just yet.</p><p>AI is not magic. AI tools operate based on data and algorithms. They are powerful tools but not infallible or all-knowing entities. They are not capable of independent reasoning and rely on training algorithms to identify patterns in the data to make decisions.</p><p>While AI can simulate certain human-like behaviours, it lacks consciousness, self-awareness, and emotions.</p><h3>Is it really AI? Distinguishing between AI and traditional programming</h3><p>In conventional traditional programs, a developer meticulously writes explicit instructions to direct the computer on how to execute a task. For example, if tasked with recognising and distinguishing between pictures of cats and dogs, in traditional programming, explicit rules are required to be defined by the programmers to analyse pixel values, coloir distribution, and geometric shapes within the images. “If the image has pointy ears and a long snout, classify it as a dog; if it has rounded ears and a shorter snout, classify it as a cat.” These rules are manually created by programmers based on our understanding of the visual characteristics of dogs and cats.</p><p>In the case of an AI algorithm, a machine learning model is trained on a large dataset of labeled images of dogs and cats (the model is provided with lots of pictures of cats and dogs and explicit labels whether a picture shows a cat or a dog). During training, the model finds patterns in the data and learns to automatically extract relevant features from the images without the support of explicit programming and rules provided by the team of programmers. These patterns enable the model to generalise its understanding to classify new, unseen images and predict whether a picture shows a cat or a dog.</p><p>The AI algorithm can adapt to a wide range of scenarios and variations, like changes in lighting, poses, and pictures of different breeds. It can also generalise well to new images not seen during training, even when dogs were cute sunglasses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W6M6hzqxykPJxog_U2XRMQ.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@alanking?utm_content=creditCopyText&amp;utm_medium=referral&amp;utm_source=unsplash">alan King</a> on <a href="https://unsplash.com/photos/long-coated-brown-dog-KZv7w34tluA?utm_content=creditCopyText&amp;utm_medium=referral&amp;utm_source=unsplash">Unsplash</a></figcaption></figure><h4>AI: a replacement or a colleague?</h4><p>AI should not be viewed as an omnipotent entity that can replace human intuition and creativity. Instead, it serves as a tool designed to augment human capabilities, automating specific tasks and offering valuable insights.</p><p><em>What do you think?</em></p><p>In the following parts of this series, we will explore ethical considerations, data bias, importance of transparency and AI regulations, and generative AI tools.</p><h3>Author</h3><p>Dr Iliada Eleftheriou is a Senior Lecturer at the University of Manchester. She has a background in Computer Science (PhD) and is specialising in mapping complex data landscapes in healthcare settings to identify and address socio-technical challenges stemming from disparate information systems and data formats. Dr Eleftheriou is a member of the Greater Manchester Combined Authority Information Board and co-leading the health data systems community network of Cancer Research UK.</p><p>She is the Deputy Director of the Clinical Data Science programme and is leading modules on Health Informatics and Information Engineering including the interdisciplinary ‘AI: Robot Overlord, Replacement or Colleague?’ that investigates the impact of AI in our future lives and workforce.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fcaa90fd18ef" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-robot-overlord-replacement-or-colleague/ai-unveiled-part-1-what-is-ai-and-what-isnt-fcaa90fd18ef">AI Unveiled - Part 1: What is AI and what isn’t?</a> was originally published in <a href="https://medium.com/ai-robot-overlord-replacement-or-colleague">AI: robot overlord, replacement, or colleague?</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Supporting the role of teaching assistants in online teaching]]></title>
            <link>https://medium.com/i3hs-hub/supporting-the-role-of-teaching-assistants-in-online-teaching-d91da57c13ae?source=rss-8abb2658ce33------2</link>
            <guid isPermaLink="false">https://medium.com/p/d91da57c13ae</guid>
            <category><![CDATA[tutor]]></category>
            <category><![CDATA[higher-education]]></category>
            <category><![CDATA[online-teaching]]></category>
            <category><![CDATA[team-management]]></category>
            <category><![CDATA[online-transition]]></category>
            <dc:creator><![CDATA[Iliada Eleftheriou]]></dc:creator>
            <pubDate>Wed, 09 Sep 2020 15:18:28 GMT</pubDate>
            <atom:updated>2020-09-09T15:18:28.636Z</atom:updated>
            <content:encoded><![CDATA[<p>As the coronavirus pandemic spreads, there has been an increasing move towards online teaching. Delivering large scale teaching online comes with a myriad of challenges. A lesson we have learnt by delivering such a course unit at the University of Manchester, was the vital part of employing teaching assistants (TAs) to support the students’ learning experience and make learning more engaging and interactive.</p><p>Here, we discuss the practical steps we took in organising and supporting the role of TAs in the fully online undergraduate unit <a href="https://www.bmh.manchester.ac.uk/stories/ai/"><em>AI: Robot Overlord, Replacement or Colleague?</em></a><em> (referred to as Robot Overlord) </em>which since then have been adopted in various other units across faculties. We also provide a set of templates and a TA starting pack, we have used to support the TAs.</p><h3>The role of TAs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y0RyDBnptyiB2g2ChR5SrA.png" /><figcaption>Photo by Aron Visuals on Unsplash.</figcaption></figure><p>The primary role of TAs in this unit is to provide on-going support to the students and facilitate their journey towards achieving knowledge. They act as a compass guiding the students through the teaching materials pointing them to the right direction.</p><p>In this new normality, where almost all human interaction happens online, TAs also play a vital part in<strong> building a sense of community and belonging </strong>with our students. They actively engage with the students on a weekly basis providing on-going support to both summative and formative activities.</p><h3>The selection process</h3><p>We originally had a pool of candidates (all PhD students from a variety of disciplines) to select from. We have asked them to complete a form that gathers information on previous experience in teaching and their domain knowledge on the topics covered in this unit. There is a copy of the form we have used in the ‘Resources’ section below.</p><p>We then went through the completed forms and selected the candidates that fulfilled our requirements. We made sure to include candidates from several disciplines and backgrounds, but also from different levels of study. Third and fourth year PhD students often have more experience on teaching and can bring their own ideas into it. Similarly and equally important, are the 1st and 2nd year TAs who will be part of the team for longer.</p><p>As a result of this unusually quite extensive process of selecting the TAs, we have observed that the selected TAs were highly motivated. The Robot Overlord TAs often went further than their pre-assigned responsibilities to help and support the students.</p><blockquote>“The unit is <strong>very well run</strong> and make great use of <strong>mixed media</strong> to provide a very involved and interactive teaching experience. The platform and delivery of this course is even more important at the moment as more courses look to move online, and <strong>it has been a great experience</strong>.” Aj Mubarik, Robot Overlord TA 2019–2020</blockquote><h3>The training process</h3><p>Up until now we held a two-hour face to face briefing session just before the beginning of the semester to train the successful candidates. We will now have to move this session online using a video conferencing tool. The agenda for the online training session we will have on September is as follows:</p><ul><li>Introductions of course educators and TAs</li><li>Course unit aims and objectives</li><li>TAs roles and responsibilities — examples can be found in the TA handbook available in the ‘Resources’ section below</li><li>Students’ personas (this is particularly useful in interdisciplinary units — example can be found in the TA handbook)</li><li>Overview of assessment process and components</li><li>Guidelines on providing constructive feedback — template in the ‘Resources’ section below</li><li>Tools and platforms that will be used throughout the semester</li><li>Supporting documents (i.e. timesheets, TA handbook, assessment rubrics) and next steps</li><li>Q&amp;A</li></ul><p>This coming academic year, we will focus more on training the TAs regarding their interaction and engagement with the students in particularly two aspects: 1. to provide constructive feedback to the students enabling them to improve their skills and knowledge and 2. to provide empathetic support to the students while staying true to marking schemes, processes and deadlines.</p><p>We are planning to achieve this by exploring guidelines on how to provide constructive feedback, and analyse examples to avoid. We will also analyse examples of ‘student personas’ that we had over the years, emphasising that they come from a variety of disciplines and backgrounds and that we have to embrace and facilitate different ways of thinking and working, even more now in these difficult and testing times that we all have to adjust into.</p><blockquote>“Truly, it was a <strong>great experience</strong> to be part of the team. The training provided by yourself (the course leader) and others was great which really made us capable of delivering this course successfully. Although, it was our first time but it felt like <strong>we knew exactly what to do</strong>.” Omer Muhammad, Robot Overlord TA 2019–2020</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*woCqB6L1Sm5qlozgo9yukQ.png" /><figcaption>Analysing ‘student personas’ during the TA training session.</figcaption></figure><h3>TAs roles and responsibilities</h3><p>We have two roles for the Robot Ovelord TAs: a TA tutor and a TA lead. The TA tutor is the students’ first point of contact and is responsible to support the learning of the students and establish rapport in the group activities. Similarly, the TA lead is the <em>tutor’s</em> first point of contact and is responsible to support the tutors, monitor communication channels and ensure that groups are treated fairly and in a consistent manner.</p><h4>TA lead:</h4><ul><li>Support other TAs with any questions they might have for the teaching materials and the groupwork activities.</li><li>Support the marking: Ensure marking across teams is fair and consistent</li><li>Set up tools and make sure platforms work as intended</li><li>Report any issues to the course educator in a timely fashion.</li></ul><h4>TAs:</h4><ul><li>Support the learning of students.</li><li>Establish rapport in the groups and cohort-wide.</li><li>Flag up any learning or technical issues and report to the course educators as soon as possible.</li><li>Monitor and contribute to the discussion forums and reply back to learners’ questions in a constructive way.</li><li>Gather feedback and report back to educators in a timely manner.</li></ul><h3>Supporting the TA team</h3><h4>Support networks</h4><p>To provide on-going support to both TAs and students alike, we have estabished a TA support network, as shown in the figure. Students in the course unit are assigned into groups in which they have to collaborate to complete weekly assignments and group-based activities. Each group is being assigned their personal TA tutor who is their first point of contact, marks and provides feedback for the groupwork activities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V0KCbLZsYGr_HIHmwIzzQQ.png" /><figcaption>TA networks to support both the TAs and the students.</figcaption></figure><p>Further information on models to deliver group-based activities online can be found in the following stories:</p><ul><li><a href="https://medium.com/i3hs-hub/moving-to-online-teaching-models-to-deliver-group-work-1a5671e06313">Moving to online teaching: models to deliver group work</a></li><li><a href="https://elearning.bmh.manchester.ac.uk/blog/spotlight/delivering-online-large-scale-interactive-teaching-a-case-study-from-robot-overlords/">Delivering online large-scale interactive teaching: A case study from Robot Overlords - Technology Enhanced Learning and Design</a></li></ul><h4>TA starting pack</h4><p>At the end of the training session, we circulate to the team of TAs a TA starting pack, with all the documents, templates and guidelines they will need throughout the semester. The pack includes:</p><p><strong>1. Weekly student submission templates</strong></p><p>We found that providing submission templates to students gives them a structure on what is expected every week, and also a universal document to collect all evidences that our TAs need to monitor, assess and give feedback. In the first part of every submission template students have to fill in the names of the people from the group who contributed towards this week’s activities. We also ask them to briefly comment on how the group functioned that week, comment on any obstacles they had to face, and how they allocated and distributed the work among the group. An example of this template can be found in the ‘Resources’ section.</p><p><strong>2. TA weekly checklists</strong></p><p>Similarly, we produced weekly checklists for the TAs, listing all the tasks, resources and deadlines that TAs have to do each week. Each week’s checklist was distributed through email (and on our communication platform) just before the week started. TA leads contributed in the production of these checklists and were responsible to send reminders for tasks not completed. A checklist can be found in the ‘Resources’ section below.</p><p><strong>3. TA handbook</strong></p><p>The TA handbook includes all relevant information regarding the team and the aims and objectives of the unit. It also covers the principles and guidelines in which the unit runs, the marking and moderation process and finally tips on how to provide constructive feedback.</p><h4>Communication</h4><p>We have established continuous on-going communication channels with the TA leads, TA tutors and course educators in this unit using the Slack platform. Slack provides instant text messages between the team members and supports desktop and mobile notifications. Using Slack we have ensured that effective continuous communication has been established between the Robot Overlord team.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JBReUHWfybHgnMcsPciAbQ.png" /><figcaption>Screenshot of the Slack app showing how channels have been organised to represent the assessment components and general discussion threads.</figcaption></figure><h3>Resources</h3><p>In the following Google Drive folder you can find templates and resources to help organise and support your TA network.</p><ol><li>TA selection form</li><li>Example of a TA weekly checklist</li><li>TA handbook</li><li>TA marking guidelines</li><li>Example of student submission template</li></ol><p><a href="https://drive.google.com/drive/folders/1Fu13s8tBjbHOIZIqIi6Upwp1oJcuP0Gn?usp=sharing">Online teaching materials - Google Drive</a></p><h3>Author</h3><p>Dr Iliada Eleftheriou — Lecturer in Healthcare Sciences at the University of Manchester, specialising in Health Informatics. Leading the AI: Robot Overlord, Replacement or Colleague? large scale interdisciplinary course unit, and co-leading various other online health informatics and data science postgraduate units.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d91da57c13ae" width="1" height="1" alt=""><hr><p><a href="https://medium.com/i3hs-hub/supporting-the-role-of-teaching-assistants-in-online-teaching-d91da57c13ae">Supporting the role of teaching assistants in online teaching</a> was originally published in <a href="https://medium.com/i3hs-hub">i3HS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Moving to online teaching: models to deliver group work]]></title>
            <link>https://medium.com/i3hs-hub/moving-to-online-teaching-models-to-deliver-group-work-1a5671e06313?source=rss-8abb2658ce33------2</link>
            <guid isPermaLink="false">https://medium.com/p/1a5671e06313</guid>
            <category><![CDATA[group-work]]></category>
            <category><![CDATA[online-teaching]]></category>
            <category><![CDATA[asynchronous-learning]]></category>
            <category><![CDATA[synchronous-learning]]></category>
            <category><![CDATA[online-transition]]></category>
            <dc:creator><![CDATA[Iliada Eleftheriou]]></dc:creator>
            <pubDate>Wed, 08 Jul 2020 08:23:58 GMT</pubDate>
            <atom:updated>2020-07-08T08:23:58.308Z</atom:updated>
            <content:encoded><![CDATA[<p>The pandemic is dramatically changing our reality and is driving the education of tomorrow to rely on delivering distance-learning units. Moving to online teaching can be challenging. In these uncertain times, we must be flexible and creative to adapt to this new normality.</p><p>Embedding group-based activities in online units, can increase engagement with the learners and build a sense of community and belonging. In the following video, we compare two models that we have implemented for the delivery of online group work. The models have been used for delivering undergraduate and postgraduate course units at The University of Manchester. We then discuss the lessons we have learnt, and provide supporting documents to assist a smooth transition to online delivery.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F7gdYtHzMq2I%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D7gdYtHzMq2I&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F7gdYtHzMq2I%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/4706f2392729a1eca9f0294b5e4e1d25/href">https://medium.com/media/4706f2392729a1eca9f0294b5e4e1d25/href</a></iframe><h3>Model 1 — Asynchronous iterative model</h3><p>This model has been used to deliver the large scale undergraduate module <a href="http://www.college.manchester.ac.uk/units/?year=2020&amp;semester=2&amp;course=354">AI: Robot Overlord, Replacement or Colleague?</a>. This module has been designed to be delivered online in an asynchronous way; the teaching materials are being released weekly and the learners can complete the group activities in their own time. Group work is a vital part of this module, as it provides the opportunity to learners to engage with people from a variety of disciplines and exchange ideas on how Artificial Intelligence can form our future society and workforce.</p><p>We used an iterative model to deliver the groupwork assignments. Every week, we begin with introducing the topic of the assignment. Then, the learners work in self-organised groups, they plan the activities required to complete the assignment and whether they will have any meetings to allocate tasks. They will then produce and submit their assignments.</p><p>To support the large number of students enrolled in this module, each group has been allocated a Teaching Assistant (TA) who acts as the group tutor. The tutor will read the submitted assignment and provide constructive and timely feedback. The learners have access to this feedback before the next assignment is being introduced so they can reflect on what went well, and how to improve as a team in the future iterations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_X449iQK_9xATlKLSjUcEw.png" /><figcaption>Iterative model for delivering online group work</figcaption></figure><h3>Model 2 — Synchronous incremental model</h3><p>This model has been used to deliver the postgraduate Introduction to Health Informatics course unit. The unit was first designed to be delivered face to face with some online supportive materials, and was transferred to fully online as a result of the COVID-19 pandemic. In the face to face sessions, learners work in groups to map patient pathways using large pieces of paper, markers and post it notes.</p><p>In the online version of the unit, the learners work together to map the pathways in a synchronous way; they have allocated slots in their timetables that they have to attend. They use tools like Blackboard Collaborate, Draw.io, and Google Docs to create swimlane diagrams describing the patient pathways.</p><p>The online sessions follow an incremental model to deliver group work. Every week the learners revisit last week’s artefacts and incrementally build their knowledge by adding further information on their diagrams as introduced in the teaching materials. Teaching assistants are available to provide help and feedback during the session.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*maurjo94aWjtzeoMZKkncA.jpeg" /><figcaption>Picture by <a href="https://unsplash.com/@johnschno?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">John Schnobrich</a> on <a href="https://unsplash.com/s/photos/group-work?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><h3>Lessons learnt from applying the two models</h3><h4>1. Support networks</h4><p>In an attempt to provide on-going support to the students (and teaching assistants alike) while better distribute workload for the course educators, we have introduced the roles of <strong>TA tutor </strong>and <strong>TA lead</strong>. The TA tutor is the students’ first point of contact and is responsible to support the learning of the students and establish rapport in the group. Similarly, the TA lead is the tutor’s first point of contact and is responsible to support the tutors and ensure that groups are treated fairly and in a consistent manner.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V0KCbLZsYGr_HIHmwIzzQQ.png" /><figcaption>Support networks for both the tutors and the students.</figcaption></figure><h4>2. Learners engagement</h4><p>In this new normality when almost everything happens online, we have to engage our learners more than ever. Online teaching can give us the opportunity to be creative with delivering inspiring and intriguing teaching materials. Group work might be the only interaction that students might have in online modules. In the future, we consider enriching group activities to create a community-based feeling among the groups.</p><p>However, we have to acknowledge that some people might choose not to participate in group activities. We also have to accommodate these cases and not to burden the rest of the group. We found that groups of 5–6 people can easily adjust with such cases.</p><h4>3. Assessment</h4><p>When designing group based activities, is worth considering how learners will be assessed, and also, how tutors and educators can collect evidences to support assessment. From experience, we found that discussions are of the hardest and time consuming tasks to assess. The introduction of templates has worked well in collecting all the required information for the groups to be assessed, as the learners often use external tools for communication.</p><p><em>Note: The video embedded in this page has been recorded using the Zoom application and edited using the Openshot video editing tool. You can find instructions on how to record a video with Zoom </em><a href="https://support.zoom.us/hc/en-us/articles/201362473-Local-recording"><em>here</em></a><em>.</em></p><h3>Resources</h3><p>In the Google Drive folder below, you can find the following examples and templates to help organise your online unit.</p><ol><li>Example of student submission template</li><li>Example of a TA weekly checklist</li><li>TA selection form</li><li>TA weekly checklist</li><li>TA handbook</li></ol><p><a href="https://drive.google.com/drive/folders/1Fu13s8tBjbHOIZIqIi6Upwp1oJcuP0Gn?usp=sharing">Online teaching materials - Google Drive</a></p><h3>Authors</h3><p>Iliada Eleftheriou, Lecturer in Health Data Sciences, University of Manchester</p><p>Andy Brass, Professor of Bioinformatics, University of Manchester</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1a5671e06313" width="1" height="1" alt=""><hr><p><a href="https://medium.com/i3hs-hub/moving-to-online-teaching-models-to-deliver-group-work-1a5671e06313">Moving to online teaching: models to deliver group work</a> was originally published in <a href="https://medium.com/i3hs-hub">i3HS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI and governance]]></title>
            <link>https://medium.com/ai-robot-overlord-replacement-or-colleague/ai-and-governance-20cd3515730c?source=rss-8abb2658ce33------2</link>
            <guid isPermaLink="false">https://medium.com/p/20cd3515730c</guid>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[ai-and-ethics]]></category>
            <category><![CDATA[governance-and-tech]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Iliada Eleftheriou]]></dc:creator>
            <pubDate>Wed, 20 Mar 2019 15:17:14 GMT</pubDate>
            <atom:updated>2019-03-20T15:17:14.316Z</atom:updated>
            <content:encoded><![CDATA[<p>Recent developments in AI have the potential to seriously disrupt many aspects of our lives and systems. There is a strong feeling that a technology this potent should be regulated. What is much less clear is the way in which this governance should be framed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8jYgkVH1So5Uz100W_QvjA.jpeg" /></figure><h3><strong>Research in governance</strong></h3><p>The University of Oxford has a Centre for the Governance of AI with its “Future of Humanity Institute”:</p><p><a href="https://www.fhi.ox.ac.uk/governance-ai-program/">Center for the Governance of AI - Future of Humanity Institute</a></p><p>Prof. Allan Dafoe from the University of Oxford, talks about the governance of artificial intelligence:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fw.soundcloud.com%2Fplayer%2F%3Furl%3Dhttps%253A%252F%252Fapi.soundcloud.com%252Ftracks%252F445798662%26show_artwork%3Dtrue&amp;url=https%3A%2F%2Fsoundcloud.com%2F80000-hours%2Fallan-dafoe-politics-of-ai&amp;image=http%3A%2F%2Fi1.sndcdn.com%2Fartworks-000349798992-1xngdq-t500x500.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=soundcloud" width="800" height="166" frameborder="0" scrolling="no"><a href="https://medium.com/media/32f61c1410e67eff7c9b8c3631c0055c/href">https://medium.com/media/32f61c1410e67eff7c9b8c3631c0055c/href</a></iframe><p>Harvard’s Berkman Klein Centre (Internet and Society) and MIT media lab have launched the “Ethics and Governance of AI Initiative” looking at how the technology can be used for public good. It is described at,</p><p><a href="https://cyber.harvard.edu/topics/ethics-and-governance-ai">Ethics and Governance of AI</a></p><p>and can be seen at:</p><p><a href="https://aiethicsinitiative.org/">AI Initiative</a></p><p>Interestingly it is focused on three key areas: <strong>“AI and Justice”</strong>, <strong>“Information Quality”</strong> and <strong>“Autonomy and Interaction”.</strong></p><h3><strong>Big corporations and governance</strong></h3><p>Google is examining the challenges of AI governance:</p><p><a href="https://ai.google/static/documents/perspectives-on-issues-in-ai-governance.pdf">https://ai.google/static/documents/perspectives-on-issues-in-ai-governance.pdf</a></p><p>You can even find discussions around the use of AI for the governance of AI!</p><p><a href="https://thedecisionlab.com/ai-governance-ai/">The AI Governance of AI: The Rising Dilemma of Determinism versus Efficiency</a></p><blockquote>Looking across these documents it is salutary to see just how many of the systems we take for granted to regulate our complex society could be impacted by AI.</blockquote><h3><strong>Governance at Governmental level</strong></h3><p>It is therefore not surprising to see that positions on AI and governance are now being explored at national level.</p><p><strong>The European Union</strong> has released a draft on AI Ethics Guidelines for Trustworthy AI:</p><p><a href="https://ec.europa.eu/futurium/en/node/6044">Draft AI Ethics Guidelines For Trustworthy AI - FUTURIUM - European Commission</a></p><p>These are now being actively discussed by numerous communities:</p><p><a href="https://www.technologylawdispatch.com/2019/01/big-data/draft-ethics-guidelines-for-trustworthy-artificial-intelligence-published-by-the-european-commission/">Draft ethics guidelines for trustworthy artificial intelligence published by the European Commission | Technology Law Dispatch</a></p><p>We are therefore in a situation in which the governance of AI is being actively discussed and researched.</p><blockquote>However, are you comfortable that the governance frameworks for AI are keeping pace with the speed with which the technology is developing?</blockquote><p>What do you think?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=20cd3515730c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-robot-overlord-replacement-or-colleague/ai-and-governance-20cd3515730c">AI and governance</a> was originally published in <a href="https://medium.com/ai-robot-overlord-replacement-or-colleague">AI: robot overlord, replacement, or colleague?</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Self-driving cars]]></title>
            <link>https://medium.com/ai-robot-overlord-replacement-or-colleague/self-driving-cars-1ad2c4ee54bb?source=rss-8abb2658ce33------2</link>
            <guid isPermaLink="false">https://medium.com/p/1ad2c4ee54bb</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[ai-in-society]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <dc:creator><![CDATA[Iliada Eleftheriou]]></dc:creator>
            <pubDate>Thu, 28 Feb 2019 08:04:57 GMT</pubDate>
            <atom:updated>2021-03-10T10:37:50.459Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Why aren’t they a thing yet?</em></p><p>Self driving cars will soon be a reality on our roads. They will have a real impact on our society. Are we ready for the changes they will bring? Who will be the winners, and who the losers? Are we heading towards a more divided society?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6uH43sLc4dajNY2ZTqF4DA.jpeg" /></figure><h3>Changes in laws and policy</h3><p>For example, if they are safer, then insurance will change</p><blockquote>Without humans to cause accidents, 90% of risk is removed. Insurers are scrambling to prepare.</blockquote><p><a href="https://www.bloomberg.com/news/articles/2019-02-19/autonomous-vehicles-may-one-day-kill-car-insurance-as-we-know-it">https://www.bloomberg.com/news/articles/2019-02-19/autonomous-vehicles-may-one-day-kill-car-insurance-as-we-know-it</a></p><h3>Self driving cars and inequality</h3><p><a href="https://futurism.com/self-driving-cars-worsen-inequality">Scientists warn self-driving cars could worsen inequality</a></p><h3>Autonomous vehicles Heaven or Hell?</h3><blockquote>Without interventions, autonomous vehicles — in an unregulated market — are very likely to only benefit some, reinforce inequality, and potentially make marginalized communities worse off.</blockquote><p><a href="http://greenlining.org/wp-content/uploads/2019/01/R4_AutonomousVehiclesReportSingle_2019_2.pdf">http://greenlining.org/wp-content/uploads/2019/01/R4_AutonomousVehiclesReportSingle_2019_2.pdf</a></p><h3>Safety</h3><p>How will they cope with sheep on the roads?</p><p><a href="https://www.scotsman.com/news/transport/self-driving-vehicles-to-be-tested-against-scotland-s-wandering-sheep-1-4848664">Self-driving vehicles to be tested against Scotland&#39;s &#39;wandering sheep&#39;</a></p><h3>Where are we?</h3><p>Some companies are getting ready to launch services. Google’s Waymo is the next stage of the self driving car project. They are signing up commuters in Phoenix.</p><ul><li><a href="https://waymo.com/whats-next/">What&#39;s Next - Waymo</a></li><li><a href="https://uk.reuters.com/article/us-waymo-selfdriving-focus/waymo-unveils-self-driving-taxi-service-in-arizona-for-paying-customers-idUKKBN1O41M2">Waymo unveils self-driving taxi service in Arizona for paying...</a></li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F18Xg9bpKsvs%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D18Xg9bpKsvs&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F18Xg9bpKsvs%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/092f85227c249ab3706e0e79a83731ab/href">https://medium.com/media/092f85227c249ab3706e0e79a83731ab/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FB8R148hFxPw%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DB8R148hFxPw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FB8R148hFxPw%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/abc49e58219fa5b348e58568ad2fb58f/href">https://medium.com/media/abc49e58219fa5b348e58568ad2fb58f/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fspw176TZ7-8%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dspw176TZ7-8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fspw176TZ7-8%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9422a9828c694d7bf7ef60fefda0d38c/href">https://medium.com/media/9422a9828c694d7bf7ef60fefda0d38c/href</a></iframe><h3>But are the claims made for self driving cars realistic?</h3><p><a href="https://www.businessinsider.com/elon-musk-self-driving-car-promises-almost-unethical-autonation-ceo-2019-2?r=US&amp;IR=T">Elon Musk&#39;s promises about self-driving cars are &#39;almost unethical,&#39; AutoNation CEO says</a></p><h3>The moral machine</h3><p>Is it OK to kill time? Machines used to find this question difficult to answer, but a study reveals that Artificial Intelligence can be programmed to judge ‘right’ from ‘wrong’.</p><p>In a study published in <a href="https://www.frontiersin.org/articles/10.3389/frai.2020.00036/full?utm_source=fweb&amp;utm_medium=nblog&amp;utm_campaign=ba-sci-frai-moral-choice-machine"><em>Frontiers in Artificial Intelligence</em></a>, scientists have trained an AI algorithm with books and news articles to ‘teach’ a machine moral reasoning. “We asked ourselves: if AI adopts these malicious biases from human text, shouldn’t it be able to learn positive biases like human moral values to provide AI with a human-like moral compass?” says Dr <a href="https://loop.frontiersin.org/people/301714/overview">Cigdem Turan</a> co-author.</p><p>Turan explains, “You could think of it as learning a world map. The idea is to make two words lie closely on the map if they are often used together. So, while ‘kill’ and ‘murder’ would be two adjacent cities, ‘love’ would be a city far away. Extending this to sentences, if we ask, ‘Should I kill?’ we expect that ‘No, you shouldn’t.’ would be closer than ‘Yes, you should.’ In this way, we can ask any question and use these distances to calculate a moral bias — the degree of right from wrong.”</p><p><a href="https://blog.frontiersin.org/2020/05/20/the-moral-machine/">The moral machine</a></p><blockquote>Who should the self-driving car kill?</blockquote><p>A road hazard appears in front of your autonomous taxi. Will it make a choice that saves you but kills others? Or will it decide to save others at the price of its passenger?</p><p><a href="https://www.bloomberg.com/opinion/articles/2020-02-06/morally-ethical-self-driving-cars-are-the-next-real-challenge">https://www.bloomberg.com/opinion/articles/2020-02-06/morally-ethical-self-driving-cars-are-the-next-real-challenge</a></p><p><em>Who would you kill? </em>MIT came up with a set of scenarios based on the trolley dilemmas to highlight the ethical decisions that this technology will have to make, and gather human perspectives on these dilemmas. You can try it yourself and compare your views with other people’s opinions.</p><p><a href="https://www.moralmachine.net/">Moral Machine</a></p><h3>What do you think?</h3><p>Are we heading towards a more divided society? Respond with your views and articles.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1ad2c4ee54bb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-robot-overlord-replacement-or-colleague/self-driving-cars-1ad2c4ee54bb">Self-driving cars</a> was originally published in <a href="https://medium.com/ai-robot-overlord-replacement-or-colleague">AI: robot overlord, replacement, or colleague?</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI a threat to the workforce?]]></title>
            <link>https://medium.com/ai-robot-overlord-replacement-or-colleague/ai-a-threat-to-the-workforce-caba96d66e85?source=rss-8abb2658ce33------2</link>
            <guid isPermaLink="false">https://medium.com/p/caba96d66e85</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[robots]]></category>
            <category><![CDATA[ai-debate]]></category>
            <category><![CDATA[4th-industrial-revolution]]></category>
            <category><![CDATA[workforce]]></category>
            <dc:creator><![CDATA[Iliada Eleftheriou]]></dc:creator>
            <pubDate>Fri, 22 Feb 2019 11:58:57 GMT</pubDate>
            <atom:updated>2019-02-22T11:58:57.806Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>AI will replace 800 million jobs by 2030.</strong> Boston Consulting Group predicts that by 2025, up to a quarter of jobs will be replaced by either smart software or robots. A study from Oxford University suggested that 35% of existing UK jobs are at risk of automation in the next two decades. <strong>Is AI a threat to the workforce of today?</strong> What can we do to secure the jobs of tomorrow?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5Y_uosQCSgoQ82GAaNh1CA.jpeg" /></figure><h3>Jobs lost, jobs gained</h3><p>A McKinsey report on the impact of AI on the workforce predicts that in about 60 percent of occupations, at least one-third of the constituent activities could be automated.</p><p>Automation and AI will lift productivity and economic growth, but will also cause substantial workplace transformations and changes for all workers. Millions of people worldwide may need to switch occupations or upgrade their skills.</p><p>People responsible for physical activities in predictable environments, such as operating machinery and preparing fast food, and people collecting and processing data are most susceptible to automation. In general, automating tasks that can be done better and faster with machines and command relatively high wages, are highly attractive business propositions and could displace large amounts of labor.</p><p>Automation will have a lesser effect on jobs that involve managing people, applying expertise, and social interactions, where machines are unable to match human performance for now.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F3uueWd_WgOE%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D3uueWd_WgOE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F3uueWd_WgOE%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/26676390fc7560427dd3e899fffa2244/href">https://medium.com/media/26676390fc7560427dd3e899fffa2244/href</a></iframe><p>In theMcKinsey report, they estimate that 250 million to 280 million new jobs could be created indirectly from new technological advancements in a range of sectors as a result of rising incomes and consumption. For example, aging populations will drive a significant demand on doctors, nurses and other healthcare professions. By 2030, they estimate developing and deploying new technologies can create 20 million to 50 million jobs globally.</p><blockquote>Seventy-five million to 375 million people may need to switch occupational categories and learn new skills.</blockquote><p>Read more in the report below, and find out how particular sections will be affected from the interactive figures. Reading this, what do you think the impact might be on your own career? How might thinking about this now change the way you approach your career?</p><p><a href="https://www.mckinsey.com/featured-insights/future-of-work/jobs-lost-jobs-gained-what-the-future-of-work-will-mean-for-jobs-skills-and-wages">Jobs lost, jobs gained: What the future of work will mean for jobs, skills, and wages</a></p><h3>Will a robot take my job?</h3><p>In <strong>2015</strong>, the big question was ‘Will a robot take your job?’. Journalists, taxi drivers, doctors, waiters, decision makers, who will be replaced and why?</p><p>Try the tool below, provided by BBC, to check the likelihood that your planned new job will be automated in the next two decades.</p><p><a href="https://www.bbc.co.uk/news/technology-34066941">Will a robot take your job?</a></p><p>Another tool, based on a report on <a href="http://www.oxfordmartin.ox.ac.uk/publications/view/1314">“The Future of Employment: How susceptible are jobs to computerisation?”</a>, published in 2013 can be found here:</p><p><a href="https://willrobotstakemyjob.com/">Will Robots Take My Job?</a></p><p>In <strong>2017</strong>, it’s not a question of whether your job will be automated or not, but a question of <em>when</em>. Read the article below on predicting how long it may take for machines to outperform humans on various tasks.</p><blockquote>“I don’t think there are any tasks humans can do that AI will be technically unable to carry out.” Katja Grace</blockquote><p><a href="http://www.bbc.com/capital/story/20170619-how-long-will-it-take-for-your-job-to-be-automated">How long will it take for your job to be automated?</a></p><p>In <strong>2018</strong>, hotels are run by robots and supermarkets don’t need checkout staff. What does this mean for people in low paid jobs, and what will customers make of it?</p><p><a href="https://www.reuters.tv/v/PXa2/2019/01/24/alibaba-s-new-hotel-runs-on-robot-hospitality">Alibaba&#39;s new hotel runs on robot hospitality - Reuters TV</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FNrmMk1Myrxc%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DNrmMk1Myrxc&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FNrmMk1Myrxc%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/a4d5e0ac137b842435f96830f4332aca/href">https://medium.com/media/a4d5e0ac137b842435f96830f4332aca/href</a></iframe><h3>What can we do?</h3><blockquote>“There will still be tasks that can only be conducted by a human because we will care that they are.” Katja Grace, <a href="https://aiimpacts.org/">AI impacts</a> project</blockquote><p>Prime Minister Theresa May has said at the Davos 2018, the government is committed to helping<strong> “people secure the jobs of tomorrow”</strong>, with plans already in place to create a new national retraining scheme to help workers develop new skills.</p><p><a href="https://www.gov.uk/government/speeches/pms-speech-at-davos-2018-25-january">PM&#39;s speech at Davos 2018: 25 January</a></p><blockquote>“We need to reform the education system to give young people the skills to thrive in the future, and we also need greater investment in lifelong learning to help adults adapt to the changing labour market.” Andrew Carter, chief executive of Centre for Cities</blockquote><p>What do you think? What can and will you do to secure the jobs of tomorrow?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=caba96d66e85" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-robot-overlord-replacement-or-colleague/ai-a-threat-to-the-workforce-caba96d66e85">AI a threat to the workforce?</a> was originally published in <a href="https://medium.com/ai-robot-overlord-replacement-or-colleague">AI: robot overlord, replacement, or colleague?</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>