Generative AI : Education and Work disrupted
Everyone agrees on one thing : Generative AI models such as ChatGPT have a lot of potential for disruption. But what will be their main impacts on Education and Work?
Will they make it easier for students to cheat? Or allow them to be better writers?
Will journalists and designers be out of a job? Or will AI enable them to be better creators?
Let’s dive into it.
A high-potential, yet immature technology
Generative AI is having a moment. After a few years of steadily improving results, the latest demonstrations from Dall-E 2, ChatGPT, and others, have put the technology in the limelight.
Why? Beyond the fun of playing with it, the technology is getting close to (or better than) human performance at a number of tasks, especially text and image generation. It is both impressive and uncanny.
Generative AI is improving extremely quickly, thanks to :
- Technological breakthroughs (starting with Transformers in 2017) and
- Very large investments from a few key players, who have both the cutting-edge knowledge and the capital to create large models (thanks to funding from Google, Facebook, Microsoft, etc…). Training sets are huge, and a single training run of GPT-3 costs millions of dollars (source)
I won’t do the now-overused trick of “an AI has written this article” because it’s only convincing for a couple of paragraphs, and even though it sounds believable, it isn’t always right or insightful, an issue organisations like Stackoverflow are now struggling with. But that’s already much better than a few years ago, when a toddler would sound more coherent than the best models.
The truth is, chatGPT is a tech demo and developer tool, not a fully-formed ready-to-use product for consumers and pros. So don’t expect it to write a full page of beautiful content based on a single prompt. But the technology, properly integrated into a product, has a huge potential.
What can it do? The models can create text, images, audio, video, but also data and code. They’re also helpful for tasks derived from there : summaries, classification (sentiment), translation, evaluation, questions (asking and answering), search, and a lot of others.
What for? It’s now good enough to automate certain tasks — evaluating simple answers in natural language or speech for example. For more complex tasks it can act as an assistant to augment human work. This approach works already at scale : Github copilot generates c.40% of new code for the 1 million+ developers using it (source).
It’s a feature, not a product : the technology will be embedded within new and existing products, as Github has done. You may not even know it’s used, like a lot of other AI applications today. And anyone can use it fairly easily, from tech giants to start-ups and traditional corporates. No need to rebuild and re-train hugely complex models such as GPT-3
It’s still a limited technology. Don’t expect one AI assistant to do everything for you. Generative AI understands language and how we communicate, but only partially grasps the meaning behind it, so it makes stupid mistakes, and is fairly easy to trick. And because it’s trained on internet content, it shares its flaws.
Our belief is that this next cycle of use cases will focus on vertical solutions. Although new breakthroughs may come quickly, as GPT-4 is just around the corner.
So let’s have a look at generative AI can do — with a focus on the Future of Education and Work, since that’s our area of expertise.
Entrepreneurs are exploring a lot of use cases
These are just part of the examples we found. You’ll see below a mapping of innovators in Education and Future of Work.
Let’s go through some use cases, from general to specific
Writing : if you’ve tried it, you know ChatGPT is very good at writing text on pretty much any subject, with complete, articulate and grammatically correct sentences and paragraphs
Use cases : an AI assistant to help you write, whether you’re a student writing an essay, a marketer or salesperson, writing an internal memo. Image models can also create images, pictograms that fit the topic
- Jasper allows you to write articles, blog posts, marketing copy, sales emails or optimize for SEO.
- Notion recently launched its own writing AI, although it’s still in beta
- Scraft focuses on helping students write essays, guiding them through the process
Improving your writing : closely related to the first ones, some tools leverage the technology to evaluate how you write and suggest improvements
The use case : Everyone. Who never makes mistakes while writing?
- Grammarly’s focus is grammatical error correction. It has been implementing new AI models to improve its core service and add other features such as correcting style, clarity or helping you write according to your company’s guidelines. These capabilities are also more and more integrated into Microsoft’s Office suite or Google Docs.
- MerciApp is focused on the same problem, and working on doing all this for languages other than English (you know, the other 7.6 billion people on earth), starting with French. A key addition : it tells you which grammar rule you got wrong.
Chatbots : I don’t know about you, but I’ll go to great lengths to avoid using a chatbot. A lot of effort went into them, but the technology wasn’t very convincing. Now? It might be. In any case, companies love using them, because they save money.
The use case : customer service is a huge market, either as a replacement or to augment humans. It also has a lot of potential for sales. In education, they can help to teach, manage student questions
- Foondamate: offers a whatsapp studybot
- Mainstay or Upswing: use their chatbots to improve student engagement
- Ada is a new generation chatbot for customer service
Evaluating, grading and giving feedback : more specific to education, grading and giving feedback is a very time-consuming and tricky task for teachers.
The use case : making grading faster and precise for teachers, or even grading automatically for certain type of questions (text, equations)
- Gradescope uses AI to streamline grading by reading student’s answers and grouping them, allowing the teacher, for example, to give full grades to all 25 students who answered correctly at once, instead of one by one
- Packback: offers instant feedback for written assignments, in addition to a chatbot to manage student questions & communities
Learning to read : by combining speech recognition with language processing, products can give live feedback to children, which enables much better digital learning
The use case : helping children learn to read faster
- Letrus: uses AI to give instant feedback on their writing to learners, supplemented by human feedback asynchronously
- Ello and Amira: with adaptive learning and speech recognition, they listen to children reading and correct them or help them when they’re stuck
Creating educational content : piecing together AI capabilities such as identifying key concepts, generating questions and false answers, one can build much more complex use cases
Use cases : creating training documentation, videos or even brand new stories
- Nolej.io can turn a YouTube video about any topic into a micro-learning, highlighting key concepts, generating quizzes in many formats and a summary
- Koalluh creates engaging children stories based on selected themes and characters, by generating a complete story and illustrations
- Duolingo used generative AI to create reading comprehension texts and questions for its standardized English test (the paper is linked)
Foreign languages
Use cases : translation or language learning
- Jasper allows user to easily translate the content they just wrote to many languages
- Speak or Blue Canoe help foreign language learners get better with instant feedback
Knowledge management :
Use cases : knowledge management, organizing work and meeting assistants
- Glean and Sana Labs connect to many apps in large companies and help users retrieve knowledge that might not be formalized by answering questions, summarizing and organizing content, etc…
- Mem aims to use AI to better organize knowledge and work
- Sembly and Otter act as AI meeting assistants. They will listen to meetings, write a transcript, summarize the discussion and identify actions. Still a challenging proposition given today’s technology
Scientific research :
- Elicit: Uses language models to automate & accelerate the research process & help anyone become an expert in anything, quickly.
- Explainpaper: Enables users to upload complex, specialised papers, highlight confusing text & get an explanation.
Looking forward
While not all of these offerings will fullfill their promises or find a market, it’s impressive to see the potential. It could become a true disruption for white collar work, automating part of the work and augmenting the rest. As investors, we’re convinced this could be a game-changer in Education and Future of Work.
For society, it will definitely raise some tricky questions. In education, students have been quick to adopt the new tool and educators are reacting already, asking whether one can develop critical thinking skills when an AI can provide ideas and writing. In the workplace, automation could mean more people end up jobless.
Overall, it’s very hard to forecast the impact new technologies will have and it sometimes takes years to become clear, for good or ill (as with social networks, the gig economy, etc…).
If you’re leveraging Generative AI to create the future of education and work, we’d be happy to speak with you (and add you to the mapping if you’re not in there). Just reach out to us!
