STUDENT CHATBOT

.: Los Angeles Unified School District (LAUSD) Launches AI Chatbot “Ed” to Serve as Virtual Student Advisor​

Summary ➜ The AI chatbot is designed to be a virtual advisor for students, aiming to simplify school navigation and support pandemic recovery. Ed can provide personalised guidance, share academic details, and make recommendations, while even assisting with non-academic issues like lost bikes. But some parents and experts have privacy and over-reliance concerns, worried it may replace human connections. For now, the chatbot is available to 54,000 students at 100 “fragile” schools, with plans to expand. Ed looks to create individual learning plans modelled on special education IEPs, touting a 93% accuracy rate over ChatGPT’s 86%.

Why this matters for education ➜ In Australia, two student chatbots are being trialled in New South Wales and South Australia, respectively. The LAUSD version’s focus is on practical student support, including quick access to assessments and grades.

So, for now, it seems to be much more like a student support service than a generative AI system for teaching and learning.

An interesting note about this development is that the initial design regarding access to information is quite closed.

From the article, it is unclear whether this is just up-to-date news or no information at all. For comparison, South Australia’s Edchat does not have real-time updated information, but students can access training data up to early 2023.

.: Other News In Brief

​

🇸🇬 Singapore university sets up AI research facility for ‘public good’

⚽️ As AI football looms, be thankful for those ready to rage against the machine

🌇 Stability AI CEO resigns to ‘pursue decentralised AI’

🤖 OpenAI shows off the first examples of third-party creators using Sora

📈 Nvidia’s AI chip dominance is being targeted by Google, Intel, and Arm

❓ Where’d my results go? Google Search’s chatbot is no longer opt-in

🎤 Can you hear me now? AI-coustics to fight noisy audio with generative AI

📜 The AI Act is done. Here’s what will (and won’t) change

🗣️ HeyGen is About to Close $60M on a $500M Post Valuation

🇮🇳 India reverses AI stance, requires government approval for model launches

​

Monthly Review

.: All the February issues in 1 PDF

​

​

Download a copy of the latest monthly roundup of Promptcraft.

What’s on my mind?

.: A Quick Critical Analysis of Student Chatbots

In this week’s reflection, I want to revisit walled garden chatbots for students.

I think this strategic path needs more attention after the news that Los Angeles schools have access to a student support chatbot and the active trials in New South Wales and South Australia.

It is also worth adding that I have a partnership with Adelaide Botanic High School, one of the first schools in 2023, to trial Edchat, the chatbot for South Australian schools.

It has been great to have access to it and work alongside some of the leadership team running the project at the school.

​

When I say walled garden chatbots, I mean chatbots that operate within a closed system or a specific domain, having access only to a limited information set.

These chatbots are designed by the school system the student belongs to, as opposed to the consumer chatbots or AI products on the open web.

For this reflection, I will use the Compass Points thinking routine from Project Zero, which starts in the East.

​

E: Excites

• Students get hands-on experience with AI tools to support their learning experience.
• There is a lot of momentum which can grow if our students are given the chance to step up.
• The LA example offers something different from the Australian trials: a chatbot more focused on support and practical companionship. I am excited to see how this develops.

​

N: Need to Know

• How do teachers feel about students gaining access to these tools?
• What professional growth opportunities exist for educators to build their capacity and understanding?
• What are the system prompts for these bots, and how do they mitigate and guard against bias?

​

S: Stance

• I support the testing and exploration of chatbots safely in schools. Many other school systems are keen to learn from their pioneering examples.
• Educational innovation like these chatbot trials must be supported, encouraged and celebrated.
• Despite the guard railing, I can see that frontier models like GPT-4 power them and can be very flexible in support of a student.

​

W: Worries

• The design, prompt, and architecture are not visible. Without transparency, it’s difficult to hold developers and operators accountable for any issues that may arise.
• Chatbots and LLMs are designed to respond with the most likely next word. They are geared towards the statistically most common. Without fine-tuning and promptcraft education, this might homogenise the message to our students.
• I am concerned about the pedagogical bias built into students’ AI systems. Imagine a student is an active user with over 100 interactions daily. 1000s every month. What’s the hidden curriculum of each of those nudges and interactions?

:. .:

​

~ Tom

Prompts

.: Refine your promptcraft

​

One of the most useful and effective prompting techniques is to include an example of what you want to generate. This is called few-shot prompting, as you guide the AI system with a model of what to generate.

For example, if you are trying to develop some model texts or example paragraphs to critique with your class and you have some from the last time, you could add these to your prompt.

When you do not include an example, this is called zero-shot prompting. Zero-shot prompting always leads to a less aligned output, which you must edit and iterate to correct.

This week’s quick tip builds on the few-shot prompting technique by adding some qualifying explanations for ‘what a good one looks like’.

We always do this when we are modelling and teaching, and it works well with promptcraft.

So the prompt structure goes:

1. Describe your instructions in clear language.
2. Add an example of what you are looking for.
3. Describe why this example is a good model.

Let’s say you’re teaching persuasive writing. Instead of asking the AI to generate a persuasive text, you could add a model paragraph from a previous lesson, followed by specific pointers on what makes it effective.

Here’s an example working with the Claude-3-Opus-200k model.

​

Note how closely the output follows my model paragraph. A little extra promptcraft goes a long way to improve the results.

:. .:

Remember to mak​e this your own, try different language models and evaluate the completions.

Learning

.: Boost your AI Literacy

​

BIAS

.: Systematic Prejudices Investigation [PDF]

The UNESCO report sheds light on gender bias within artificial intelligence, revealing biases in word association, sexist content generation, negative content about sexual identity, biased job assignments, and stereotyping.

The study emphasises the need for ethical considerations and bias mitigation strategies in AI development, including diverse representation in teams and training datasets to ensure fairness and inclusivity.

​

ELECTION ​
​.: How AI companies are reckoning with elections​

A helpful short article giving a rundown of how some of the largest tech companies and AI platforms are grappling with the impact of AI tools on the democratic process.

Although this focuses on the US, in 2024 billions of people will go to the polls across the world. According to a Time article:

How these popular generative AI companies respond will impact so many of us.

The “Seven Principle Goals” of the AI Elections accord. Image: AI Elections accord

​

LABELS ​
​.: Why watermarking won’t work

This Venturebeat article discusses the challenges posed by the proliferation of AI-generated content and the potential for misinformation and deception.

Tech giants like Meta, Google, and OpenAI are proposing solutions like embedding signatures in AI content to address the issue.

However, questions arise regarding the effectiveness and potential misuse of such watermarking measures.

​

Ethics

.: Provocations for Balance
​

The Chatbot That Knows Too Much

“Ed” is meant to be a helpful companion, guiding students through their day. But imagine a school system where every question asked, every late assignment admitted to, every awkward social situation confessed to the chatbot becomes permanent record. Mistakes can’t be erased; the AI analyzes your word choice for signs of distress.

What if this data isn’t just used for support, but for discipline, even predicting future “problem” behavior? Is it okay for an AI to judge the private thoughts of a teenager, and worse, limit their opportunities based on what it predicts they might do?
​

Inspired by some of the topics this week and dialled up.

How would you rate this issue of Promptcraft?

(Click on your choice below)
​

​⭐️⭐️⭐️⭐️⭐️ Loved it!​
​
​⭐️⭐️⭐️ Good​
​
​⭐️ Poor​

​

If you have any other kind, specific and helpful feedback, please reply to this email or contact me at tom@dialogiclearning.com​

​
.: Tom Barrett