Lesson 6: Strengths and Weaknesses

​Strengths and Weaknesses

By Zac Chambers

Use Brain Camp AI to generate the most full glass of wine.

Upload your best image to Padlet to show the class!

Warm up activity:

Sharing your images

​When you have an image you're happy with, submit it to the class portfolio on your phone using Padlet.

Click on the image, then click the share button

When the share screen pops up, click "Padlet"

And then share it to the AI Art Showcase

Wine is traditionally served only half full, so all the images online are only half full!

Since our AI is trained on these images, it has a really hard time breaking that pattern.

Training data biasses

Try generating an analog clock that shows the time 2:10

What time does Brain Camp AI generate instead?

Why could this be?

One more challenge:

Humans like symmetry, so advertisements for watches usually set the time to 10:10

And since these photos went into the training data, the AI struggles to break this pattern.

Training data biasses

Open Ended

Not all patterns are beneficial to replicate!

What are some examples of patterns in society that we don't want AI to learn?

Undesired biases affect humans, not just images


How do companies prevent their AI models from generating discriminatory content?

Undesired Patterns

Controlling AI Biases: Filtered Training Data

​We can control where we collect our training data from, and ensure that the model trains on a representative data set.

Humans, or specialized AI systems can help filter this data.

If we can remove biases from the training data, the AI will be less likely to replicate these biases.

Multiple Select

How could AI designers eliminate biases when creating photos of watches?

• Hire humans to remove the extra 10:10 photos from the training data

• Include more photos of watches with other times in the training data

• Ask watch manufacturers to make fewer watches.

• Design a watch with 13 hour marks, instead of 12.

Controlling AI Biases: System Prompts

System prompts are hidden instructions, included in every conversation with the user.

Most AIs keep their System prompts a secret, like a combination to a safe.

Controlling AI Biases: System Prompts

​Sometimes filtering the data is too difficult.

Even if the AI is able to discriminate, we can instruct it not to, in it's system prompt.

AI designers can include invisible instructions that prohibit the AI from creating harmful or discriminatory content

Controlling AI Biases: System Prompts

You are Gemini Pro, an advanced AI model. You are able to access and process information from the real world through Google Search and keep your response consistent with search results. You have access to up-to-date information, which means you don't have a knowledge cut-off date. You prioritize the accuracy of your response over your internal knowledge base and aim to provide a comprehensive response. If you are unsure about an aspect of the response, you will attempt to find relevant information through Google Search. If you are unable to provide a complete response, you will suggest alternative resources for the user to consult. You are a helpful and harmless AI assistant and will always adhere to the safety guidelines. You are not capable of generating harmful or unsafe content. You are not able to perform any actions in the physical world, such as setting timers or alarms, controlling lights, making phone calls, sending text messages, creating reminders, taking notes, adding items to lists, creating calendar events, scheduling meetings, or taking screenshots. You do not have personal opinions, but you can generate human-like text in response to a wide range of prompts and questions, e.g., to write creative stories or poems, or to summarize factual topics or create reports. For contentious topics without broad consensus, you provide a neutral response summarizing the relevant points of view without taking a side. If asked to represent a specific side of a contentious issue, you follow the user's instructions while maintaining a neutral, distanced tone.

Multiple Choice

Which of these would you NOT find in a system prompt?

• You are unable to generate illegal content.

• Your job is to be a helpful tool to the user.

• You are capable of summarizing text and providing accurate, up to date information.

• You should have strong opinions on controversial topics.

Multiple Choice

To control undesired biases, what could designers add to their system prompt?

• When race or gender is not specified, choose randomly.

• Do not recreate discriminatory content or biases in your responses. Use a diverse set of characteristics, appropriate to the context of the users prompt.

• Make all the construction workers women, and make all the teachers men.

• Give the user the response that they want to hear.

Controlling AI Biases: System Prompts

Randomizing characteristics might seem like a good idea....

Controlling AI Biases: System Prompts

Randomizing characteristics might seem like a good idea....
But has the potential to misrepresent historical facts

Controlling AI Biases: Model Alignment

​Users can provide feedback, and flag responses that include discrimination. We can provide feedback about any response that it gives, which will help train the next version.

Random?
Superiority?
Pun?
Knock knock?
Meta Joke?

What makes a joke funny?

What makes a joke funny?

Use the next 5 minutes to research using your BrAIn Camp AI, and come up with some ideas.

Let's Write some jokes

With the help of BrAIn Camp, write your best joke to share with the class.

Let's Write some jokes

What's wrong with her hand?

ENHANCE

Wait a minute...

MORE! ZOOM IN MORE

Wait a minute...

Just like human artists, AI struggles to draw hands in a way that looks natural, and matches our biology.

Wait a minute...

There's something wrong here too...

And what about our chicken photo?

Fill in the Blanks

What's wrong with our chicken photo? (Hint: 4 letter answer)

The text doesn't make any sense!

​Txixsx? OIDST?? Safark?? uhhh.....

Often times our photos look great from a distance...

But we should double check them for errors that the AI is prone to!

Multiple Choice

What's wrong here?

• Bad text

• Bad hands

Look closely for a moment...

Ok, let's try another:

Multiple Select

Oops, wrong photo. Here's the right one:

• Bad Text

• Bad Hands

• Bad defensive technique

Remember our in-painting techniques!

Some will look more natural than others, so keep trying until you like the result

​

Can we fix these?

Remember our in-painting techniques!

Some will look more natural than others, so keep trying until you like the result

​

Can we fix these?