If you haven’t seen or heard, Meta launched Meta AI using the Llama 3 platform. It is an AI assistant platform designed to assist and converse with humans and can be used through their website or via social media platforms such as Instagram, Facebook, Messenger, and WhatsApp.

Meta’s goal with the platform was to make peoples life easier by allowing users to access real-time information from the internet without having to switch between apps. The primary function is to understand and respond to human input by answering questions, providing information, generate text, offer suggestions, and create images. CEO of Meta Mark Zuckerberg has expressed that Meta AI is “the most intelligent AI assistant” that is free to use.

How does it work?

Meta AI has been integrated into the Meta ecosystem, making it easily accessible and convenient choice for many users. Meta AI can be used in feeds, chats, search bars and more. The goal of Meta AI was to make peoples life easier by allowing users to access real-time information from the internet without having to switch between apps. The primary function is to understand and respond to human input within seconds by answering questions, providing information, generate text, offer suggestions, and create images. The platform retrieves and analyses information using Google, providing a reference to allow the user to validate where the information came from.

You may be wondering, how are the models trained to be able to respond to users prompts? Meta AI uses unsupervised learning to train the model, which means they don’t need human-labelled data to learn. Instead, the model learns from large amounts of text from scientific articles, the internet and news articles. With this, there has been concerns raised on how the AI model will respond to topics on politics and social issues because of how its trained. Meta have responded stating they will use feedback to continue training the models, so they are less likely to produce potentially harmful responses and to improve their systems that automatically detect policy violations.

Exploring Meta AI

We tested the new Meta AI platform on Instagram and Messenger. We asked Meta AI questions based on Brisbane, such as what time will the sunset, and what activities are on in Brisbane. The platform delivered quick and accurate results, providing references for both prompts. Whilst it seems convenient and useful for answering quick prompts, such as what time the sun sets. A user is still required to switch apps to click on the link Meta AI provides if the user wants to find out more information. Whilst this is handy, it also defeats the purpose of not having to leave the app.

We thought we would also test the image generating tool to see how it performed. Meta used 1.1 billion publicly visible Facebook and Instagram images to train image generating AI model and is supposedly meant to be one of the most accurate images generating platforms. We found that generating an image took longer to deliver in comparison to generating answers but delivered an accurate image when asking simple prompts. When asking the image generator to create a user persona or an image of a person, it responded well. Although you can tell the image has been AI generated because of the rendering, the faces look relatively realistic. Additionally, the faces don’t look warped or cross-eyed, which was initially a common error for a lot of AI image generators.

After having successful results testing the AI image generating tool on basic prompts, we tested slightly more complex prompts. We asked the tool to /image Brisbane City River with people in boats and fish surrounding. Originally, the first image it generated looked relatively realistic apart from the fish. The fish looked out of scale in proportion to the boats and people, and it wasn’t likely for Brisbane to have koi fish within the river. For the second prompt, we asked to make the fish smaller, and to change the fish to something more likely to be within the Brisbane River.

The second response it gave looked even less realistic, most likely because of the shark’s head out of the water. It also didn’t really look like Brisbane City River anymore as the image was now mostly focused on the people. In a third prompt, I asked Meta AI to make the image look more like the Brisbane City River and well… I surely hope the Brisbane City River isn’t as shark infested as what it looked like below. The background of the city and river looks more realistic than the previous images, but the number of sharks make it look fake. It seems like the Meta AI tool is great at focusing on one prompt but when given several, it can’t focus on generating both at scale.

In comparison to previously tested image generating tools like DALL-E and Craiyon, Meta AI consistently delivered faster results. However, it remains evident that like many AI image generators, all struggle with more complex requests that require multiple prompts. While these tools excel at basic tasks like creating user personas or generating ideas, they are not yet suitable for replacements for stock photography.

Protecting your Data

Meta may have reasoned that integrating the AI tool into popular social media apps used daily by many people was a logical step. However, many users of social media are expressing that they feel like the Meta AI integration feels invasive because users currently don’t have a choice to opt out or remove the feature from their social media apps.

An additional concern with Meta AI being integrated into the Meta ecosystem is the confusion with a lack of end-to-end encryption. Meta AI’s messaging feature on Instagram is NOT encrypted. WhatsApp claims to have end-to-end encryption, but this is unverified. End-to-end encryption guarantees that only you and your intended recipient can access your messages. This ensures confidentiality even from the platform itself. However, the Meta AI and Instagram integration lacks this essential feature. An absence of end-to-end encryption means your data could potentially be exposed to third parties, such as law enforcement and the government. What does this mean for users? Think twice before deciding what to share with your Meta AI assistant.

Meta AI relies on large amount of data for training and improvement, often obtaining information from third parties. Whilst the information obtained from third parties is publicly available on the internet or license information, it is still important to consider protecting your data. While this data is often accessible through public internet sources or licensed channels, prioritising the protection of your own data is essential. Consider implementing the following two methods to safeguard your personal information.

1. Delete personal information that Meta uses

• Users on Facebook now have the option to delete some personal information that Meta uses to train their generative AI models through the Generative AI Data Subject Rights form. This allows users to request deletion or raise concerns about any personal information obtained by third parties to train Meta’s generative AI models. However, Meta has made it clear that they do not directly link personal data to any Meta account. Find the form below
• Generative AI Data Subject Rights Form.

2. Delete your data when using Meta AI

• Meta AI saves conversations to provide more personalised responses. To avoid Meta AI retaining and using the information that you share, consider deleting your data, you can still see a copy of the chat, but the platform will delete the data.
• This can be done by typing /reset-ai

Overall

Whilst the new Meta AI messaging can bring convenience for many users, it’s crucial to remain vigilant about the data we share and be aware of the potential concerns regarding user privacy. Stay tuned for part 2 of this series, where we will compare ChatGPT to Meta AI