If you’ve updated your mobile apps since the weekend you may have noticed a shiny new feature on your Facebook, Messenger, Instagram and WhatsApp apps, courtesy of parent company Meta.
On April 18 the tech leviathan began rolling out the integration of Meta AI – an artificial intelligence (AI)-powered, English language chatbot – into these apps in a dozen countries, including SA, Canada, Australia, Nigeria and New Zealand. Meta AI was available in the US last year, and can also be accessed via the website, meta.ai.
Look for the blue-ish circle icon at the top of your app interface or in the search bar. Tapping these opens up a prompt, with suggestions to ask Meta AI for things like personalised suggestions (workout plans, recipes using what you have on hand), simplifying concepts (like quantum physics), planning (a low-cost trip, a family game night), and so on.
You can also use the chat to generate images using plain words to describe what you want. The photo accompanying this article was generated by Meta AI, using the prompt “imagine a person using meta ai on their phone” – although I must confess it took two further refinement requests to get an image that didn’t feature a person wearing a headset reminiscent of Meta’s own.
The images are pretty good, judged against the bar of other generative AI image tools, but with that particular glossiness and colour depth that should be a red flag for anyone paying attention. Look a little closer at the hands or the screen of the phone, and the illusion falls apart.
The chatbot uses Llama 3, a large language model trained using public information including Facebook and Instagram posts, with some effort to exclude private content such as posts shared with family and friends only. Speaking to Reuters in September 2023, Meta president of global affairs Nick Clegg said private chats were excluded in an effort to respect consumer privacy. Such exertions must always be applauded, while acknowledging that Meta’s definition of respecting privacy may not meet your own tolerance for such. The group’s history is littered with privacy failures, after all.
In addition, you can’t really opt out. Meta confirmed to Fast Company that there is no way to remove the integration from your search bar, “but you can search how you normally would engage with a variety of results” [sic]. Practically, that looks like searching for a Whatsapp chat about getting a flu shot and getting Meta AI prompts – including “is it effective this year?” – above chats in which those words occur. In other words, instead of quickly navigating to the chat with a family member about her annual flu vaccine, I am prompted to question the validity thereof.
For the record, the chatbot goes on to suggest getting vaccinated anyway, but built as it is on user posts and scrapped data, you can see the potential for the insidious influence of misinformation therein. Meta has, however, attempted to limit some hot-button topics. In a comically awkward pre-emptive step, typing in “trump” got me a handful of prompts about buying trumpets.
As with all of the gen AI chatbots currently available, Meta AI can and does hallucinate. When I asked Meta AI how to disable it, it provided a neat list of plausible but entirely inaccurate instructions. This isn’t just my experience if the number of Reddit chats on the same topic is anything to go by.
Next, I asked Meta AI if it could change the background of an image I upload to it. Cheerily, it replied, “yes, I can … using my image editing capabilities… and I’ll be happy to assist”. How do I upload?” I persisted, which prompted this u-turn: “I apologize [sic] for the confusion. I currently do not have the ability to directly receive images”.
CNBC says users have reported instances of the virtual assistant joining a private parent group on Facebook and claiming to have a “gifted and disabled child”, and joining a “buy nothing forum” where it attempted to give away “nonexistent items”. That’s a fail on the metrics of reality and privacy, so I’d say if your Facebook feed isn’t already full of users mistaking AI content for real life, expect it to be.
So, this is a good opportunity to reiterate that no-one should be treating the outputs of gen AI as authoritative; double-checking and multiple sources are your friends. Honestly, that should be the rule for information gathering in general. Meta AI is clearly a very capable tool with remarkable natural language abilities, but while a chatbot can tell you what to cook with pork and bananas, it still takes human intelligence to say “don’t”.
This is just one of several big bold moves from Meta this week, including a decision to collaborate with our technology companies on its virtual reality offerings, and big plans for Meta-designed chips for AI processing.
TechCrunch say the integration of AI into these popular apps “underscores Meta’s efforts to stake out a position as a mover and shaker amid the current hype for generative AI tools among consumers”, but to my mind it is also a way to compete with AI specialists such as Open AI, ringfencing user time and attention by bringing the trendy and fun generative AI functionality right into the apps we already use.
And, of course, Meta is already training Llama 3’s successor, with our help. Accessing the tool involves agreeing to terms including that Meta receives the prompts we feed into the tool while using them – as most of the gen AI options do. Unlike most though, Meta has a touchpoint into the lives of something like half the population of the planet. That’s one helluva training corps.
• Thompson Davy, a freelance journalist, is an impactAFRICA fellow and WanaData member.
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.