Meta AI Review: A Convenient but Unimpressive Virtual Assistant

6.0/ 10
SCORE

Meta AI

Pros

• Integrated into Instagram, Facebook, Messenger and WhatsApp
• Connected to Google and Bing
• Text and image generator

Cons

• Frequent hallucinations
• Often requires clarifying or follow-up prompts

Basic info:

• Price: Free
• Availability: Web, Meta platforms
• Features: Connected to Google and Bing, multilingual text, image and code generation
• Image generation: Yes

If we had Meta AI before ChatGPT, I might have been more impressed by Meta's virtual assistant. Or if we didn't already have Microsoft Copilot, which is also connected to the open internet to use real-time information in its responses. Or Perplexity, with its Reddit database. Even Meta's defining differentiator, its Imagine image generator, isn't mind-blowing compared to Dall-E 3 and Adobe Firefly. It's better than "hallucination station" Google Gemini, but it's certainly not better than CNET's current reigning champ, Anthropic's Claude -- and Claude isn't even connected to the open internet.

The beta version of Meta AI was announced in September 2023, after it was developed into a full-blown conversational assistant from the "large language model Meta AI" AKA Llama. I don't think there's anything wrong with Meta being a bit late to the party, especially if they put the extra time to good use. But unfortunately, Meta AI's flaws were all the more apparent to me because it had to stack up against so many competitors. 

That said, I do believe there is a place in our digital lives for Meta AI, for one simple reason: It's already where we are. On Instagram, Facebook, Messenger, WhatsApp and even in Meta's Ray-Ban smart glasses -- Meta has integrated the chatbot widely across its products and platforms. And Meta plans to continue building on this foundation.

That's good news for users like you and me because while Meta AI is convenient, it isn't perfect. Here's how well the newest model of Meta AI powered by Llama 3 fared as a shopping aide, recipe creator, research tool and image generator.

How CNET tests AI chatbots

CNET takes a practical approach to reviewing AI chatbots. Our goal is to determine how good it is relative to the competition and which purposes it serves best. To do that, we give the AI prompts based on real-world use cases, such as finding and modifying recipes, researching travel or writing emails. We score the chatbots on a 10-point scale that considers factors such as accuracy, creativity of responses, number of hallucinations and response speed. See how we test AI for more.

Meta writes in its AI terms of service that it may "retain and use" any information you submit to its generative AI chats, like in prompts. It specifically tells users "Do not share information that you don't want the AIs to retain and use." It also says Meta can share your information with third parties "who help us provide you with more relevant or useful responses," such as third-party search engines, in accordance with those company's privacy policies. You can delete your previous information from Meta AI's history by entering the prompt /reset-ai, which Meta AI told me deletes the AI's copy of your conversation from Meta's servers. You'll still be able to see the chats from your side.

Generating images with Imagine

Using Meta AI's "Imagine" function to create AI-generated images is cool, but the results aren't always picture-perfect. Meta spit out the images very quickly, in less than a minute. But the real fun is that while you're writing your first prompt, Meta will generate images and update them as you type. 

This iterative process is great for helping you see what Meta has in mind so you can adjust as needed. For example, the video below shows the process I went through while writing a prompt, and I was able to steer the AI toward a specific kind of look. I wanted an image that was more photorealistic than the cartoon style Meta was leaning toward, and the live preview saved me from having to type a follow-up message.

Once you've finished your prompt and hit enter, you get four images, all with a small watermark in the lower-left corner that says "Imagined with AI." Many of the images I generated looked all right from afar, but zooming in revealed some telltale quirks that many AI image generators face. I generated a batch of images of dogs chasing frisbees in which none of the dogs looked real, for example. These problems were more apparent in photorealistic images, but more often than not, I was pleased with the result.

If you don't like what Meta generates, you can still send follow-up messages to refine your images. As a bonus, you can animate images into short looped videos like GIFs simply by clicking the Animate button below each batch of images.

Shopping

When you're deciding what to buy, like a new smartphone or TV, you might feel overwhelmed by the amount of information to consider. Sorting through official reviews, spec sheets, multiple retailers and even your friends' or random TikTok recommendations can be a huge time commitment. That's where AI can come in handy and help you narrow down your search.

I've been casually browsing for a new phone, so I wanted to see what Meta AI would recommend. I told Meta that I currently have an iPhone 11 and that I want a new phone with excellent camera quality and storage without paying an exorbitant price. Meta produced a list of the best smartphones for 2024 -- which, unsurprisingly, was just the latest models from major manufacturers, including the iPhone 15 Pro Max and Samsung Galaxy S24 Ultra.

Call me old-fashioned (or cheap), but I don't automatically assume newest means best. So I followed up with many questions: What's the best phone for camera quality? What iPhone would be a significant upgrade from the 11? What do I need to know if I want to jump ship from Apple to Android? Will Apple ever make a foldable iPhone? Meta handled all these questions well, citing reputable sources and pointing out things I hadn't considered before.

Meta is a great shopping comparison tool, pulling specs from different models and breaking down jargon when asked. It also helped me understand the key differences between my current phone and the options I was considering. And I'm not sure why, but there were fewer hallucinations and sourcing quirks while I tested shopping queries than other types.

Meta performed similarly to Perplexity and Gemini. Claude and Copilot in creative mode were good AI shopping assistants. I would caution against using chatbots without recent information, like ChatGPT 3.5, as you'll likely miss out on new reviews and products.

Recipes

If you're like me and "jump to recipe" is your favorite feature of food blogs, using an AI for recipes might be an unexpected perk for you. 

Like other chatbots, Meta cuts through all the BS on recipes and food blogs and just delivers the ingredients and instructions in an easy-to-digest format (pun intended). I asked Meta to find a recipe for one of my favorite meals recently, Gigi Hadid's viral spicy vodka pasta. Meta nailed it, pulling up the exact recipe I had seen on social media. I did tell Meta to generate more general recipes, like pasta salads with a lot of variety and options, and it still produced a reasonable recipe. Meta performed similarly to competitors like Perplexity, Copilot and ChatGPT 3.5, but Google Gemini excelled at this task, pulling in more robust ingredient lists with more authentic ingredients.

One bonus tip: You can ask Meta to add each ingredient's measurement to the instructions so you don't have to scroll back and forth between ingredients and instructions -- a major time-saver and something I'm going to do with my recipes going forward.

Research and accuracy

Anyone who has done a Google search knows that two of the most important elements of research are the timeliness of the information and the credibility of the source -- in other words, is this information the latest, and is it true? Meta's real-time integration with Google and Bing and the ability to cite its sources should theoretically make Meta AI an excellent research tool. In practice, it falls short.

I asked Meta to find me good academic journal papers about the relationship between gender roles and power structures in Native American communities, specifically peer-reviewed and recent ones. In return, Meta gave me a list of five sources, including an elementary school lesson plan, papers from 1989, 2000 and 2002 and a law journal article without a date that a quick Google search told me was probably from 2006. Only a couple were by Native scholars, which would be essential authorities and perspectives on this topic. A Google search with the same search terms brought up much more recent sources.

When I asked Meta to summarize those sources, it produced a thorough, nuanced result. But because I know the sources it used to generate the information weren't exceptional, I have to take the notes with a grain of salt. When I asked Meta to create a list of Native scholars, it did give me a list of real people who have published some type of work, but their expertise wasn't always in the area I asked for.

Meta AI might be a good place to start research -- coming up with key search terms, finding names of scholars in particular fields and getting acquainted with a topic more broadly. Still, hallucinations and misunderstandings mean you can't entirely rely on it. It's like a worse, AI version of Wikipedia -- good to get you started, but you'll need to double-check its sources carefully. 

Claude and Copilot were the best-performing research tools, producing usable results with linked sources. Perplexity didn't hallucinate as much as Meta did, but it had trouble synthesizing information. ChatGPT 3.5 (not internet-connected) and Gemini (internet-connected) shared Meta's struggle with hallucinations.

Summarizing

I asked Meta to summarize an article I wrote earlier this year explaining the concept of the fediverse. I copied and pasted the article URL into my prompt and Meta was able to summarize it. The initial summary Meta produced was fine, though bare-bones. When I asked it to expand its explanation, I got a few more key details, but it still omitted a lot of the necessary context I had in my original article.

Meta can analyze a good chunk of text, so I copied and pasted the entire text of the article into Meta AI and asked it to summarize, which gave me a much better response. For example, in the first summary, Meta told me that the fediverse is a collection of social media platforms that can communicate with each other, which is true. But when it analyzed the whole article text, it noted that it's a system of decentralized social platforms, which is a crucial detail. It also included the instructions I wrote for adjusting your Threads account settings to enable fediverse sharing, a big part of the story that was missing in the initial summary. 

Meta summarized as well as Gemini did, both producing good summaries when given the whole article's text and decent summaries with just the URL link. ChatGPT 3.5 had some issues and has a character limit that will make it hard for you to use it to summarize long documents. Claude and Perplexity failed to produce sufficient summaries.

Traveling

Some people may enjoy trip planning, sorting through reviews, guidebooks, TikToks, and all the travel sites available for every major vacation destination. I am not one of those people, so I was excited to use AI as a virtual travel agent. To put Meta AI to the test, I had it create several itineraries and plans for a trip in a touristy, though not too popular, locale: Banff, Alberta, Canada.

Because Meta AI uses real-time information from Google and Bing, I had higher expectations for its results compared to chatbots without internet access. And boy, was I disappointed when Meta recommended locations that had permanently closed, made up names of several restaurants and considered everything on the same side of a mountain "within walking distance." And for the options that did exist, Meta recommended the same places multiple times throughout the trip, which wasn't quite the exploratory trip I had in mind.

Safe to say, I won't be using the itinerary Meta created for my upcoming trip. What Meta was better at was helping me find specific restaurants, hotels and hiking trails. Even then, there were still issues and hallucinations, but I found that asking Meta to narrow down the best choices for each activity helped avoid some problems. Copilot worked best as a travel planning assistant, followed by Claude and ChatGPT 3.5. Gemini shared Meta's penchant for making stuff up.

It's also worth pointing out that I don't feel like Meta fully weighed the safety of the activities and travel methods it recommended. Meta would add notes at the bottom of its plans encouraging me to check forecasts and local transportation schedules and to plan or drink responsibly, but it was a legitimate afterthought. When I'm traveling, safety is a top priority for me, especially if I'm alone. Combined with Meta's other hallucinations and flukes, I wouldn't rely on any Meta AI-generated travel plan.

Writing emails

Meta's email writing ability is average. My only concern with using Meta AI for writing emails is that its baseline tone is very monotonous -- robotic, if you will -- and doesn't mimic human tone without being asked to do so. Meta performed reasonably well when I asked it to improve emails I had already written. I had it generate a basic email asking for a meeting, and when I asked it to adjust the tone -- more professional, more friendly, etc. -- it responded accordingly. 

I also gave the chatbot the more difficult task of composing a pitch email about the ethical considerations of AI use in university settings. Its pitch was pretty good, taking into consideration the different uses it may have and the potential ethical ramifications. It wasn't the most original sounding -- it felt like Meta was plugging relevant topics into a template -- but it was a thorough brief. If I were to use it, I would edit it some and put my own spin to make it less monotonous.

As the most conversational chatbot, Claude excelled at this task. Meta performed alongside competitors like Google Gemini, ChatGPT 3.5 and Perplexity. Microsoft Copilot refused to write a similar pitch for CNET's Imad Khan in his review, saying the subject matter was too sensitive. While Meta is on par with other chatbot services, I noticed that Grammarly's generative AI produced more human-sounding emails.

Meta AI meets you where you are

The biggest thing Meta AI has going for it is that it's meeting you where you are. For better or worse, Meta owns and operates some of the biggest social media platforms in the world, and by integrating its AI into those platforms, Meta has catapulted itself above the competition and created one of the most accessible AI products.

But just because you can easily access Meta AI doesn't mean it's worth using. Meta struggled with hallucinations and other false or misleading answers -- and for a chatbot with an internet connection, that was pretty disappointing. Meta excelled as a shopping and recipe assistant, and the Imagine image generator was undoubtedly a bright spot. But its travel plans were a disaster, and its performance as study and research aide was only standard at best.

So, is Meta AI better than ChatGPT 3.5, Copilot, Perplexity, Gemini and Claude? It definitely outperformed Gemini, and ChatGPT's lack of recent data also moves Meta AI up in my rankings. Claude, Perplexity and Copilot are all worthy competitors, and for certain tasks, they are likely the better choice. But I can't deny that if I'm scrolling through my Instagram feed and need a quick question answered, I'm going to DM Meta AI before doing a Google search or opening another chatbot. Convenience isn't everything, but it sure is nice.

CNET is using an AI engine to help create a handful of stories. Reviews of AI products like this, just like CNET's other hands-on reviews, are written by our human team of in-house experts. For more, see CNET's AI policy and how we test AI.