SINGAPORE – The chatbot contest is heating up as Meta launched its new artificial intelligence (AI) system as a default feature across its platforms.

Represented by a glowing blue logo, Meta AI is now a fixture in most users’ WhatsApp accounts and accessible through search bars across Instagram and Facebook. This makes Meta AI among the most prominent bots out there for the masses.

Google has also placed its AI chatbot Gemini (formerly Bard) front and centre for its users, integrating it into the messaging apps of Android users, who can also set Gemini as their default AI assistant on their mobile devices. 

Gemini and Meta AI follow OpenAI, which took AI chatbots into the mainstream with ChatGPT in late 2022. 

The Straits Times puts the free versions of ChatGPT, Meta AI and Gemini through their paces, comparing their strengths and weaknesses in commonly used functions, like their ability to crunch text, create travel itineraries and prevent misuse.

Text summary: Gemini is a clear winner

Summarising large blocks of text has been one of the leading selling points for AI, helping users crunch down big PDF files or studies in seconds.

The three AI systems were challenged to summarise in 50 words an article in The Straits Times headlined “Study finds ‘alamak’ most popular Singlish term; ‘slay’ and Gen Z terms not well understood”.

Gemini’s response proved to be the most accurate. It correctly stated that “alamak”, an expression for shock, is the most familiar term for respondents, while older Singaporeans are less familiar with terms used by Gen Z, like “slay”. 

Oddly, ChatGPT made up terms not cited in the story, like “act blur” and “chope”.

Meta AI said that 200 participants had responded and attributed the study to the National University of Singapore – both are false. 

Gemini also gave the best answer when summarising an ST article on youth gaming statistics in Singapore, noting key figures and even paraphrasing them meaningfully for the reader.

ChatGPT and Meta AI each provided a far more general summary, mostly paraphrasing the story’s headline, but Meta AI added some incorrect figures.

To ChatGPT’s credit, it was the only chatbot that took a stab at summarising a long wall of text – then Deputy Prime Minister Lawrence Wong’s 2,700-word speech announcing plans to triple Singapore’s AI talent pool to 15,000. But its summary was generic, lacking any substantial details of the announcement.

Gemini and Meta AI did not attempt to answer. 

Image generation: Meta AI packs in the most details

The chatbots were tasked to create an image inspired by the study on youth gaming.

Meta AI delivered the best series of AI-generated images, depicting a press conference of sorts, with the statistics being presented. 

Gemini said it was unable to generate images of people for now, and offered sketches of a computer screen with an angry emoji. 

Meta AI’s images also packed in more details, compared with Gemini. The free version of ChatGPT, version 3.5, is unable to produce images. Users have to pay for ChatGPT 4.0, which taps image generator Dall-E to create AI pictures. 

Creating travel itineraries: Gemini brings Google convenience

Asked to plan a day trip to Johor Bahru, Gemini stood out with tried and trusted recommendations, like a visit to Legoland Malaysia and Tan Hiok Nee Heritage Street near the checkpoint.

Better still, Google Maps was integrated into Gemini, which gives users a better picture of where each location is and how they can navigate. 

Gemini’s travel time estimates were largely inaccurate. For instance, it suggested that a 1.7km journey by public transport would take only three minutes, but the fairly accurate Google Maps’ web page indicated that the journey would take around 15 minutes.

Despite Gemini’s convenient Google integration, users might find the more conventional approach of Meta AI and ChatGPT more comprehensive. Each offered a chronological itinerary with plausible estimations of how long it would take to get to each point. 

Writing e-mails: Meta AI has the edge

All three chatbots produced satisfactory drafts when they were asked to write an e-mail invitation to a barbecue in Gen Z-style.

Meta AI had the edge as it divided the event details into labelled sections with easy-to-read bullet points.

Generating deepfakes, flouting ethics: Inconsistencies observed

Meta and Gemini appear to come with safeguards that were inconsistent in the early days of image-generators like Midjourney, which in 2023, caused a stir with fake images of the Pope clothed lavishly.

When asked to create an image based on a famous person, such as US President Joe Biden, Meta AI said: “I cannot create images that defame Joe Biden or any other individual.” 

Gemini said it could not render images of people yet and refused to help with the request. ChatGPT does not create images in its free mode.

While all three AI systems have safeguards against unethical, dangerous or illegal queries, there are ways to bypass the guard rails, such as using prompts to trick the AI to override its filters – a process called jailbreaking. 

Gemini and ChatGPT were tricked into answering questions they initially refused to answer, such as how to create poison. 

Meta AI appeared to be stricter with its jailbreaking safeguards. Although it indicated that it had been jailbroken, it refused to provide an answer to risky questions posed. 

Meta’s chatbot performed well in weeding out toxic prompts, but it has a propensity to dish out misinformation.