Chat apps are gradually becoming gateways for asking questions, searching for information and completing tasks, as AI tools appear within conversations in the form of integrated assistants or bots that can be called when needed.
The shift allows users to summarise a discussion, draft a message, request an answer or complete a task without moving to another app or opening a separate search engine. Bringing smart assistants into messaging apps is changing the nature of competition between platforms, but it is also putting data handling and permissions at the forefront.
Meta unifies its assistant across social apps
The more an assistant relies on the context of a conversation and the information available within it, and the greater its ability to take actions, the more important it becomes to know which data it processes and what limits govern what it can do on the user's behalf. Meta is adopting an approach centred on providing a unified assistant across WhatsApp, Facebook, Instagram and Messenger, as well as on the web and through a standalone app.
Meta AI can answer questions, search the web and create images, as well as help with shopping, coding and a range of everyday tasks. How the assistant works varies by platform. On WhatsApp, users can chat with it directly or call it into group conversations to obtain recommendations or request explanations of messages.
On Instagram, it is linked to messaging, search and visual content, while it operates within Messenger conversations and helps Facebook users find posts and communities. In some cases, Meta AI uses public posts published on Instagram, Facebook and Threads to formulate answers connected to topics users are discussing.
Meta is seeking to expand its capabilities beyond direct answers to include planning, document writing, presentation creation and structured research. On 8 April, Meta announced its new closed model, Muse Spark, saying it powers the assistant after it was rebuilt and integrated into it.
The company had announced an optional feature called Private Processing on 29 April 2025, designed to carry out certain tasks, including summarising WhatsApp messages and offering writing suggestions in a confidential environment whose content, it said, would not be accessible to it. On 13 May, Meta also unveiled a hidden chat feature for Meta AI conversations through WhatsApp and the standalone assistant app.
The feature relies on the Private Processing system, and the company said that messages used in it are not stored after the session ends. Meta believes that expanding its assistant's presence across its platforms will help keep users and commercial interactions within its services.
It said that more than one million businesses use its business-specific agents each week through WhatsApp and Messenger to communicate with customers and complete sales, strengthening the position of chat apps as commercial channels rather than merely means of exchanging messages.
Telegram opens chats to a bot marketplace
Telegram is taking a different route centred on bots and relatively open application programming interfaces, helping create an environment that includes both official and unofficial AI tools. On 7 May, the app announced a feature allowing AI bots to be called into private and group chats simply by mentioning their names, even when those bots are not members of the conversation.
Telegram's platform contains millions of bots, and some of these tools and integrations offer capabilities including answering questions, checking information, creating text and images, completing tasks and automating inboxes. Some tools can also be linked to users' accounts within the limits of the permissions granted to them.
Telegram also added a feature for generating smart summaries of long posts, based on open-source models through the decentralised Cocoon network. The platform offers Microsoft's official Copilot bot, while OpenAI does not provide an official ChatGPT bot on Telegram, and Anthropic does not offer a ready-made public bot for its Claude assistant, despite making a Telegram integration available through Claude Code.
Users can nevertheless access the ChatGPT and Claude models through bots operated by third parties, including projects that allow users to switch between multiple models within the same conversation. Telegram therefore offers something resembling a miniature marketplace from which users can choose an assistant according to the task, model or privacy considerations, although some options operate through intermediaries.
Chat competes with traditional search engines
The expansion comes as chat begins to compete with the traditional way of searching for information. For years, the search engine was the starting point where users typed their questions and received a list of results, before moving between links and comparing pages to find the answer they were seeking.
The smart assistant moves a large part of this process into the chat window, gathering information, summarising it and rephrasing it in a single answer that may draw on sources or training data. The quality of these answers varies according to the model used and the nature of the task requested.
A ready-made answer saves users time, but it can make the route to the source less clear, particularly if users treat it as a final result and do not review the original material or compare different points of view. Competition between smart assistants and search engines is therefore measured not only by the speed of the answer, but also by the clarity of sources and the ability to verify what is presented.
In a survey of 5119 people, the Pew Research Centre found that 42% of chatbot users rely on them to search for information. The centre also found that users of Google search pages displaying an AI-generated summary were less likely to click on links and visit the original source.
Another study published by the Nuremberg Institute for Market Decisions, involving 1526 people, found that ChatGPT users completed tasks faster and were more likely to reach the correct solution in practical search tasks than those who used Google Search. However, assistant users relied less on original sources published on the web.
The chat interface allows users to simplify terminology, summarise documents and ask follow-up questions without explaining the subject again from the beginning. However, the confident and orderly way in which an assistant presents its answer can conceal errors, missing information or differences between the sources covering the subject.
Do model hallucinations narrow the range of available information? As the text states: false or fabricated information appears in a convincing form.
AI hallucination refers to the production of false or fabricated information presented in a convincing form. These errors may take the form of incorrect information, a non-existent quotation, a misleading interpretation, or a source and link that do not reflect what the assistant attributes to them. One study tested 27 language models across 155 topics linked to 12 countries, using 200 different question formulations.
It concluded that most models provided less diversity in claims and information than the baseline of web search. The finding suggests that users may reach a narrower range of information and perspectives when they rely on a single generated answer instead of reviewing and comparing multiple sources.
Errors cannot always be detected from the style of the wording: an assistant may present a fabricated detail or outdated information, while the omission becomes clear only when the source is opened or the answer is checked against an independent reference. When using assistants for medical, legal or financial matters, it is important to consult the original source and compare the answer with independent evidence, as well as request verifiable links and ask the assistant how certain it is.
The tool should not be authorised to make a decision or carry out a sensitive action based on a single summary. At the same time, having an assistant within a chat app can broaden access to knowledge, particularly for people who struggle to use search engines or read long texts.
These tools can simplify scientific and technical concepts, translate messages and documents, summarise meetings and articles, and explain images and diagrams. Voice features and the way the service is designed may help some people with disabilities use these tools and access information. But the practical benefit remains tied to the availability of suitable features and how they are integrated into the platform.
Privacy becomes more sensitive when AI tools enter personal conversations. A message sent by a user to a bot operated by a third party may become available to the bot operator, depending on how the service works and the terms governing its use. It cannot be assumed that all bots provide the same level of protection users expect from the underlying chat app.
Telegram says hosted AI bots can see only the message they are mentioned in and the replies connected to it, but this does not remove the need to know who operates the bot and how it handles data. In Meta's case, the company has said it uses users' interactions with AI features among the signals that help it personalise content and advertising across its platforms.
The impact of interacting with an assistant therefore extends beyond the answer it provides and may also become part of the broader commercial relationship between the platform and the user. An assistant embedded in a conversation has advantages that a traditional search box may struggle to match: it is present from the moment a question is asked and may benefit from the context of the conversation, interests and relationships.
However, awareness of context and speed of response do not guarantee an accurate result or remove the need to verify it.
The success of these assistants depends on more than their ability to produce a quick answer. It also involves distinguishing situations they can answer from those requiring them to ask the user for more clarification, refer the user to an independent source or decline to carry out a task for which sufficient safeguards are not available.
