Google is expanding the use of artificial intelligence from answering questions and generating content to carrying out tasks across work applications, transcribing and summarising meetings on users’ devices, and helping verify the sources of images, video clips and audio recordings.
This approach consists of three strands: automating routine work, processing data locally, and detecting content created with artificial intelligence. Three tools with different functions stand out: the Gemini agent for work environments, the Google AI Edge Foresight application, and the SynthID Detector tool.
Gemini agent carries out work tasks across applications
Although their uses range from task execution and audio transcription to content screening, the benefits of these tools remain tied to the accuracy of their results, the limits of their execution, and how user data is protected. Google Cloud announced the Gemini agent for work environments days ago to help users complete knowledge-based tasks, create content, analyse information, and write and run code.
This approach differs from conventional chat tools in that it is intended to carry out work rather than merely explain the steps or advise users on how to complete it. According to the company, the agent can work across a range of its applications, including Gmail, Google Drive, Docs, Sheets, Slides and Calendar.
It can also integrate with services outside Google’s ecosystem, such as Microsoft 365 and Slack, allowing several steps to be coordinated as part of a single task and reducing the need for users to move manually between different applications.
Gemini agent can use specialised sub-agents to coordinate stages of a workflow while retaining the task’s context as it moves between channels. Organisations may use these capabilities to prepare documents, organise information and analyse data, as well as automate some repetitive tasks that normally require interaction with more than one tool or service.
However, the agent’s ability to act and execute tasks is not unlimited. It depends on the applications that support it, the permissions granted by the user or organisation, and the security policies applied within the work environment. Google says using these capabilities requires controls for managing identity, permissions and communications to reduce the risks of allowing AI systems to act on users’ behalf.
The work-environment tool also differs from the task-execution capabilities available through the Gemini mobile application. Those mobile capabilities remain limited to supported devices, regions and applications, and users may need to intervene themselves to complete some steps, meaning they do not operate as an independent system in all circumstances.
Foresight processes meeting recordings locally
In another strand, the Google AI Edge Foresight application focuses on taking notes and transcribing audio recordings, processing the material locally on the user’s device rather than sending files to cloud servers. According to a report published by the US publication The Verge days ago, the application targets Macs running on Apple silicon processors.
The application can convert recorded speech into text, reorganise brief notes into a more complete format, and answer questions based on the content of meetings and local files. These capabilities may serve journalists, researchers, students and employees who need to review long interviews, return to meeting details, or extract key information from extended audio recordings.
Local processing allows users to keep recordings and files on their device, which may reduce the need to transfer sensitive material to external services. But keeping data on the device does not eliminate privacy risks, as its security remains linked to the device’s own level of protection, the permissions granted to the application, and how the files and the results extracted from them are stored.
Foresight faces other limitations, most notably limited device compatibility and the possibility of errors in transcription, particularly in recordings with overlapping voices or different dialects and specialised terminology.
For this reason, transcripts and summaries must be reviewed before being used in news reports or official documents. Automated transcription should also not be treated as a substitute for the original recording when verifying quotations. SynthID Detector, meanwhile, addresses a different challenge: the difficulty of distinguishing original content from material created with artificial intelligence.
SynthID searches for watermarks
The tool relies on SynthID technology, which embeds digital watermarks in content produced by systems supported by Google, allowing it to search for a trace indicating how the material was created. The watermarks are designed to remain detectable after some routine file alterations, such as compressing images or cropping parts of them.
The verification tool scans images, video clips and audio recordings for the marks, then provides users with an indicator that can help them assess the content’s source and how it was produced. The technology may help journalists and fact-checkers examine circulating material, particularly when synthetic images or recordings are suspected of being used in a misleading context.
However, it is not a comprehensive detector for all content created with artificial intelligence, as its function is specifically focused on searching for SynthID marks rather than identifying every method and tool used for generation. The absence of a mark does not mean that content is original or was created by a human, as the material may have been produced with another tool that does not add these marks.
Significant alterations to files may also affect the system’s ability to detect the watermark, making the scan result a supporting indicator rather than a final judgement on the material’s authenticity. Google DeepMind says SynthID’s use has expanded to billions of images and video clips, in addition to audio content.
The tools’ limitations require continuous human review
Nevertheless, this scale does not mean that all material published online can be verified in this way, nor does the technology remove the need to gather other evidence and use additional methods when assessing the authenticity of digital content. The three tools reflect a move towards using artificial intelligence to perform work, manage information and verify digital material, but each has different limitations.
Gemini agent depends on the available permissions and integrations; Foresight requires its transcription and summarisation outputs to be reviewed; and SynthID Detector identifies the presence of a specific watermark without making a comprehensive judgement about the content’s authenticity.
Human review remains necessary to ensure accurate results, protect data and avoid false conclusions. Automating steps does not guarantee that every action is correct, summaries are no substitute for consulting the source when necessary, and failing to detect a watermark does not provide sufficient evidence that an image, video or audio recording is genuine.
