The new feature is optional and is not enabled by default. It is currently available for some Mac users depending on account type and region.
Once enabled, ChatGPT can use input events while using the device, such as mouse clicks, typing, keyboard shortcuts, and switching between apps.
Instead of these activities remaining as separate, unrecoverable data points, the system converts them into memory and a timeline that can be used in later conversations.
For example, a user can ask ChatGPT about activities they performed during a specific time period or ask it to help complete work they previously started, without needing to explain all the details again.
ChatGPT does not take screenshots or record audio
Although descriptions of this feature might give the impression that ChatGPT “monitors” the user’s screen, technically its mechanism is different.
According to OpenAI’s documentation, the Computer History feature does not rely on taking screenshots or recording video of the display’s contents. The feature also does not record the microphone or system audio and does not include private browsing activity.
This distinguishes the new ChatGPT feature from Microsoft’s Recall feature, which relies on taking periodic screenshots to help users remember activities performed on the device.
However, the absence of screenshots does not mean only a small amount of information is collected. Data such as keystrokes, clicks, and app switching can reveal significant information about the nature of a user’s activities and their digital habits.
Where is the security problem?
OpenAI has stated that event files are temporarily stored on the user’s device and will be deleted after a maximum of 48 hours. During this time, the data is used to process the user’s activities and convert them into “memories” that ChatGPT can use.
The company also says it does not retain event files after processing them, unless legally required to do so. It also says these data are not used to train AI models.
Nevertheless, some of the memory files created on the Mac are not encrypted; an issue that has raised concerns among cybersecurity experts. According to these experts, malware present on the device could attempt to access this information.
In such circumstances, the data collected by this system could become an additional target for attackers if the computer is hacked.
What does ChatGPT know about the user’s activity?
Depending on the type of data this feature records, ChatGPT can gain a more detailed picture of how a person uses their device; including which apps the user switches between, which websites they use, the documents they work on, and the tasks they perform.
This information can be very useful for increasing productivity; because the AI assistant, instead of waiting for the user to explain every step, can better understand the context of their activity.
Conversely, a report of the user’s digital activities also becomes part of the data that must be protected.
According to the British newspaper Metro, security experts believe the biggest risk is not necessarily hacking ChatGPT directly, but malware accessing the user’s device and abusing residual information from this feature.
If attackers gain access to this data, it could be used in phishing and social engineering attacks, credential and identity theft, and even corporate espionage.
The user still has control over this feature
To mitigate these risks, OpenAI has provided tools that give users more control over how this feature operates.
A user can specify which apps and websites are allowed to participate in activity reporting, stop the collection of activities, view stored items, and delete data they do not want retained.
The feature also only starts working after the user enables it; therefore the decision to use it is up to the user.
As a result, people who do not want ChatGPT to have broader access to their daily device activities can disable the feature or exclude specific sources from the activity recording process.
From “assistant” to “contextual task overseer”
The Computer History feature is a sign of a broader trend in the evolution of AI assistants; a shift from assistants that wait for a user’s question toward assistants that try to understand what the user is doing.
In such a scenario, instead of the user telling ChatGPT “where is the file I was working on?”, the assistant can, relying on activities performed on the device, attempt to find the file or understand the intended task.
The paradox is that the more information the AI has about how the user works, the more precise help it can provide; but conversely, the price may be giving it greater visibility into the user’s digital life.
Therefore, the main issue is not only how useful this feature is, but how willing the user is to allow the AI broader access to the activities they perform daily with their device.