OpenAI adds screen‑capture “Chronicle” to Codex for Mac, sparking privacy debate

OpenAI unveiled Chronicle as part of a major update to its Codex for Mac app, turning the coding‑focused assistant into a broader AI workspace. The feature runs background agents that capture screenshots of the user’s display at regular intervals. Those images travel to OpenAI’s cloud, where optical‑character‑recognition and visual analysis convert them into plain‑text summaries. The resulting notes are stored locally in a hidden directory as Markdown files, allowing Codex to draw on recent activity when answering prompts.

Chronicle is a research preview, limited to Apple Silicon Macs running macOS 14 or later and to ChatGPT Pro users who pay at least $100 per month. It does not ship in the European Union, the United Kingdom or Switzerland, a restriction that hints at OpenAI’s awareness of GDPR‑related data‑minimisation rules. Users must grant macOS screen‑recording and accessibility permissions before the feature can operate.

OpenAI’s president, Greg Brockman, described Chronicle as “an experimental feature giving Codex the ability to see and have recent memory over what you see, automatically giving it full context on what you’re doing.” In practice, the assistant can now infer which apps are open, what documents are being read, and even which lines of code are being edited, without the user having to restate that information.

Unlike competitors that keep processing on‑device, Chronicle relies on cloud‑based analysis. Raw screenshots reside temporarily in a system‑level temporary folder and are deleted after six hours. OpenAI asserts that the images are not retained on its servers and are not used for model training. However, the text memories remain on the computer indefinitely as unencrypted files, potentially exposing sensitive data to any process that can read the directory.

The privacy trade‑off has drawn sharp criticism. OpenAI’s own documentation warns that Chronicle “increases risk of prompt injection” because malicious content captured in a screenshot could be interpreted as instructions. The company also advises users to pause the feature before meetings or when handling confidential material, effectively shifting risk management to the end user.

Chronicle’s approach contrasts with Microsoft’s Recall, which captures screenshots on Windows devices but stores them in an encrypted local database and processes them on a neural processing unit without sending data off‑machine. OpenAI’s decision to send images to the cloud prioritises utility and rapid context generation over a privacy‑first architecture, a stance that may limit adoption in regions with stricter data‑protection laws.

Industry observers note that screen‑aware AI assistants have a turbulent track record. Early entrants like Rewind AI were acquired and discontinued, while Microsoft’s Copilot has seen subscriber churn partly attributed to trust issues. Open‑source alternatives such as Screenpipe offer entirely on‑device capture and processing, positioning themselves as privacy‑centric rivals.

Chronicle arrives amid a broader push toward ambient computing, where AI systems aim to understand user intent without explicit commands. Apple is reportedly testing AI‑enabled smart glasses, Slack has integrated deeper AI context into its platform, and Gartner predicts that more than 40 % of large enterprises will pilot ambient intelligence by 2026. OpenAI’s gamble is that the productivity gains from seamless context will outweigh privacy concerns, a balance that will be tested as regulators and users scrutinise the feature’s data handling practices.

For now, Codex users who opt into Chronicle enjoy a smoother workflow, but they must weigh that convenience against the possibility that sensitive screen content could be captured, transmitted, and stored in a format that lacks encryption. Whether the feature’s promise of “magical” assistance translates into sustained adoption remains an open question.