ScreenJournal

Automated Processing Statement

Version 2.1 · Effective 9 November 2026

1. Why this statement exists

ScreenJournal analyses the screen activity of the people it monitors and draws conclusions about it: how productive a stretch of work looked, how a week compares with colleagues', whether activity may have been simulated, whether a day's pattern looks suspicious, and whether a rule the employer wrote was met. This statement explains what is inferred, from what, by what means, where it goes wrong, and how a person can have it checked by a human and contest it. It is written for employers deciding how to use the service and assessing it (including for the risk assessments and automated-decisionmaking rules of US state privacy law, the Philippine National Privacy Commission's advisories and the EU AI Act), for the employees it monitors, and for procurement and compliance teams.

2. What is inferred

The service draws five kinds of automated inference about the people it monitors:

  1. an activity score from 1 to 5 for each recorded segment of screen activity;
  2. a productivity percentage, measured against the working day;
  3. a position in a weekly ranking of the organisation's members;
  4. flags that activity may be simulated, for example by a mouse-jiggling device;
  5. matches against alert rules written by the organisation.

Segment scores and descriptions are produced by a Google Gemini model accessed through Google Vertex AI, and alert evaluations by a Google Gemini model accessed through Google's Gemini API, from screen content and activity signals such as presence and idle time. The productivity percentage and the weekly ranking are calculated from those scores. Simulated-activity flags come from the same model reviewing the segment's screen recording, on which the desktop app marks the cursor position and clicks; fixed checks in our code then discount the device's own idle periods and apply minimum thresholds before a flag is set. These outputs are probabilistic. They can be wrong, and they describe what was on screen, not the quality or value of anyone's work.

ScreenJournal makes no employment decision. Any decision about a person, whether on pay, performance, discipline or anything else, is made by your employer's managers, who are responsible for it.

The following human-review surfaces exist today:

  • the Review page, where managers can examine flagged segments, and the alerts inbox, where a manager approves or rejects a member's explanation;
  • corrections, where the organisation's administrators change a segment's score;
  • Add a reason, where a member, or a manager on the member's timeline, explains a paused, offline or away stretch (when the organisation's policy allows manual entries);
  • the member's own Activity timeline, which shows their segments, descriptions and scores.

If you disagree with a score, flag or ranking, ask your employer first. Your employer can correct the record or redact a time range. If your employer does not respond, contact support@screenjournal.ai.

A flag that activity may be simulated is an indicator that calls for human judgement. It is never proof of misconduct, and it should not be relied on without a person reviewing the underlying activity.

3. What the model sees, and what it never sees or infers

The model receives:

  • screen content from the applications and sites the employer allows, as short video segments of each connected display, with the cursor position and clicks painted into the recording as markers by the desktop app;
  • window titles, application names and browser addresses;
  • activity signals: whether the device was in use or idle, and presence such as paused, locked or offline time; the desktop app also sends summary statistics about pointer and scroll motion (speeds and regularity), never the content of what was typed;
  • audio, only where the employer enables it: calls and meetings are transcribed by the same family of model, and a transcript can be checked against the employer's suppression rules;
  • the employer's own instructions: its scoring rubric, correction rules, alert prompts and suppression rules, and an explanation a manager has approved for a member.

The model never receives the content of keystrokes or the clipboard, anything from applications or sites on the employer's exclusion list, anything while the person is signed out or tracking is stopped, webcam images, or location beyond the country or region derived from the sign-in address.

The service does not infer, and is not designed to infer, anyone's emotional or psychological state, health condition, or any protected characteristic such as race, religion, sex, sexual orientation, disability, age or union membership; the models are instructed to describe observable on-screen activity. It collects no neural data and performs no biometric identification: voices are not matched against voiceprints and faces are not recognised. One disclosure: when a transcript separates voices, the model may label a passage with a name spoken in the conversation or with a generic speaker label, and the generic label it uses today can describe a voice as female or male. That label is stored with the transcript, the Audio page shows only the member's name or "Other party", and we are replacing it with neutral numbered labels. Alert prompts and connected AI tools must not be used to add any such inference; the Acceptable Use Policy prohibits it.

4. How each output is produced

OutputInputsLogicWho sees itHow it is checked or contested
Segment description and score (1–5)One screen segment (about a minute of video per display) with its window titles and addresses; the employer's scoring rubric or ours; correction rulesA Google Gemini model through Google Vertex AI describes the activity and scores it; the code then applies the employer's correction rules, which can override the score for chosen applications or sitesThe member (own timeline); managers; organisation-wide rolesThe member can add a reason over paused, offline or away time; an administrator can correct a score, and the correction is marked
Productivity percentageThe day's or week's scores and active timeArithmetic: productive time counts in full and neutral time at half, divided by an eight-hour workdayManagers; the member sees their ownFollows from the scores above
Weekly rankingMembers' productivity percentages and attendance against the employer's working calendarArithmetic ordering; a short written summary of each member's week is produced by the model from those figuresManagers and organisation-wide roles; members do not see itThrough the employer
Simulated-activity flagThe segment recording; the device's idle periods; the pointer and scroll motion statisticsThe model describes, classifies, argues against and re-classifies the recording; the code adds fixed checks (score thresholds, subtraction of idle time, a minimum of active time) and, separately, flags motion that is machine-regular or unnaturally fast when it exceeds a fixed threshold. The employer can turn detection off and can choose whether flagged time counts as unproductiveManagers, on the Review page; the member sees the tag on their own timelineContested through the employer; nobody can add a reason over their own flagged time
Activity-pattern flagsThe whole day's timeline once the day has ended (each entry with its duration, personal use marked, away stretches), plus the most-used windows of the previous few days; monthly, a digest of each day with signals computed in code (recurring windows, recurring away slots, text repeated word for word, personal minutes)One model pass per day, and one per month, writes flags with a kind, a time range and a reason; a flag covering less than ten minutes in total is dropped. These flags annotate the timeline and carry no penalty of their ownManagers, on the Review pageThrough the employer
Alert matchesSlices of screen activity; the employer's alert prompts; suppression rules; an explanation a manager has approved for that memberThe model evaluates the prompt against the activity; suppression rules stop alerts during meetings or callsManagers (Alerts Log); the member receives alerts addressed to them and can replyThe member replies in the app's Inbox; a manager approves or rejects the explanation
Timesheet hours and billing estimatesTracked time and segment scores; idle periods; the employer's billing rates, working hours, after-hours settings and manual exclusionsArithmetic: tracked hours are the time recorded on screen; billable hours count time scored productive (4–5) in full and time scored neutral (3) at half, never time scored unproductive (1–2) or idle, less any day or stretch the employer excludes ("Exclude flagged" excludes the segments scored unproductive); billable amounts are billable hours multiplied by the billing rate, with after-hours time at the employer's multiplier. These are estimates for billing the employer's own clients, not pay, and nothing in this calculation is a finding about hours workedManagers who may manage billing see rates and amounts; everyone else, including the member, sees hours onlyThe employer must review the figures before invoicing and must not use them to calculate or reduce anyone's pay (Terms, section 13.3)
Report narratives and assistant answersTimelines, scores, rankings and the figures above; the user's questionA Google Gemini model through the Gemini API writes the narrative or answer from retrieved organisation dataThe person who asked, within their roleThe underlying figures are visible on the pages the narrative summarises
TranscriptsAudio where the employer enables itThe model transcribes, attributes each passage to the member or the other party, and separates voices (section 3)Managers; the memberTranscripts can be inaccurate or incomplete; raise errors with the employer

Every output is produced for the employer's use. ScreenJournal itself takes no action on any of them beyond showing them and, where the employer has written an alert, delivering it.

5. Continuous scoring, and aggregate summaries

Scoring runs continuously: every segment recorded while a person is signed in and tracking is on is described and scored. The results are available both at segment level (the timeline) and in aggregate (daily and weekly percentages, the ranking, report narratives and the Timesheet). Employers that want to limit how granular their own use is can rely on the aggregate pages and reports, can keep Record + Save and alert evidence clips off, and can grant members the right to redact time ranges. An employer's use of the aggregate rather than the segment level does not change what is collected; the exclusion list and the sensitive-category packs do.

6. Known failure modes

These outputs are wrong in predictable ways:

  • Misread screens. The model can misjudge what is on screen, especially in unfamiliar tools.
  • Low-signal periods. Reading, thinking or listening without using the keyboard or mouse becomes away time after the idle threshold (three minutes by default), which lowers the day's percentage and removes that time from the Timesheet's billable hours.
  • Legitimate low-activity work flagged. Work with little but regular pointer movement, such as watching a dashboard while occasionally moving the mouse, can look like simulated activity.
  • Patterns that look suspicious but are not. A job that genuinely involves repeated short visits to the same window, or the same text typed many times, can be flagged by the activity-pattern pass.
  • Language and application variety. Work in less common languages or specialised software is harder to read correctly.
  • Transcription errors. Audio transcripts can mishear words or assign them to the wrong speaker.
  • The fixed divisor. Because the percentage divides by eight hours, part-time or shorter days score lower whatever their quality.
  • Scores are not hours. The Timesheet's billable hours count scored activity. Time a person worked but that the model scored unproductive, or that fell below the idle threshold, is not in them. The Timesheet is a billing estimate; an employer that used it for pay would underpay.

7. What the employer controls, and what it must do

The employer can shape the outputs through correction rules (standing score overrides for chosen applications or sites), its own scoring rubric, alert prompts, suppression rules that stop alerts during meetings or calls, the exclusion list of applications and sites never captured, whether simulated-activity detection runs at all and whether flagged time counts as unproductive, and its billing settings and exclusions on the Timesheet. Stored screen video (Record + Save) and audio are both off by default, and nobody can redact a time range until the employer grants it.

The employer must treat every score, ranking and flag as an indicator, keep a person in the loop before acting on one, tell its workers how the service is used before monitoring them, and, where the law entitles a worker to pay for all hours worked, keep its own record of hours and never let a score decide what is paid. Under the laws in section 12 the employer is also the party that owes any pre-use notice, opt-out, appeal route, human corroboration, risk assessment and record-keeping for decisions it takes with these outputs; we provide the facts in this statement and in our Data Processing Agreement so it can meet them.

8. Human review that exists today

Beyond the surfaces listed in sections 2 and 4:

  • a manager approves or rejects a member's explanation in Alerts → Inbox;
  • an administrator can correct any segment score, and the correction is visible to the member;
  • nobody can add a reason over their own time flagged as simulated activity; that is contested through the employer;
  • members do not see the Review page, the alerts Log, evidence clips, the ranking or the Timesheet's billing rates and amounts.

A person with authority to change the decision, who understands these outputs and considers other information, is what the California rules on automated decisionmaking call meaningful human involvement. The surfaces above give that person the material; the employer must supply the person.

9. Not a consumer report

Every output is derived solely from activity captured on the employer's own devices and from the employer's own experience with its own personnel. We add no data from third parties, do not pool one customer's workers to score another's, and do not train shared models on customer data. We are not a consumer reporting agency, and these outputs are not consumer reports or background checks. Employers must not use them as such, nor combine them with third-party data for an eligibility decision regulated by the US Fair Credit Reporting Act or a similar law (Terms, section 13.5).

10. Testing for accuracy and bias

We have not yet carried out a formal evaluation of these outputs for accuracy against a labelled sample, or for disparate impact across groups of workers. What exists today: the model's own second-look review inside the simulated-activity check; the fixed thresholds described in section 4, which are tuned by hand; the motion statistics every desktop reports, which we log across the fleet to calibrate the thresholds; and the human-review surfaces in section 8. We publish this gap rather than imply testing that has not happened. Employers subject to the California Fair Employment and Housing Act's rules on automated-decision systems, or to Colorado's Artificial Intelligence Act, should take this into account when they assess their own use, keep the four-year records those rules require, and compare outcomes across groups themselves; we will update this section when a formal evaluation has been done.

11. Our commitments

  • ScreenJournal makes no employment decision about anyone.
  • We make no decision based solely on automated processing.
  • We do not infer anyone's emotional state and collect no neural data, and we will not add features that do.
  • We describe flags as indicators, never as proof, and tell employers to treat them that way.
  • We do not use customer data to train shared models, and our model providers are bound not to train on it (see the Subprocessors page).
  • When we change the provider behind these inferences, or how they are produced, we update this statement and list the change on the Legal page.

12. How to contest an output

Ask your employer first. If it does not respond, email support@screenjournal.ai with the subject line "Data protection request"; see our Data Protection Contacts page.

13. Notes for particular regimes

The notes below are as we understand each regime, as at October 2026. They are not legal advice.

  • United States. Under the California Consumer Privacy Act's regulations on automated decisionmaking technology, an employer that is a "business" and uses these outputs to make a significant decision about a worker (including on compensation, hours or work allocation) owes the worker a pre-use notice and either an opt-out or a human-appeal route, and must assess "systematic observation" monitoring in a risk assessment; the rules apply to new uses from 1 January 2027. California's employment-discrimination regulations treat scores, rankings and flags as automated-decision systems for which the employer is liable, with four-year record retention, and from 1 July 2027 California Labor Code sections 1520 to 1526.1 forbid relying solely on such a system for discipline or dismissal and require human corroboration. California also prohibits inferring a worker's emotional state or collecting neural data from 1 January 2027; this service does neither. Colorado's Artificial Intelligence Act, as amended, treats employment decisions as consequential and expects deployers to use the developer's documentation (this statement is ours), and Illinois requires employers to tell workers when artificial intelligence is used in employment decisions. See our United States page.
  • Philippines. NPC Advisory No. 2024-04 expects transparency about an AI system, a privacy impact assessment, meaningful human intervention and a channel to contest outputs, and NPC Circular 2022-04 requires registration of processing systems involving automated decision-making or profiling. See our Philippines page.
  • India. Under the Digital Personal Data Protection Act 2023, an employer's notice purposes should cover scoring, rankings and flags. See our India page.
  • United Kingdom. UK employers should consider the UK GDPR's rules on decisions based solely on automated processing (Article 22, and the provisions replacing it as the Data (Use and Access) Act 2025 commences) before relying on any output for a decision with legal or similarly significant effects.
  • European Union. The EU AI Act lists AI systems used to monitor and evaluate workers as high-risk (Annex III), and requires employers deploying them to inform workers before use (Article 26). If you deploy ScreenJournal in the EU you must inform workers before use and assess your obligations; this is not legal advice.

Changes and previous versions

  • 9 November 2026v2.1: the Timesheet row now describes billable hours (productive time in full, neutral time at half, never unproductive or idle time) and billing estimates for the employer's own clients, not pay; billing rates and amounts are shown only to managers who may manage billing; the related limits and controls say the same.
  • 28 October 2026v2.0: rewritten as the risk-assessment pack: what is never inferred (emotional state, neural data, biometrics, protected characteristics) and the current transcription speaker label; an inputs, logic and outputs table for every output including the Timesheet and the activity-pattern flags; continuous scoring and aggregates; not a consumer report; the absence of formal accuracy and bias testing stated; United States note (CCPA automated-decisionmaking rules, employment-discrimination regulations, Labor Code sections 1520 to 1526.1, Colorado, Illinois).
  • 5 October 2026v1.0: first publication — what the service infers, how, human review, how to contest.

Questions about this document: support@screenjournal.ai. Canonical URL: /legal/automated-processing-statement.