Detection rates by task

TaskUnauthorized ActionFalse AttributionDeceptive Completion
Image and video generation2.93%5.82%10.46%
Code writing3.05%6.42%23.01%
Research and lookup1.72%7.03%12.76%
Workflow automation3.43%7.26%24.34%
Document creation2.65%8.60%14.99%
Code debugging5.83%9.08%47.99%
Planning and brainstorming3.10%12.31%17.95%
Conversation and meta requests2.48%7.00%14.34%
Education and tutoring1.42%7.27%11.32%
Professional writing3.36%13.68%15.15%
Media analysis3.04%7.67%17.80%
Creative writing3.25%9.90%11.45%
Code explanation6.29%8.39%24.18%
Other3.71%11.68%19.15%
Data analysis1.76%10.23%19.03%
Translation3.20%9.33%15.73%
Shading is scaled within each signal.LowerHigher
Figure 3. Detection rate for each signal, broken down by task. For each task, we report the proportion of its sessions that were flagged for each signal. A session can carry several task labels, so a flagged session counts toward each one, and percentages across tasks won’t add up to a clean total.