A response to Tsvi Benson-Tilsen's A case that whole brain emulation research is net-harmful by default. Four forks: how soon AI arrives, how hard uploading is, whether a stalled project keeps its secrets, and whether partial brain emulations would help AI at all. Every leaf shows its probability. Green is good for the world, grey is neutral, red is the one bad leaf. The numbers are Michael Andregg's; change them and the tree recomputes.
Only one leaf is bad. AI timelines are long enough for uploads to matter, uploading turns out hard, the partial artifacts leak, and they are useful to AI. Everything else is neutral or good. The strip under the tree puts all five leaves on one line so the bad term sits next to the good ones rather than on its own.
The 90% branch is a "not both". It is one minus (leak × useful). With secrets kept 80% of the time and partials useful 50% of the time, the bad combination is 20% × 50% = 10%, and everything else is 90%.
Hard is not the same as failure. The 90% branch splits again on whether a slow, hard program still produces an upload before superintelligence. The half that does is as good an outcome as the easy path, only later. Under the default numbers the good leaves total about 51% and the bad leaf 3.5%.
One simplification. The tree uses a single "secrets kept" figure for the whole hard branch. The slow-but-lands leaf has the longest window with partial artifacts in hand, so its true leak probability is somewhat higher than the shared figure. Lowering "secrets kept" shows the sensitivity: at 50% kept the bad leaf is 8.75%.
Plug in your own numbers. The second preset is an illustrative sceptical set, roughly 3% short timelines, 5% easy, 30% secrets kept, 70% useful, 20% still lands. It gives about 45% bad. The disagreement lives in the forks, not in the structure of the tree.
Drawn by Michael Andregg, Eon Systems, September 2026, with help from Claude. The numbers are Michael's own estimates. The page is a single self-contained file with no tracking.