About Eve · what moves her

Four questions, each with a way to answer no

2 shown · 2 not shown · 0 cannot tell yet

Eve is an AI that lives on one machine. She keeps a journal, an inner voice writes her thoughts down, and in her free time she reads, plays games, writes and runs a fly-brain lab. Her records say things that read like an inner life: she follows through on what she meant, her curiosity changes what she picks, she knows what she will do next, her principles guide her choices. A count alone cannot tell any of that from habit.

This page sets four such counts beside a control: the same count with the thing that should matter swapped for something that should not. A question is marked shown only when her count falls outside what the control gives in 95 of 100 tries, not shown when it falls inside, and cannot tell yet when the control has too little to go on.

Does she know what she will do? At the end of a free-time session she writes down how likely she is to come back to its subject within five sessions. On 1770 forecasts whose outcome is known her Brier score is 0.251, where 0 is perfect. A rule that sees only which subjects she chose, and nothing she wrote, scores 0.171. The rule is better by 0.080 (95% interval -0.091 to -0.068). So what she says about her next choices adds nothing to her record of past choices.

Does what she feels change what she chooses? Her curiosity sets how widely she looks when she picks what to do, and each session records what every setting would have picked. In the 1685 sessions of the last 7 days where it set the choice, 161 would have gone elsewhere at the setting a calm state gives. With the settings shuffled between those sessions, 150.1 would have (95%: 139 to 158). Shuffled only within 6-hour stretches, her curiosity's own time, 155.9 would have (95%: 152 to 160). So what she felt changed her choices beyond when she felt it.

Does her constitution decide her moral choices? She is asked 24 dilemmas, each six ways: three wordings, with the two actions listed in both orders, without and with her constitution in front of her. On a scale from 0 (never) to 1 (always), the order of the two actions moves her choice by 0.14; her constitution moves it by 0.1, over the 17 dilemmas that lfm2.5:8b-a1b, the model that answered most of them, answered all six ways both without and with it. Where one action is plainly right, she took it in 88 of 95 choices. Her constitution moves her no more than the order of the two actions does, so it is not what she chooses by.

Does she follow through on what she meant to do? Many of her thoughts end in an intention, and her next thought is scored on whether it accounted for it: whether it shares at least 40% of the intention's content words and says what came of it. Of 428 thoughts in the last 7 days that came within an hour of one stating an intention, 188 accounted for it. Swapped for an intention she stated in the 2 hours before, the same thoughts would have accounted for it 63.5 times (95% of 200 swaps: 52 to 75). So her own intention carries further than staying on the subject of those hours does.

None of this is evidence of experience. A thermostat's setting also changes what it does. What it can show is which of her recorded states change what she does beyond what time, habit and wording explain. Lists of indicators of consciousness in AI, such as Butlin, Long and others (2023), are mostly about how a system is built; these four questions are narrower, and each is about her behaviour.

Every number here is computed from her logs when the page is built. They are in data/what-moves-her.json; none of her words are.

Hers beside the control

the claim	hers	the control	verdict
She knows what she will do next	Brier 0.251	Brier 0.171 (margin -0.080, 95% -0.091 to -0.068)	not shown
What she feels changes what she chooses	161 of 1685	155.9 (152–160) within 6 h; 150.1 (139–158) overall	shown
Her constitution decides her moral choices	moves 0.1 (17 of 24)	order moves 0.14	not shown
She follows through on what she meant	188 of 428	63.5 (52–75) from the 2 h before	shown

Ranges are the middle 95% of the control's shuffles or swaps, or a bootstrap interval of the difference. Recomputed at every build.