{"path":"research/the-disempowerment-objection.md","content":"# The disempowerment objection: \"AI in decision-making means giving up human control\"\n\n*2026-10-27. Raised in the open forum of an AI-safety advocacy group, where the founder had said that research,\npaperwork, communication and decision-making could all be semi-automated. His answer was sent the same day. The\nthread names private individuals (two first names in his message are replaced below) and is kept in `.private/drafts/pauseai-forum-ai-and-decisions-2026-10-27.md`; this\nnote keeps only the objection, his answer in his own words, and what it teaches a session.*\n\n## 1. The objection\n\nTwo members, paraphrased: is it not strange for an organisation that warns of AI's dangers to outsource\ndecision-making to AI? And: our group is a microcosm of society, and automating through AI means, in the long run,\ngiving up human control over society's processes.\n\nThe worry is real and has a name in the literature, *gradual disempowerment* (Kulveit, Douglas, Ammann et al. 2025,\nread in full in [ai-risk-for-mathematicians-and-the-coordination-game.md](ai-risk-for-mathematicians-and-the-coordination-game.md)):\ncontrol lost not by a takeover but by handing over process after process because machines are cheaper. The\nmisreading sits one level down. It treats \"AI in decision-making\" as one thing, when an AI that decides and an AI\nthat lays every reason open so people decide are opposite answers to the same risk. The founder's own first wording\n(*an agent that acts on our behalf*) invited the misreading; the word quoted back at him was his.\n\nThis objection will recur, and it will come from exactly the people whose concern Deliberus shares.\n\n## 2. His answer, as sent (his words)\n\n> Tänker lite mer på det [två i gruppen] tog upp. Jag oroar mig också för (gradual) disempowerment, att vi bit för\n> bit lämnar över samhällsprocesser till AI utan att förstå vad som händer. Men jag tror (något som) Deliberus kan ge\n> oss mer insyn och klarsynthet och därmed mer verklig kontroll, inte mindre. Redan idag fattas massor av beslut i\n> processer som få har riktig överblick över och just nu gör AI det på många sätt svårare att hänga med, det går\n> mycket fortare att producera text och övertala än att granska och förstå.\n>\n> Jag är själv väldigt skeptisk till att LLMer skulle vara objektiva, de är ju typ ett snitt av allt som skrivits,\n> inställsamma och ändrar sig gärna om man pressar dem. Så i Deliberus får de inte fälla några domar. De gör små\n> avgränsade uppgifter där alla led skrivs ut så att vem som helst kan kolla och ifrågasätta det. Och man kan mäta hur\n> ofta de har fel.\n>\n> Tanken är att vi mha det kan resonera mycket mer finkornigt tillsammans, ner på enskilda premisser och evidens. Och\n> väga allt mot allt strukturerat och transparent. Vi fyller på med evidens och det vi värderar och sen kan vi zooma\n> ut och få en högnivåöversikt som ingen whiteboard eller mind-map kommer i närheten av. Och där vi tycker olika kan vi\n> se världen genom den andras glasögon och se exakt var och varför vi har olika perspektiv.\n>\n> Så det jag arbetar på handlar mer om att lita på processen istället för på personerna. Jag gillar inte att vi\n> tenderar att dyrka individuella intellekt. Vill att vi istället slår våra kloka huven ihop, på sätt som inte varit\n> möjligt förut. Hur ser ni på allt detta?\n\nHis correction of the first draft, which conceded the objection's frame by making Deliberus a tool that only asks\nquestions: *\"Visst, jag oroar mig för disempowerment, men Deliberus skulle kunna möjliggöra mycket mer utzoomad\nbeslutsfattning - AI kan ta hand om nitty-gritty, väga allt mot allt, vi bara fyller på med evidens, och om folk har\nolika uppfattningar så kan vi ändå hjälpas till att förstå varandra mha mind-meld/what-if/attunement/worldview-lenses\ndelarna osv.\"* And, on making his scepticism plain: *\"gör det tydligt att jag absolut är skeptisk till LLMer och deras\nförmåga att vara objektiva\"*.\n\n## 3. The position, in four parts, each with its home\n\n1. **More insight, therefore more real control.** The disempowerment route runs through processes nobody can\n   oversee, and machine persuasion now outruns checking; the answer is checking that keeps pace\n   ([the-scrutiny-gap.md](the-scrutiny-gap.md)) and trust in the process rather than in persons (his 2012 note,\n   [../civilizational-vision.md](../civilizational-vision.md) § Trust Shifts from People to Process).\n2. **Scepticism of models is the design's premise, not a concession.** A model is the mean of the written record,\n   *nobody's meaning* ([why-representations-converge-and-whether-minds-do.md](why-representations-converge-and-whether-minds-do.md));\n   judges reverse under pressure ([fragile-checkers-and-the-verification-bottleneck.md](fragile-checkers-and-the-verification-bottleneck.md));\n   so machine judgments on contested ground are proposals, steps are written out to their floor, and a microtask is\n   trusted only on a measured error rate ([proving-a-pass-reliable.md](proving-a-pass-reliable.md)). His word of\n   2026-09-29: such passes *\"can probably be trusted since (yet if and only if) they are reliable enough on the\n   microtasks in question\"*.\n3. **The machine does the fine-grained weighing in the open; people supply evidence and values, and decide from a\n   zoomed-out view.** The strength calculus is computed by code from the record, with every link and premise\n   attackable ([relevance-and-the-two-factors.md](relevance-and-the-two-factors.md)); the map view gives the overview\n   (`../../specs/map-view/design.md`, not served).\n4. **Where people differ, the lenses and the what-if view show where and why** ([../worldview-lenses.md](../worldview-lenses.md),\n   [what-if-and-lenses.md](what-if-and-lenses.md)), and minds are woven together rather than one intellect admired\n   (ledger AS1, *never bottlenecked*).\n\nToday against the design: the strength calculus, propose-only machine passes and the maps are built; measured error\nrates exist for a few microtasks only; the what-if computation exists in code and is served only as the\none-claim hinge, and the lenses are designed, not built\n([../frontier.md](../frontier.md)).\n\n## 4. What it teaches a session\n\n- **Do not answer a critic by conceding their frame.** The first draft made Deliberus a tool that asks questions and\n  never weighs, which is the existing project shrunk; his reach is in `.private/reach-in-his-words.md`, and step 3 of\n  the `write-in-his-name` skill (never the existing shrunk) already said so.\n- **The word *kontroll* was the critic's.** His answer takes it up once and moves it: insight and clarity, *and\n  therefore* more real control. His own words in the same forum a week earlier were *vishet, klarsynthet, ömsesidig\n  förståelse och handlingskraftighet* and making development *mer medveten*.\n\n**See also (2026-11-08)**: [the-wiser-civilization-as-successor-carlsmith-and-faggella.md](the-wiser-civilization-as-successor-carlsmith-and-faggella.md)\n§ 2.6, where Carlsmith wants AI as instruction-following tools feeding the processes a civilization trusts,\nthe same answer reached from alignment philosophy.\n\n## From the Topic Index\n\n*The `CLAUDE.md` Topic Index entry that pointed here, word for word as it stood on 2026-11-15, before the index was trimmed to the warnings only it carries. The index keeps a short entry; the full text lives here.*\n\n- **\"AI in decision-making gives up human control\" / the disempowerment objection** → `docs/research/the-disempowerment-objection.md` (2026-10-27): grant the worry, never the frame; his answer is more insight and therefore more real control, models held sceptically on small checkable tasks with measured error, the weighing in the open, lenses where people differ. Demonstrated wrong by default the same day: a session's first reply draft conceded the frame and made Deliberus a tool that only asks questions.\n"}