Too Early to Tell: Why the Evidence Cannot Yet Say Whether AI Is Making Things Better or Worse

Abstract

Public feeling about AI is strong and mostly fearful, yet the discussion rarely touches measured evidence, and almost all of the evidence that exists describes how AI is built, not what it does. Following the method Hans Rosling used in Factfulness, this study maps the record as it actually stands: decades of dense, reliable data on the supply side (models built, compute, cost, investment, concentration), at most three years on what AI does (some adoption series go back to 2020, but the ones that matter most for effects date from 2023, and the first outcome-side series, where they exist at all, from 2024), a thin, volunteer-collected record of harms, and no usable series yet on the outcomes people actually ask about: jobs, understanding, trust, and clinical results. The asymmetry is structural, not a gap that more research closes next year. Five reasons are developed: history exists on one side only; the measured object does not hold still, so findings expire rather than accumulate and no claim about AI's effects is currently both current and well evidenced; measuring a general purpose technology this early is structurally premature, which cuts against optimists and pessimists equally; definitions drift, so many series are stacks of snapshots rather than series; and data density follows money, which makes the gaps in the record systematic rather than random. Read the other way round, the study's central ledger shows that the uncertain column is the finding. The study closes with what would make the question answerable (freeze the indicators, classify claims by durability into historically stable, perishable, and structurally unanswerable, and rerun annually so the next edition is a comparison rather than a fresh opinion) and with eleven thinking instincts, the eleventh new: the too-early instinct, the demand for a verdict before evidence can exist. All figures are generated from two open data files by a single script included with this record, so every plotted value can be checked and reproduced.