Petar Juršić: ideas and stories

Context building and AI

5 min read

Knowing the context is what AI cannot supply for you, and the evidence for that is mixed

Hook

In one study, consultants who used AI on a task built to trip it up got worse: 84.5% right without it, 60% and 70.6% with it.

Claim

Opinion: in the AI era a key skill is judging context: knowing what the task is for, what the data leaves out, and where the tool is likely to be wrong. The research found does not test “context building” directly. It supports a narrower point: people do badly when they cannot tell where a tool is reliable, and gains depend on the task and the user’s familiarity with it.

Evidence:

  • BCG, Harvard, MIT and Warwick study (Dell’Acqua, McFowland, Mollick, Lifshitz, Kellogg, Rajendran, Krayer, Candelon, Lakhani; Organization Science 37(2), 403-423, 2026, DOI 10.1287/orsc.2025.21838, online 11 March 2026, earlier circulated as a 2023 working paper): a preregistered experiment with 758 BCG consultants (about 7% of its individual-contributor consulting workforce), randomly given no AI, GPT-4, or GPT-4 plus a prompt-engineering overview. On 18 tasks inside the AI’s ability, AI users completed 12.2% more tasks and 25.1% faster, with significantly higher quality (Mollick’s 2023 post gives 40% higher quality and a 43% gain for the lowest performers; I did not find those two numbers in the published text, so treat them as working-paper figures). On one managerial task built to be outside the frontier, the control group was correct 84.5% of the time against 60% and 70.6% in the two AI groups, an average drop of 19 percentage points (the abstract says “19% less likely”). Bottom-half performers gained more than top-half on the inside-frontier tasks.
  • METR randomized trial (July 2025; Becker, Rush, Barnes, Rein): 16 experienced open-source developers, 246 tasks in repositories they had worked on for about five years. With AI allowed (mostly Cursor Pro with Claude 3.5/3.7 Sonnet) tasks took 19% longer (interval +2% to +39%). Developers expected a 24% speed-up beforehand and afterwards still believed AI had sped them up by 20%; economics and ML experts had forecast about 39% and 38%. METR says it does not claim AI slows most developers, the sample is small, and learning effects cannot be ruled out.
  • METR follow-up (24 February 2026): 57 developers, 143 repositories, 800+ tasks. Result for returning developers -18% time (interval -38% to +9%), new developers -4% (-15% to +9%); negative means faster, and neither is clearly different from zero. METR calls this “an unreliable signal” because developers refused to work without AI and skipped tasks they did not want to do without it (30-50% said so), so it is probably a lower bound on the speed-up. The 2025 result is therefore dated and the direction may have flipped.
  • Survey of 319 knowledge workers (Lee et al., Microsoft Research and Carnegie Mellon, CHI 2025; 936 self-reported examples): higher confidence in the AI went with less critical thinking, higher self-confidence in the task with more. Self-reported and correlational, so it supports the point only loosely.

Reading (opinion): the METR developers knew their repositories well, and the BCG consultants were skilled professionals meeting a task built to mislead. The thread is that the user’s knowledge of the situation decides whether the tool helps. That is an inference, not a finding in either paper.

What cannot be shown: that “context building” is a distinct skill that can be trained, or that it matters more than prompting skill or domain knowledge.

Story

A consultant in the BCG study reads an AI answer that looks well written and accepts it. On the task designed to trap the tool, it was wrong. The missing skill was not typing; it was knowing when to doubt.

Sources

Check before posting

  • Resolved 2026-11-04: volume/DOI, METR February 2026 post opened, all numbers except 40% and 43% confirmed in the primary sources.
  • Open: 40% quality and 43% gain are not in the published text found; either cite Mollick’s 2023 post as the source or leave them out.
  • Open: SSRN page and the INFORMS page could not be opened directly (403); the published details come from the Crossref record and the open-access PDF.
  • The 84.5/60/70.6 split is one task; do not generalise it to “AI makes people worse”.

Counterargument

Models and tools are improving; the lessons from 2023 and early 2025 may not hold. And “know your context” was good advice before AI.

Post angles

Facebook (Croatian) Jedno istraživanje na 758 savjetnika pokazalo je nešto neobično. Na zadacima u kojima je AI dobar, pomogao je. Na zadatku koji je namjerno bio zamka, ljudi bez AI-ja imali su 84,5 posto točnih odgovora, a s AI-jem 60 i 70,6. Zaključak: nije presudno koliko dobro pišete upute, nego znate li kada ne vjerovati odgovoru. A to znaju ljudi koji poznaju situaciju.

LinkedIn (English) Two studies worth knowing. BCG’s: on tasks inside AI’s ability, consultants were faster and better; on one task built to expose its blind spots, accuracy fell from 84.5% to 60-70.6%. METR’s 2025 trial: experienced developers working on repositories they knew took 19% longer with AI, while believing they were faster (METR’s February 2026 update found a possible speed-up but called its own data an unreliable signal). Neither tests “context building” directly, but both suggest the person who knows the situation is the one who can tell when the tool is wrong.

Technocracy for the people means experts whom citizens can check Systems fail when true information stops reaching the people who can act on it

Graph View

Backlinks