Cheap Fallible Cognition
B · Useful, limited A carefully built conceptual framework from two practitioners (a software CEO and a Morgan Stanley investment chief) that collects no new data and is not peer-reviewed, so it frames the debate rather than settles it.
Summary based on the full text·Original: about 45 minutes to read
In short
An arXiv paper (2608.11512, 11 August 2026) by Christophe Kolb, CEO of Taller, and Jim Caron, CIO of the Portfolio Solutions Group at Morgan Stanley Investment Management. It treats generative AI as cheap, scalable and fallible cognition and argues that "will AI destroy jobs" is the wrong question. Labour effects depend on verification costs, liability, question selection, workflow redesign, demand elasticity, apprenticeship and who captures the rents. The central worry is expertise formation: automating junior tasks can raise output today while cutting the pipeline that produces tomorrow's accountable experts.
Key numbers
- 5design inequalities that must each hold for cheap machine cognition to raise welfare
- 16%relative employment decline for workers aged 22-25 in AI-exposed jobs, as cited from Brynjolfsson, Chandar and Chen (2025)
- 19 ppdrop in correct solutions on a task outside the AI frontier in the 758-consultant field experiment the paper cites
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- A task vulnerability index built from eight task traits: digital observability, output regularity, ex-post evaluability, retrieval intensity and error tolerability push exposure up; tacit context, liability and coordination-and-trust push it down.
- An adoption condition: a task moves to AI only if some level of verification meets the error and quality floors and the full machine cost (model, verification, integration, governance, expected error loss) is below the human cost.
- Professional cognition has three stages: question selection, answer generation and accountable use. AI mostly cuts the cost of the middle stage, which raises the value of the other two.
- A labour-demand accounting equation splits the change in hours into task compression, scale expansion through lower prices, and new plus complementary human work.
- Junior tasks have two outputs, current work and learning. Firms that capture only part of the future value of expertise will over-automate training tasks (the "apprenticeship wedge").
- Verified output is capped by skilled verification capacity, so cheap drafts can raise premiums for senior verifiers while making entry harder.
- Cited evidence: Noy and Zhang (2023) 40% less time and 18% higher quality on writing tasks; Dell'Acqua et al. (2026) 12.2% more tasks and 25.1% faster inside the frontier, 19 percentage points less likely to be correct outside it; BLS projects 15% growth for software developers, QA analysts and testers, 2024-2034.
- Seven policy implications, including "question rights" inside firms and an AI balance sheet that sorts tasks into automate, compress, augment, preserve and teach.
How they got it
A stylized conceptual model: formal definitions, local approximations and accounting decompositions built on the task-based labour economics of Autor, Acemoglu and Restrepo. The authors state it is neither estimated nor fully solved. Empirical claims come from published studies, not from the authors' own data.
Limits
- No new data: every number is borrowed from other studies, and none of the equation parameters is measured.
- Both authors work in industry (a technology services firm and an asset manager); the claims are argued, not tested.
Why it matters for companies
The paper gives managers a checklist for any AI deployment: what verification costs, who is liable, and who frames the questions. Its "automate, compress, augment, preserve, teach" split is a usable way to review task inventories before cutting junior roles. It warns that removing entry-level drafting work may save money now and leave the firm short of reviewers and experts in five years. It also argues that firms should decide explicitly who may frame problems for AI and who can stop a workflow when answers are unsupported.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org