AI researchers are coming: What will happen to science?

Sakana.ai's updated 'AI scientist' produced 3 AI-generated papers, with 1 of the 3 accepted at a top-tier machine-learning workshop (overall ~70% of submissions passed the first round). The system can generate a paper for roughly $6–$15 and reduce researcher time from an estimated ~20 hours to ~3.5 hours, but outputs still suffer from underdeveloped ideas, structural issues, outdated or fabricated numerical results (hallucinations). The piece flags systemic risks — loss of human critical oversight and a potential 'monoculture' of science — leaving implications for research quality and standards uncertain.

Analysis

Commoditization of idea-generation will reallocate budget and bargaining power from human labor to compute, provenance and verification. Expect R&D groups to run many more experiment-design cycles per dollar spent — a plausible 2–5x increase in model training/validation runs per lab within 12–24 months — which lifts demand for GPUs, cloud credits, and experiment-tracking software more than for incremental headcount. This shifts margin capture toward hardware/cloud providers and specialists that can instrument and certify outputs.

A concentrated risk is reputational and regulatory: a handful of high-profile, harmful or fabricated AI-generated findings would catalyze fast policy responses (retraction waves, funder rules, or mandatory provenance) inside a 6–24 month window. That outcome would rapidly create paid demand for auditability, signed metadata, and tamper-proof logs — a winner-takes-most market for governance tools and enterprise vendors that integrate provenance into R&D workflows. Conversely, absent strong human oversight the system is prone to systematic biases embedded in training data, which are hard to detect without independent verification.

Competitive dynamics favor incumbents who can bundle compute, workflows and compliance: cloud providers (platform + marketplace), GPU vendors, and analytics firms that own citation and provenance flows. Traditional publishers and analytics companies can monetize “trusted” badges and gatekeeping, while small CROs and academic services face a squeeze if AI reduces billable human hours per paper; however, wet-lab execution remains a moat for the foreseeable future. The consensus fear of a homogenized “monoculture” is directionally valid, but underestimates the countervailing force: easier access to ideation may broaden the long tail of niche projects run by smaller teams, increasing overall experiment volume even as quality variance rises.

Key Decisions for Investors

• Long NVDA (NVDA) — buy shares or a 6–12 month call spread. Rationale: direct beneficiary of incremental training/validation cycles and inference demand. Risk/reward: target 15–30% upside with downside volatility risk of 25–35% if macro or AI regulation compresses multiples.
• Overweight Microsoft (MSFT) or Amazon (AMZN) cloud exposure — tactically via 12–18 month call spreads on MSFT/AMZN. Rationale: capture higher cloud spend for model hosting, MLOps and provenance services. Risk/reward: expect 10–20% incremental revenue lift in cloud AI customer cohorts over 12 months; downside if competing pricing pressure or enterprise pushback emerges.
• Long Palantir (PLTR) — 12 month horizon. Rationale: enterprise workflow, logging and provenance capabilities position it to sell governance layers to pharma, national labs and large corporates. Risk/reward: asymmetric 2:1 upside if adoption accelerates; risk of slow enterprise cycles and execution drag.
• Long Clarivate (CLVT) or RELX exposure (RELX.L) — 12–24 months. Rationale: companies that control bibliometrics and publisher relationships can monetize verification, provenance badges and premium curation. Risk/reward: modest 20–40% upside if publishers pay for trust infrastructure; regulatory pressure or reduction in paid subscriptions is the primary downside.