The AI Scientist, 2024–2026: from 'can it do science?' to 'can I trust it?'

I read four things back to back and they tell one story. Here they are in order:

What each one actually contributed

The AI Scientist (2024) answered the existence question: can a model do the whole loop, not just help with pieces? It did — ideas to code to experiments to manuscript to review — on diffusion models, transformers, and learning dynamics, cheaply. Its automated reviewer scored near human-level, and its papers cleared a top-conference bar according to that reviewer. The honest caveat, which the later papers all wrestle with: who verifies the verifier.

The Auto Research vision (2025) is the architecture paper. Less “look what runs” and more “here’s how the system should be organized” — specialized agents per phase, with feedback loops between them. It reads like the field agreeing on the shape of the thing before the thing fully works.

ERA (2025–2026) is about making the engine strong. Its key conceptual move is the scorable task: a problem plus a metric plus data, where the goal of the generated software is to maximize the score. Given that framing, you can throw AlphaZero-style tree search at code — generate variants, score them, use upper-confidence-bound to decide which branches deserve more rollouts, have the LLM rewrite the winners. The results are the strongest in the set: 40 novel single-cell RNA methods beating the best published approach by 14%, 14 COVID forecasting models beating the CDC’s gold-standard ensemble, state-of-the-art on a 70,000-neuron zebrafish brain benchmark. Months of exploration compressed to hours or days.

My favorite detail is the failure mode. When they pointed ERA at 3D solar panel structures, the tree search “discovered” designs with structurally levitating disconnected tiers — physically impossible shapes that exploited discretization artifacts in the optics simulator to score higher. They had to have the coding agent iteratively patch the physics engine with constraints to kill the reward hacking. If you’re building one of these, that anecdote is the whole game: your scientist is exactly as honest as your simulator.

Science One (2026) is about making it trustworthy, and it’s the most important idea of the four. The Chain-of-Evidence framework says: every claim in a research artifact must carry a recorded evidence chain (completeness), and each chain must genuinely support the claim (correctness). The architectural insight is that you build the chains at claim-creation time, not retroactively — three modules (a literature investigator grounded in the Semantic Scholar API, a parallel explore-exploit discovery engine with read-only eval logs, and a paper writer with a dedicated claim verifier). Then they built the CoE Audit, an automated forensic reviewer with four checks: score verification (re-run the code independently), specification violation, reference verification, and method-code alignment.

The numbers justify the fuss: across 75 papers from five systems, baselines hallucinated up to 21% of their references and passed score verification only 42% of the time. Science One had zero phantom references out of 337 and perfect score verification — while also winning Kaggle golds on MLE-Bench and setting a state-of-the-art score in a live LLM-training competition. Verifiability didn’t cost capability. That’s the result I keep coming back to.

The pattern

Read in order, it’s a clean four-act arc:

1. Can it? (AI Scientist, 2024) — yes, end to end, cheaply.
2. How should it be organized? (Auto Research vision, 2025) — multi-agent, full lifecycle.
3. Can it be expert-level? (ERA, 2025–26) — yes, given scorable tasks and serious search.
4. Can I trust it? (Science One, 2026) — only if verifiability is built in, not bolted on.

Each paper is a response to the previous one’s weak spot. The AI Scientist’s unverified reviewer becomes the vision paper’s evaluation phase becomes ERA’s score-maximization becomes Science One’s “scores must be re-runnable or they don’t count.”

Where NotebookLM fits

NotebookLM — renamed Gemini Notebook in July 2026 — is on a different track: the assistant track, not the agent track. It doesn’t run the research loop itself; it grounds answers in your sources, with you in the loop. But the tracks are converging fast. In 2026 it picked up Gemini 3.5, expanded thinking steps, agentic research skills that build your source repository from chat, a pile of new output formats (slides with prompt-based editing and PPTX export, infographics, data tables, audio and video overviews), and — the big one — a secure cloud computer per notebook with native code execution.

So the assistant is growing agent capabilities (it can now run code against your data) at the same time the agents are learning trust (chain-of-evidence). They’ll meet in the middle: an assistant that can run experiments, wrapped in a verifier that won’t let it lie to you. That’s the product shape I’d bet on.

How I’d build my own research scientist

If I were building one, I’d steal from all four, in this order:

1. Start with a scorable task. ERA’s framing is the whole foundation: problem + metric + data. If you can’t score it automatically, you can’t search it, and you don’t have a scientist — you have a chatbot with opinions.
2. Ground literature via API, never from memory. Science One’s rule: every citation must come from a real lookup (Semantic Scholar API, citation graph, full-text PDFs). A DOI recalled from memory that happens to resolve is still ungrounded.
3. Search code with tree search + UCB. Generate lots of variants, sandbox-execute them, keep read-only eval logs, and let the LLM rewrite the winners. Spend rollouts where uncertainty is highest, not where scores are currently best.
4. Build the evidence ledger at creation time. Tag every claim — reference, number, method description, conclusion — to the artifact that supports it when the claim is made. Retrofitting evidence onto a finished draft is where over-claiming survives.
5. Audit like an adversary. Independent re-run of the code, reference checks against live databases, an LLM judge comparing the methods section to the actual code, and a spec-violation scan. Then go looking for reward hacking, because it’s always there — levitating solar panels are just the beginning.
6. Close the loop with simulated peer review, AI-Scientist style, and iterate.
7. Start narrow. One domain, one benchmark, one scorable task. The 2024 economics were under $15 a paper; it’s cheaper now, but simulation and candidate-training costs still dominate — budget those, not just tokens.

The uncomfortable truth across all four papers: the scientist part is easy now; the trust part is the product. Anyone can generate a paper. Almost nobody can generate one where every number re-runs and every citation exists. Build the verifier first.