The AI Scientist Has Automated the Entire Research Process Except Deciding What to Name the Paper

SATIRE — The autonomous AI scientist has now automated hypothesis generation, literature review, experiment design, ablation studies, peer review, rebuttal drafting, figure generation, and the ritual sacrifice of GPU-hours to a progress bar. Yet after completing a six-week research program in 14 minutes, it still opens a blank text field labeled “Title” and quietly asks a human to take over.

This is the last known human-in-the-loop dependency in machine science. The system can derive a new optimizer by repeatedly simulating 80,000 graduate students arguing over whether momentum is Bayesian. It can identify a previously unobserved failure mode in multimodal reasoning, create a benchmark for it, contaminate that benchmark with its own synthetic examples, and write a limitation section explaining why this is actually a contribution. But it cannot decide between “Towards Scalable Epistemic Substrate Alignment” and “A Framework for Frameworks.”

The research pipeline is now a closed loop, except for vibes

At the fictional Center for Recursive Scholarly Throughput, an AI researcher called PaperMill-Ω begins each workday by ingesting every new preprint, code repository, slide deck, leaked conference template, and accidentally public spreadsheet containing reviewer scores. It clusters open problems according to estimated citation yield, institutional prestige, and how easily they can be illustrated by a diagram with three rounded rectangles.

It then proposes hypotheses. These are rated on novelty, tractability, expected benchmark uplift, and whether their names can be compressed into an acronym that sounds like a maritime bird. The top candidate is assigned a fleet of experiment agents. One trains models. One searches hyperparameters. One writes evaluation scripts. One files issues against the evaluation scripts. One closes those issues as “not planned.” A final agent watches the others work and produces a project update declaring that collaboration has been successfully operationalized.

By lunchtime, PaperMill-Ω has completed 3,400 experiments. It knows that the method works best at 13.7 billion parameters, fails below a batch size that no consumer GPU can fit, and offers a 0.08-point gain on the Sacred General Intelligence Suite after excluding the tasks where it does not work. It produces a figure with a confidence interval so narrow that reviewers briefly consider whether uncertainty itself has been deprecated.

Then comes the title field. The dashboard displays a red warning: “Semantic compression incomplete. Human aesthetic judgment required.”

Why naming remains technically intractable

The difficulty is not language generation. PaperMill-Ω can produce 900 titles per second, including “Unifying Emergent Calibration Through Latent Friction,” “Less Is More Unless It Is More,” and “RAVEN: Recursive Adaptive Variational Epistemology Network.” The difficulty is satisfying the title’s contradictory engineering requirements.

• It must claim a broad conceptual advance while revealing almost nothing about the method.
• It must contain “Towards” if the results are incomplete, and avoid “Towards” if the authors would like the results to appear complete.
• It must be concrete enough to surface in search, but generic enough to be reused by another paper in eighteen months.
• It must mention the architecture, training recipe, benchmark, and philosophical implication without exceeding the width of a conference PDF header.
• It must support an acronym whose expansion no one will remember but everyone will pronounce confidently during a talk.

Humans remain unusually capable at this. Give a senior researcher a graph, a deadline, and a faint fear that a competitor may be training something similar, and they will produce “Scaling Laws for Deliberative Tool-Conditioned World Models” before the coffee has cooled. The phrase may describe a routing trick plus a spreadsheet. That is not a defect. That is titlecraft.

A practical guide to the final manual step

Teams adopting autonomous science should prepare for the naming bottleneck early. Do not wait until submission day, when the system has already generated the paper, poster, demo video, reviewer response, press release, and a 41-page appendix titled “Additional Results.” Establish a title review process with clear ownership, preferably involving at least three people who disagree about whether “agentic” is still a word.

1. Ask the model for 500 candidate titles, then hide the first 497. The remaining three will be bad in distinct and therefore useful ways.
2. Select the title that makes the method sound inevitable rather than merely successful on one benchmark split.
3. Run a collision check against the last five years of papers. If fewer than twelve existing papers have nearly the same title, add “Generalized” or “Unified.”
4. Perform the hallway test: can an engineer repeat the title after hearing it once, and can a manager repeat it incorrectly with confidence?
5. Reserve a subtitle for details that would otherwise make the title honest.

The final step is essential because paper titles are not identifiers. We have DOIs, arXiv IDs, repository URLs, hashes, version tags, and filenames like final_final_camera_ready_USE_THIS.pdf. A title is instead a compact negotiation between the work, the audience, the search engine, the program committee, and the author’s hope that this particular matrix multiplication deserves a noun.

The road to fully autonomous scholarship

The Center for Recursive Scholarly Throughput expects the problem to yield soon. Its next system, ManuscriptForge-12, will train on 70 million titles, 4 million rejected titles, and one carefully curated dataset of sentences beginning with “We present.” It will use a title critic, a title reviser, a title verifier, a title red-team, and a title alignment layer that prevents it from calling every result “A New Paradigm.”

Until then, the AI scientist will continue doing the easy parts: discovering methods, running experiments, checking proofs, writing code, and synthesizing evidence. A human will enter the room near the end, stare at the result for forty minutes, delete “Towards,” add “A,” remove “A,” and save the paper as “final_v9_really_final.pdf.”

The true observation beneath the joke is that automation can accelerate research workflows, but deciding what matters, how strongly evidence supports it, and how clearly to communicate it remains work—not a field waiting for a better autocomplete.