Within Five Years We May Have AI That Does Science
Science is moving into a strange new phase: AI already helps write papers, review them, and summarize them for broad audiences. The next step is bolder-AI proposing the questions we study and the hypotheses we test.
That future isn't abstract. Ágnes Horvát and Robert West point to clear signals from social platforms, journals, and conferences that this shift is underway-and accelerating.
What's happening to science online
Most research attention now flows through feeds and video platforms. That reach comes with tradeoffs: extreme compression, incentives for hype, and a steady influx of AI-shaped text.
Evidence is piling up. Horvát's team found unmistakable LLM traces in 2024 biomedical abstracts-roughly 13% showed signs of AI "massaging," flagged by a lexicon of several hundred telltale terms. West's group observed a similar pattern on the peer-review side: at ICLR 2024, at least 16% of reviews used LLM assistance, implying a loop where AI drafts papers, AI reviews them, and people read AI summaries afterward.
- Social media still boosts citations for scientists, but the benefit is trending down.
- Compression favors certainty. AI systems amplify that by default-they must return an answer.
Net effect: more output, smoother prose, higher certainty signals-regardless of true uncertainty.
Misinformation risk and persuasion dynamics
The information supply chain is fragile. Content is remixed with weak provenance, while AI systems can generate persuasive text at superhuman scale and near-zero cost. If bots push false claims, speed and volume do the rest.
Detection is not a safety net. Current AI detectors undercount. Treat any single tool's "AI-written" score as a weak signal, not a verdict.
One open question worries both researchers: homogenization. If LLMs smooth style and structure, they may also narrow the space of ideas that make it onto the page.
If AI starts proposing hypotheses
Five years is a credible timeline for AI-generated research ideas to enter mainstream practice. That raises two hard questions: Will those ideas be good? And do they reflect human priorities?
AI can read everything and surface patterns humans miss. But it doesn't care about values by default. If we let it steer questions, we need guardrails that keep meaning, ethics, and societal impact in view.
A practical playbook for research teams
- Disclosure and governance
- State where and how AI was used (ideation, writing, code, review). Keep prompt logs and model/version notes.
- Adopt a simple rubric for "allow/flag/ban" use cases across your lab.
- Writing without false certainty
- Calibrate language. Replace confident claims with effect sizes, CIs, and limitations up front.
- Run a "hedging pass" after any AI-assisted draft. Add uncertainty and failure modes explicitly.
- Provenance and reproducibility
- Link datasets, code, preregistrations, and model cards. Use persistent IDs where possible.
- Record data lineage. Note synthetic data and augmentation steps.
- Channel strategy that doesn't reward hype
- Anchor everything in a long-form artifact (paper, prereg, protocol). Share shorter posts that point back.
- Publish negative results and ablations. They travel less-but build long-term trust.
- Review with AI, not by AI
- Use LLMs for checklists (claims-to-evidence mapping, reference sanity checks), then make the judgment yourself.
- Rotate spot checks: re-run analyses, replicate small sections, and compare to prereg goals.
- Misinformation countermeasures
- Prebunk likely misreads in a short FAQ attached to the preprint.
- Rate-limit auto-posting and label team-run bots. Keep source links prominent.
- Metrics that matter
- Track preprint downloads-to-citation ratio, saves/bookmarks, replication interest, and time-to-first-misquote.
- Use altmetrics as early signals, not success criteria.
Working with AI-generated hypotheses
- Define selection criteria: value to the field, societal impact, feasibility, cost, and falsifiability.
- Stress-test ideas before you code: adversarial prompts, negative controls, and prior work triangulation.
- Preregister study plans where suitable. Commit to pass/fail thresholds in advance.
- Run small pilot studies to measure signal quality before scaling.
The leadership agenda for the next five years
- Set a lab policy for AI use across the research cycle (ideation to outreach).
- Invest in data governance: permissions, consent, and documentation for human and synthetic data.
- Revisit incentives: reward careful uncertainty reporting, replication work, and public clarifications.
- Upskill your team in prompt design, evaluation, and disclosure norms.
Two closing realities can coexist. First, AI can make scientific writing clearer and peer review faster. Second, it can flood the zone with persuasive nonsense. Our job is to build processes that keep the good and filter the bad-before AI starts picking our questions for us.
Related references:
- International Conference on Learning Representations (ICLR)
- Nature editorial on responsible use of generative AI in research
If your team is formalizing AI practices or training new staff, you can review focused resources for researchers here: AI courses by job and prompt engineering.
Your membership also unlocks:
