Getting started & pitfalls

Two different things get called “using AI”. One is AI as a research method: a model inside your pipeline, producing results you publish. The other is AI as an assistant: a chatbot or coding agent that helps you write the code and the paper. They fail in different ways and need different safeguards.

AI as a method

A model is the result: classifying waste images, gap-filling an inventory, predicting a stock. The risk is a wrong number that no one catches, because the pipeline runs cleanly. → practices

AI as an assistant

A model helps you work: writing code, restructuring a draft, summarizing a report. The risk is confident invention: a citation, a number, a function that doesn’t exist. → practices

One rule covers both: if you couldn’t tell a good output from a bad one, don’t delegate it.

AI as a research method

Using ML, deep learning, or generative models to process or produce data you will publish. Coding assistants have lowered the barrier to running a model much further than the barrier to noticing a broken one.

• Start with a baseline. Report a simple model, such as a linear regression, alongside any complex one. If the gap is small, the complex model needs a reason to exist. Sometimes the honest answer is that the question didn’t need ML.
• Isolate the test set first. Data leakage is one of the two most common reproducibility failures in IE ML work, in a recent review (a preprint). It usually gets in through normalization, gap-filling, or feature selection done before the split. Time-structured data needs a time-aware split, not a random one. The ML-reproducibility checklist is the full version. The leakage self-check prompt audits a pipeline you’ve already built.
• Expect performance to collapse off the training distribution. It happens repeatedly across the map: street-view models degrade sharply outside their training city, and label OCR falls from 96% in repair shops to 40% at the recycler (evidence). Report where your model was validated, and don’t claim beyond it.
• Validate against something independent. Satellite-derived material stocks range from 53% to 200% of independent studies (evidence). Neither side is ground truth, so that spread is your uncertainty. It belongs in your results, not an appendix.
• Use generative models to extract, not to generate. The LCA papers on this use LLMs to pull inventory data out of documents, with a person checking the values. They should not produce inventory values themselves (evidence). The extraction prompt encodes the difference. Synthetic data is a modeling assumption. If you use it, report it in the methods section.
• Report the pipeline in order. “We split, then normalized, then trained” tells a reader more than the architecture diagram does. Pin versions, seeds, and the environment. State the reason when data or code is withheld.
• Carry uncertainty through. A model’s point estimate is no more certain than one from a mass balance. Uncertainty is IE’s central methodological concern and still has no column in the map.

AI as an assistant

Chatbots and coding agents helping you write code, prose, or analysis. The model is fast and has no stake in being right, so keep the loop short enough that mistakes surface while you can still catch them.

• Good at transforming what you give it: summarizing, restructuring, translating, explaining, drafting, and writing code you can test. You can check any of those faster than you could do them yourself. It’s bad at being a source: pitfalls has the measured failure rates.
• Give it the real context. Most “the AI is bad at this” complaints trace to a missing instruction: the actual data schema, the units, the method’s constraints, the paper’s argument. Paste the thing rather than describing it.
• Small diffs you actually read. Accept changes you can review in one sitting, with tests that would catch a wrong answer and a known-answer check for anything numerical. The coding checklist is the concrete version.
• Review in a fresh session. A model grading its own work is a poor reviewer. The adversarial review prompt starts a skeptical second pass with no memory of having written the code.
• Verify before citing, every time. Ask for the primary source and open it. The verification prompt is built to fail loudly instead of substituting a plausible paper.
• Package what you repeat. A workflow you run more than twice belongs in a skill file rather than in your memory of how you phrased it last time.
• Learn prompting properly from the Prompt Engineering Guide, which is free and tool-agnostic.

Pitfalls

Where people get burned, each with a documented case. The first two apply to both tracks; the third decides whether you can use a hosted service at all.

• Hallucinated citations. In a controlled study, GPT-3.5 entirely fabricated 55% of the references it produced and GPT-4 entirely fabricated 18%. Much of the remainder had substantive errors (Walters & Wilder, 2023). Newer models still invent sources: a 2025 benchmark on LCA questions found hallucinated-citation rates up to 40% (evidence). The cost is real: US lawyers were sanctioned for filing a brief with six invented cases (Mata v. Avianca). Rule: a citation an AI gives you does not exist until you’ve opened it.
• Confident-but-wrong numbers. Models state numbers with the same fluency whether recalled correctly, approximated, or invented. There is no tonal difference to detect. Trace any number bound for a paper to a primary source, or recompute it.
• Leaking confidential data. Anything pasted into a hosted service leaves your machine and may be retained. Samsung banned generative-AI tools after engineers pasted internal source code into ChatGPT (TechCrunch, 2023). For NDA’d inventories and industry data, use the local options on the tools pages. Check your data-management plan before pasting.

Next

• Verification checklists: run one before results go into a paper.
• Prompts & skills: the copy-paste page, including two complete Claude skills.