Deep Dive Nº 07 · Research

AI for your dissertation, without getting caught out

The right tools can find, sort and summarise real papers in minutes. A general chatbot will happily invent the ones that do not exist. Here is how to tell them apart, and the single rule that keeps you out of the misconduct office.

It is the week before your literature review is due, and a chatbot has just handed you ten perfectly formatted references: real-sounding authors, a plausible journal, a tidy DOI on the end. You paste them into your bibliography and move on. The problem is that some of those papers were never written. The model produced the shape of a citation, the way it produces the shape of a sentence, and it cannot tell the difference between a study that exists and one that merely sounds like it should.

This is the trap at the centre of using AI for research, and it is entirely avoidable. There is now a small class of tools built specifically to search the actual scientific literature, ground their answers in papers you can open, and show you the sentence a claim came from. Used well, they turn a fortnight of database wrangling into an afternoon. Used carelessly, or confused with a general chatbot, they are a fast route to an academic misconduct panel. The difference is a workflow, not a secret.

A chatbot does not look things up and then tell you. It predicts what a plausible answer looks like. Sometimes that plausible answer is true.

Why the chatbot lies so wellThe fabricated-citation problem

Start with the failure mode, because it explains everything that follows. A general model such as ChatGPT, Gemini or a plain Claude chat is a next-word predictor. Ask it for sources and it generates text that resembles a reference list, assembled from patterns in its training data rather than retrieved from a database. The result is what researchers politely call a hallucinated reference: a citation that looks impeccable and points at nothing.

The rate is not trivial. A 2025 Deakin University study that checked references generated by GPT-4o found that roughly one in five citations were entirely fabricated, and more than half were either fake or carried factual errors. Earlier work published in Scientific Reports documented the same behaviour: invented titles, real authors bolted to papers they never wrote, DOIs formatted flawlessly and leading nowhere. The citations are convincing precisely because sounding right is the only thing the model is optimised to do.

The one rule

Never cite a source you have not opened and read yourself. If an AI mentions a paper, treat that as a lead, not a fact. Find it on the publisher's site, Google Scholar or your library, confirm it exists, confirm it says what the AI claims, then cite the paper. If you cannot find it, it very likely does not exist, and a fabricated citation in your bibliography is misconduct whether you knew or not.

Tools that cite real papersThe discovery engines

The fix is to use tools that retrieve first and write second. Instead of inventing text, they search an index of genuine publications, then summarise what they actually found, with a link back to each one.

Grounding on your own readingNotebookLM and Claude Science

A second category does something subtly different. Rather than search the whole internet, these tools answer only from the documents you give them, which is exactly what you want once you have gathered your papers and need to interrogate them.

Google NotebookLM lets you upload up to fifty sources, PDFs, your own notes, lecture slides, and then ask questions across them. Because it is grounded in your uploads, it will not wander off into invented facts, and every answer carries inline citations you can click to jump to the exact passage. It is superb for "what did these six papers say about my variable" and for turning a stack of reading into a revision summary. One honest limit: it points you to the source, but you still write the page-level reference yourself.

At the heavier end, Anthropic launched Claude Science at the end of June 2026, an AI workbench that connects Claude (Opus 4.8 included) to more than 60 scientific databases such as UniProt, PDB and ChEMBL. A coordinating agent parcels work out to specialists, and, notably, a separate reviewer agent checks every citation and calculation as it goes. It is aimed at working scientists in genomics, chemistry and structural biology rather than a first-year essay, but it signals the direction of travel: retrieval and verification built in, not bolted on.

Five tools · what each is for
ToolBest forFree access?
ElicitFinding papers, extracting data into comparison tablesFree tier, then paid credits
ConsensusEvidence-based answers to a specific questionFree tier, then ~£8/mo
Semantic ScholarConfirming a paper exists, citation graphsFully free
NotebookLMGrounded answers from papers you uploadFree
Claude ScienceDeep science across 60+ databasesPaid subscription

A workflow that keeps you honestDiscover, triage, read, cite

The tools only help if you keep them in their lane. A sane research loop looks like this, and the order matters:

Do it in that order and the AI never touches the part that has to be trustworthy. Skip to citing before reading, and you are one hallucinated DOI away from trouble.

By the numbersWhy verification is not optional

1 in 5
GPT-4o citations found entirely fabricated in a 2025 Deakin study
200M+
Papers Consensus searches to ground its answers
125M+
Papers covered by Elicit for search and extraction
60+
Scientific databases wired into Claude Science

Read the rules firstWhat your university actually requires

None of this is a licence to hand in AI-written work. UK policy is fragmented but converging, and the Russell Group principles, adopted across its 24 universities, are the clearest guide. The starting assumption in most schemes is blunt: unless an assessment explicitly permits it, generative AI is not allowed, and its output must be treated like any other person's work, acknowledged and cited, never passed off as your own.

In practice that means three things. Check the rubric for the specific assignment, because permission varies by module. Where AI use is allowed, declare it as your institution requires, usually a short statement of which tool you used and how. And understand that two things are always misconduct regardless of policy: submitting AI-generated text as your own writing, and citing sources you never verified. A fabricated reference is not a grey area. It is a fabrication.

The honest summary is cheering, not scary. These tools are genuinely good at the tedious 80% of a literature review: casting a wide net, summarising abstracts, organising what you find. They are not a shortcut through the 20% that is actually your degree: reading critically, thinking, and writing it up in your own voice. Keep AI on the discovery side of that line, verify every source with your own eyes, and you get the speed without the risk. Cross it, and the tool that was meant to save your dissertation is the thing that sinks it.

✦ End of issue ✦

Get the next one in your inbox

One AI tool, one prompt, three quick wins, every week. Free, and under 3 minutes to read.

Subscribe Free →