An AI answer is a starting point. Follow the evidence before you submit.

Two risks that deserve separate checks

An AI assistant can give a plausible description of an event, journal or research paper. The suggested item might exist but be unsuitable, or the item itself might be invented. These are different failures: one requires judging the source and its fit; the other requires establishing that the source exists at all.

A recommendation also has a time dimension. An authentic conference may have a closed deadline, an old edition or a different submission route. Save the edition and date you actually checked. A current-looking answer can be built from older material.

What a published study measured

Walters and Wilder’s 2023 study examined 636 citations in 84 generated literature reviews across 42 topics. It reported fabricated-citation rates of 55% for GPT-3.5 and 18% for GPT-4 in that experiment. [1]

The chart below reproduces those reported percentages. It is a historical bibliographic-citation experiment—not a 2026 performance score, a conference fraud estimate or a comparison of today’s search products. Model versions, prompts, retrieval tools and tasks matter when interpreting such results.

We use the study to motivate source checking. We do not infer how many conference mills appear in current chatbot recommendations, and our local conference-screening pilot did not measure that question.

Use a source-first prompt

Give the assistant your topic, intended contribution, travel constraints and target time window. Ask for a small set of candidates, official URLs and separately stated evidence for each deadline and publication claim. Require unknown fields to remain unknown instead of filling them with a guess.

An example instruction is: “Find candidate events for this topic. For each, provide the official edition page, the source for the submission deadline, the organiser’s own contact page and any uncertainty. Do not treat a directory listing as proof of proceedings indexing.” This is an editorial workflow suggestion, not a validated accuracy improvement.

Use the response to reduce your search space. It should not decide where you spend money or submit unpublished work. If the answer cannot provide a checkable source, move the candidate into an unresolved list.

Inspect the primary page and the claim behind it

Open the official source yourself and compare the event name, edition, venue and dates. Then check whether the programme and review process suit your contribution. Think. Check. Attend. offers a public checklist covering organisers, programme information and proceedings. [2]

For a cited paper, match the title, authors and identifier against the actual publication. For a claimed institutional partnership, check the institution’s own page. A logo on an event page does not answer whether the institution has confirmed involvement.

Do not upload a confidential manuscript merely to make a recommendation more specific. Use a topic summary unless the service and your institution’s policy permit sharing the underlying work. UNESCO’s guidance emphasises human-centred validation and privacy considerations in education and research. [3]

Make the decision reviewable

Keep a simple log: candidate, official URL, source checked, date checked, unresolved question and next action. Share that log with your supervisor or research office before an irreversible commitment. The final decision should be explainable without reopening the AI conversation.

ScholarVault’s Conference Finder and SCVS workflow can help structure candidate and evidence review. Read coverage and assessment limits alongside any result. Automated assessment and human verification are separate, and neither should be misrepresented as a guaranteed publication outcome. [4]

Your next step: take one AI recommendation and reconstruct its evidence trail. If you cannot establish the edition or the publication route, ask for clarification from the organiser through an independently confirmed contact channel.